[CS61C FA20] Weekly Lecture 06.LIVE - SDS & CL — Transcript

Full transcript

1. 0:00look at me
2. 0:03we've had many times where we forgot to
3. 0:04push record
4. 0:07all right welcome over here what's wrong
5. 0:09with my um
6. 0:11we just lost your there we go okay the
7. 0:13magic is being revealed here we go
8. 0:16cool now can you borrow when you try to
9. 0:19change slides what happens when you do
10. 0:21that
11. 0:21can you try to change because i have it
12. 0:23right now set to my
13. 0:25my screen like i have like eight slides
14. 0:28in a row
15. 0:29look at that now you're at static slides
16. 0:31right okay
17. 0:32that's fine static slides but if i go to
18. 0:35watch me if i go to the first one
19. 0:38then here this is now animation
20. 0:41and then watch here we go now it'll play
21. 0:43the movie watch should we go
22. 0:45and it'll play the movie see can you
23. 0:47guys hear any music anything at all
24. 0:49uh-uh not here no not not right yeah i
25. 0:52don't think mm-hmm
26. 0:53sends that sends the uh i don't think
27. 0:55him sends that
28. 0:56um the audio over but it's okay i can i
29. 1:00can get to there fast enough
30. 1:01all right so borah you're you're you're
31. 1:03i'll put it back on the static slides
32. 1:04that way you can control it as well
33. 1:06welcome students come on in i think we
34. 1:08have 54 folks here welcome everybody
35. 1:11good to see you all all right
36. 1:15let me see i think for your host i don't
37. 1:18think i can change your name
38. 1:20underneath i can i need to change it
39. 1:22myself
40. 1:24okay perfect
41. 1:29our faculty candidate is coming to your
42. 1:30talk this week i'm excited about that
43. 1:32that's
44. 1:32that's an exciting case so i'm looking
45. 1:34forward to that as well
46. 1:36yeah i'm also looking forward to that
47. 1:39all right folks we'll get started in
48. 1:41exactly 10 seconds we'll get rolling
49. 1:43here we go this should be really fun
50. 1:46uh let's see
51. 1:53okay all right 410.
52. 1:56wonderful welcome everybody to cs61c
53. 1:58week six
54. 1:59digital systems of million logic
55. 2:03great to have you all here great travel
56. 2:05to have you all here
57. 2:06we've got a wonderful guest already here
58. 2:08welcoming uh vosm sophia xiao as our as
59. 2:11our faculty guest
60. 2:12for today thank you so much for joining
61. 2:15us
62. 2:15um here is our brief agenda so we're
63. 2:18going to welcome you to week six
64. 2:20do a complete quick a quick computing in
65. 2:22the news
66. 2:23uh then we'll interview sophia and talk
67. 2:25to her about some of her research and
68. 2:26some of the exciting things she's
69. 2:27working on and
70. 2:28where she sees the space um we'll then
71. 2:30we'll let her take her leave and we'll
72. 2:32go back to our week's plan
73. 2:33uh and then have some whatever
74. 2:35announcements we have and then we'll do
75. 2:36the ask us anything at the end of this
76. 2:38perfect all right boy we'll take us into
77. 2:40computing the news
78. 2:42sure am i pinned i need to
79. 2:45feel free free to pin yourself yeah here
80. 2:47we go
81. 2:48pinned there you go hopefully this works
82. 2:51okay so um so
83. 2:55we we have been saying that now this uh
84. 2:58it is a really exciting time to be in
85. 3:03chip design in computer architecture in
86. 3:05general
87. 3:07because we are entering this era of
88. 3:10domain specialized computing and the
89. 3:12first thing that is really
90. 3:14has shaking up things is this need
91. 3:17to support machine learning machine
92. 3:19learning at a large scale
93. 3:22so we are seeing um you know there are
94. 3:24two
95. 3:25things that people need to do
96. 3:28two ways how people need to support
97. 3:30machine learning one
98. 3:32is training the data the other one is
99. 3:34running the inference on the train data
100. 3:37and there has been a number of companies
101. 3:40now
102. 3:41that have built custom chips for that
103. 3:45and in this picture um there is uh there
104. 3:48are
105. 3:49two uh clips that uh we took from
106. 3:52uh something it's a conference that is
107. 3:54called hot chips that was held in august
108. 3:57about a month ago
109. 3:59where some of them have been presented
110. 4:02some really exciting chips so google
111. 4:04talked about tpu
112. 4:05v2 and v3 that's their tensor processing
113. 4:08unit
114. 4:09that they're using to accelerate
115. 4:10training and inference in the cloud
116. 4:13you know tpu is what processes when we
117. 4:16do speech recognition or many other
118. 4:18facts
119. 4:19tpu was what uh uh beat
120. 4:22the goalmaster uh and it was an older
121. 4:26tpu i think
122. 4:26it might have been a tpu one or two um
123. 4:30that beat him um alphago i think it's a
124. 4:33name right alphago is the name
125. 4:36exactly um you know they don't publish
126. 4:40these as they make them
127. 4:41so we believe that tpu3 is probably
128. 4:43about two years old now
129. 4:45there must be tpo4 and tpu5 but they
130. 4:48talk about
131. 4:50the stuff that is already deployed there
132. 4:53is another
133. 4:53crazy chip that cerebrus has
134. 4:57presented and it is in their second
135. 5:00generation
136. 5:01you know when you build a chip um
137. 5:05you know they built something that is
138. 5:07called a wafer and wafer is like this
139. 5:08it's 300 millimeters
140. 5:10size but you know when we take our chips
141. 5:14we dice that wafer into tiny tiny pieces
142. 5:17so if you look at the cell phone uh cell
143. 5:19phone
144. 5:20chips like the ones that
145. 5:24apple makes are seven by eight
146. 5:25millimeter so they're called dies
147. 5:27so this whole wafer is diced into seven
148. 5:30by eight millimeters
149. 5:32if you take a desktop pc there may be
150. 5:34you know
151. 5:3515 millimeters on the side some really
152. 5:37big server chips are like
153. 5:39an inch on the side like a postcard size
154. 5:43basically yeah
155. 5:44uh not the it's a stamp size
156. 5:48yeah i meant a postage stamp i said i
157. 5:49said postcard i meant to say posted
158. 5:51stamp i meant to say
159. 5:52i know that it's supposed to stamp yeah
160. 5:53yeah and in fact this is a big deal the
161. 5:55larger area you have the
162. 5:56worse yield you have the higher chance
163. 5:58there is a imperfection on that so
164. 6:01you really try to care about having the
165. 6:02smallest area as possible to have the
166. 6:04highest yield from a yield means the
167. 6:06number of good
168. 6:06good chips you get out of a total slice
169. 6:09yep
170. 6:10and this place cerebros is building
171. 6:12chips that
172. 6:13are the entire wafer so there are 300
173. 6:16millimeters
174. 6:17on the diagonal which is insane why
175. 6:20don't why they need these big chips
176. 6:23because they want to do
177. 6:25training for machine learning on them
178. 6:27those things have to be
179. 6:28pretty pricey but they're really hoping
180. 6:31to make it in there
181. 6:33there is a number of other companies
182. 6:35that have their solutions
183. 6:37uh people might have heard that habana
184. 6:39was acquired by intel
185. 6:41uh then there is a few more startups and
186. 6:44some of them already gone down
187. 6:47but the new ones are are popping up so
188. 6:50this is an exciting
189. 6:51field of the domain specific computing
190. 6:53that is creating a lot of
191. 6:55interest out there um so speaking of
192. 6:59that
193. 7:00uh can we have uh can we switch over to
194. 7:04sofia
195. 7:06let me jump over
196. 7:09yeah uh sophia is a new faculty
197. 7:12in ecs that is really focusing on this
198. 7:15on
199. 7:16specialized architectures for computing
200. 7:21so she graduated from
201. 7:24harvard in 2016 and then she spent a few
202. 7:28years
203. 7:29in industry she worked in nvidia and
204. 7:32nvidia is the
205. 7:33leading place that has been selling
206. 7:36these more
207. 7:37general purpose gpus for machine for
208. 7:40training
209. 7:40and influence in machine learning
210. 7:44and last summer she joined berkeley
211. 7:48so welcome sophia hello everyone
212. 7:52and we is it okay if we ask you a few
213. 7:54questions maybe you can tell us
214. 7:56how did you get here um you know how did
215. 7:59you end up here and let's try to put us
216. 8:01let's try to did you always know you're
217. 8:03going to be an academic that kind of
218. 8:04question like
219. 8:05how did you even find find your way to
220. 8:06berkeley and all those decisions you had
221. 8:08to make yeah
222. 8:09uh that's those are really good
223. 8:10questions so first
224. 8:12good good afternoon everyone good to see
225. 8:14you all and
226. 8:15really good to uh thanks laura and then
227. 8:17for this wonderful opportunity to get to
228. 8:19interact with all of you
229. 8:21so as for a mission i'm sophia xiao i'm
230. 8:23an assistant professor here at uc
231. 8:26berkeley so how did i come here
232. 8:29so i grew up in china i did my undergrad
233. 8:32in georgia university it's a very
234. 8:36university in the southern part of china
235. 8:39and during
236. 8:40my undergrad i got really interested in
237. 8:43basically microcontroller programming
238. 8:45and fpga and i participated
239. 8:47in the embedded system and the robotic
240. 8:49competition
241. 8:50during the summer before my senior year
242. 8:52that's completely changed
243. 8:54the way i see hardware and software and
244. 8:56the way i see the possibility
245. 8:58in this area so after that competition
246. 9:01after that summer i decided that i want
247. 9:03to learn more i want to do grad school
248. 9:05i want to know more about computer
249. 9:07architecture and hardware design
250. 9:09so during 2009 i decided okay i want to
251. 9:12do grad school
252. 9:14so then i moved to boston i did my phd
253. 9:17at harvard
254. 9:18working on thinking about hardware
255. 9:20design with a special focus
256. 9:22in domain specific accelerators
257. 9:24understanding the challenges
258. 9:26with moore's law and then their skating
259. 9:28i assume you are really very familiar
260. 9:30with that
261. 9:30and then thinking about what we can do
262. 9:32in the hardware space to further improve
263. 9:34the performance
264. 9:36so after i finished my phd in 2016
265. 9:40i i was also thinking about okay whether
266. 9:42i want to do academia
267. 9:44or industry back then i really want to
268. 9:46already interned at both intel and ibm
269. 9:49before i see
270. 9:50a glimpse of how industry works but i
271. 9:52wanted to see more to really actually
272. 9:54build something
273. 9:56in this area and this is a very exciting
274. 9:58area
275. 10:00especially thinking about the amount of
276. 10:01effort in both academia and
277. 10:04industry so since i have already been in
278. 10:06academia for a few years
279. 10:07so i decided to okay let's move to
280. 10:09industry to see
281. 10:11what people are working on there so
282. 10:13that's why i decided to
283. 10:14join nvidia after my
284. 10:18phd and i spent three years
285. 10:21at nvidia it's a really exciting time as
286. 10:24boron mentioned there are a lot of
287. 10:26efforts in thinking about hardware
288. 10:28design
289. 10:28especially for hardware for machine
290. 10:31learning
291. 10:32so i was at media research where we are
292. 10:34also very interested
293. 10:36in thinking about what's the hardware
294. 10:37what's the hardware implications
295. 10:39of supporting all those important
296. 10:41emerging machine learning applications
297. 10:43and what kind of optimization that we
298. 10:45can do not only in architecture but
299. 10:47also in circuit maybe even underlying
300. 10:49device that we can do to improve the
301. 10:51performance
302. 10:52so i was there for three years with
303. 10:54wonderful mentors and
304. 10:56colleagues and working on a couple of
305. 10:58very interesting
306. 10:59and proud projects to actually build the
307. 11:01hardware and see the implications
308. 11:03uh and also the possibilities of
309. 11:05hardware for this area
310. 11:07and uh two years ago i think that's
311. 11:10roughly where we
312. 11:11just finished your paypal we finished
313. 11:13the breakout process and our trip worked
314. 11:15and that got me thinking okay what's my
315. 11:17next project what do i want to do next
316. 11:19i think that's when i started thinking
317. 11:21about okay
318. 11:23what what about academia and whether
319. 11:26i want to not only building a course
320. 11:28cool and interesting projects
321. 11:30but also interact interacting with
322. 11:32students both undergrad and grad
323. 11:34students
324. 11:35to really impact the next generation
325. 11:38students and also hardware engineers
326. 11:40so that's actually one of the major
327. 11:42factors i decided to
328. 11:44move back to academia and i'm really
329. 11:47glad that
330. 11:48i'm here at berkeley with wonderful
331. 11:50colleagues and students
332. 11:51so it has been an amazing year the past
333. 11:54year but i'm looking forward
334. 11:56to interacting with many of you in the
335. 11:58future
336. 11:59so i guess that's how i got here thank
337. 12:02you that's great it's wonderful
338. 12:03i have a question um another question so
339. 12:06how does one become a faculty at
340. 12:08berkeley i mean dan and i i think forgot
341. 12:10uh what does it take it has been a while
342. 12:1320 years at least for both of us
343. 12:15okay that's a good question i also try
344. 12:17to forget about that
345. 12:19so it's still pretty fresh in my mind
346. 12:22uh let me refocus this really quickly um
347. 12:26that's a good question so how to become
348. 12:29a faculty at berkeley
349. 12:32so so first uh you need to get a phd
350. 12:35typically for most other cases at least
351. 12:37some grad
352. 12:38grad degree i will i want to say that
353. 12:41this is a really good time to think
354. 12:43about
355. 12:44going to grad school maybe like at first
356. 12:46if you are not sure maybe try out
357. 12:48master degree first maybe also
358. 12:50eventually
359. 12:51get a trial the phd program this is a
360. 12:53really interesting time we definitely
361. 12:56see
362. 12:56a lot of exciting areas in both hardware
363. 12:59and also applications
364. 13:00and there is a strong demand in thinking
365. 13:03about
366. 13:03new innovative ideas in those areas and
367. 13:07a lot of those ideas actually coming
368. 13:09from academia who can actually
369. 13:12like stay a little bit away from like
370. 13:14the practical
371. 13:15very near-term deadline pressures but
372. 13:18really
373. 13:18thinking about what are the major
374. 13:20problems what are the big concern what
375. 13:22are the things that we
376. 13:23really need to think in five years maybe
377. 13:25even 10 years so this is a really good
378. 13:27time to think about
379. 13:28uh grad school so to become a faculty at
380. 13:31berkeley
381. 13:32typically you need a phd degree i guess
382. 13:35that's a step
383. 13:35one second i would say you need to do
384. 13:39all of us we really need to think about
385. 13:41not only
386. 13:43internal research when once you get into
387. 13:45the grad school you start thinking about
388. 13:46research
389. 13:47and then develop your own research taste
390. 13:50and also develop your own project
391. 13:52i would say one thing very important in
392. 13:54thinking about
393. 13:56becoming especially a berkeley faculty
394. 13:58is the impact
395. 13:59of your research sometimes it's very
396. 14:02easy to get lost because oh
397. 14:04i want to publish more papers i want to
398. 14:06uh
399. 14:07like get in papers in like years and
400. 14:10hopefully
401. 14:11that will actually make my research
402. 14:12stand out but
403. 14:14most of the time it's not about the
404. 14:16quantity it's really about the quality
405. 14:18especially for institutions like
406. 14:20berkeley it's really
407. 14:22important for for faculty members and
408. 14:24also for grad students like for
409. 14:26for all of us working in this area not
410. 14:29only just thinking about how many papers
411. 14:31we publish
412. 14:32really think about what kind of impact
413. 14:34we make for this area for berkeley
414. 14:36for our research community and for the
415. 14:38entire
416. 14:39society so i think the impact-driven way
417. 14:42of doing research is actually really
418. 14:44important i would say that's definitely
419. 14:47based on my experience
420. 14:48as someone who goes through the process
421. 14:50and interact with
422. 14:52all the berkeley faculties and mentors
423. 14:54and also seeing different
424. 14:55and career paths of colleagues and
425. 14:58mentors
426. 14:59i think having an impact-driven research
427. 15:02mindset is really important uh in
428. 15:05in in overall your your career paths no
429. 15:07matter whether you come to berkeley or
430. 15:09go elsewhere really think about your
431. 15:11impact of what you are doing
432. 15:12instead of just some quantifiable
433. 15:15metrics
434. 15:16uh i guess the last thing i want to say
435. 15:18especially become
436. 15:20a faculty at berkeley this is partially
437. 15:22like i mentioned earlier partially the
438. 15:24reason i decided to
439. 15:26when i came back to academia is to
440. 15:28really thinking about interacting with
441. 15:30students
442. 15:30both undergrads and grad students and
443. 15:33also this is something i
444. 15:34i observe here at berkeley is that to be
445. 15:37a faculty member here you really need to
446. 15:40care about teaching
447. 15:41care about mentoring care about advising
448. 15:44because we spend a lot of time
449. 15:46interacting with students
450. 15:48both undergrads and grad students and
451. 15:50the reason we are here
452. 15:52instead of being elsewhere in the
453. 15:53industry where that can be
454. 15:55a very well-paid job is really the
455. 15:58benefit
456. 15:58of interacting with the younger
457. 16:00generation and see what we can do
458. 16:02to actually learn with them together and
459. 16:04to
460. 16:05to as i mentioned earlier to do high
461. 16:07impact research and to impact to change
462. 16:09the field
463. 16:10so having a strong drive and also k and
464. 16:14caring
465. 16:14about students both undergrads and grad
466. 16:17students i think that's also
467. 16:19a very important factor oh that's my
468. 16:21understanding we also have two other
469. 16:22faculty members here i think
470. 16:24be interesting to hear their perspective
471. 16:26as well
472. 16:27i i've heard on your second point uh in
473. 16:30terms of just not just number of papers
474. 16:31but actually impactive papers i've heard
475. 16:33the analogy made
476. 16:34don't just get singles this is the
477. 16:36baseball analogy don't just hit singles
478. 16:37and get on base try to swing for the
479. 16:38fences try to have
480. 16:40a couple of home runs in there to make
481. 16:41some make a difference so i appreciate
482. 16:43that and i agree with that that being
483. 16:44the thing that's most important it's
484. 16:45just
485. 16:46not just paper trail of like well i did
486. 16:47a hundred papers but nobody reads them
487. 16:48has to if i did three papers that
488. 16:50everyone reads that actually has more
489. 16:51impact
490. 16:52than the hundred papers nobody reads so
491. 16:53that makes a lot of sense i appreciate
492. 16:55that
493. 16:57um yeah speaking of that
494. 17:00[Music]
495. 17:02you know this boolean logic thing the
496. 17:04mapping of boolean logic onto gates and
497. 17:06and switches uh that dan mentioned was
498. 17:09invented by claude chen and quan shannon
499. 17:11one of the best known
500. 17:13electrical engineers and information
501. 17:15theorists didn't publish many papers i
502. 17:17mean
503. 17:20because you know he only went he only
504. 17:23went for the
505. 17:24big stuff no question for sophia his
506. 17:26master thesis
507. 17:27so he didn't yeah and if you watch
508. 17:30today's video if you watch today's video
509. 17:32i say that today in today's
510. 17:33video exactly
511. 17:37a question for sophia um
512. 17:41tell us something about your research so
513. 17:43how to swing for defenses in this uh
514. 17:46in this domain all right well that's a
515. 17:49good question so as well mentioned i'm
516. 17:51really interested in domain specific
517. 17:53hardware and we're gonna talk about some
518. 17:55of the really exciting
519. 17:56developments in the industry these days
520. 17:59in thinking about hardware for machine
521. 18:02learning which is a very important
522. 18:04applications in the industry today
523. 18:06and it's really exciting to see not only
524. 18:08this new
525. 18:09development from traditional hardware
526. 18:11companies but also startups
527. 18:13and also software companies like google
528. 18:15so we definitely see a lot of
529. 18:16excitement in this area so definitely we
530. 18:19are really interested in this area in
531. 18:21thinking about
532. 18:22understanding applications behaviors and
533. 18:24explore different
534. 18:27hardware acceleration strategies
535. 18:29specifically we have three focus
536. 18:31in the way we think about this hardware
537. 18:33for machine learning research
538. 18:35the first one is definitely related to
539. 18:37individual
540. 18:38algorithm acceleration what are the
541. 18:40emerging algorithms before i mention
542. 18:42their training and inference and even
543. 18:43within training and inference there are
544. 18:45different networks different
545. 18:46applications
546. 18:47and there are also emerging algorithms
547. 18:49showing up on a daily basis maybe even
548. 18:51on hourly basis in actually
549. 18:53important play a very important role in
550. 18:56in the machine learning process
551. 18:58so we are definitely really interested
552. 18:59in understanding the application
553. 19:01behaviors
554. 19:02and also explore potential hardware
555. 19:04mechanism
556. 19:05to make them run more efficiently
557. 19:07especially in power constrained
558. 19:09devices so that's the first uh angle the
559. 19:12second one
560. 19:13what we are really interested is also
561. 19:15thinking about not only individual
562. 19:17accelerators but how those different
563. 19:19accelerators
564. 19:20work together so we talk about the
565. 19:22importance of having domain specific
566. 19:24accelerators
567. 19:25for those emerging applications and they
568. 19:28all have different behaviors
569. 19:30and we'll see how the different
570. 19:31applications and accelerators need to
571. 19:34really work together
572. 19:36especially in today's complex soc where
573. 19:38we have all the different
574. 19:40like 30 or 40 different accelerators
575. 19:42they need to interact with each other
576. 19:44and passing data from and to each other
577. 19:46so how to actually make sure
578. 19:48they can coordinate in a consistent way
579. 19:51and also achieve performance and the
580. 19:53efficiency benefit is also very
581. 19:55important on our agenda
582. 19:57so we talk about individual accelerator
583. 19:58and also how they work together
584. 20:00and finally i think another area we i i
585. 20:03personally always have a soft spot um
586. 20:05it's really thinking about
587. 20:07from a methodology point of view how we
588. 20:09can from
589. 20:10thinking about all the different efforts
590. 20:12need to pour into
591. 20:14design and also specialize different
592. 20:16applications
593. 20:17on different hardware platform what we
594. 20:19can do in the methodology space
595. 20:21to help designers to navigate this space
596. 20:24and make it easier and also more
597. 20:26productive
598. 20:27to to design new hardware i think those
599. 20:29three
600. 20:30directions are what we are really
601. 20:31interested in and also actively working
602. 20:33on thinking about
603. 20:35accelerator individual acceleration
604. 20:37system integration
605. 20:38and also design methodology here
606. 20:42all right thank you what's your opinion
607. 20:45about these two chips that we have shown
608. 20:47the
609. 20:48tpus and and uh reverses they're in your
610. 20:50space
611. 20:51aren't they right that's a good question
612. 20:54i think
613. 20:55as i mentioned earlier first it's really
614. 20:57exciting to see the development
615. 20:59from like non-traditional hardware
616. 21:02vendors so we see
617. 21:03most of the time when we talk about
618. 21:04hardware designs intel
619. 21:06amd ibm and the nvidia those are
620. 21:09basically the major players
621. 21:11and it's interestingly although everyone
622. 21:12is actually participating in the machine
623. 21:14learning space
624. 21:15the two examples that borah mentioned
625. 21:16earlier one is from google
626. 21:18which is actually more software company
627. 21:21uh and another one is actually from a
628. 21:23startup surprise so i think the area is
629. 21:26getting really interesting these days
630. 21:28with not only uh participation from
631. 21:30major hardware vendors but also from
632. 21:33new players both traditional software
633. 21:36companies
634. 21:36and also startups all of those all of
635. 21:40them actually
636. 21:41are participating in this hardware for
637. 21:43machine learning
638. 21:44design so i think one thing definitely
639. 21:47stands out
640. 21:48in both two examples that borah showed
641. 21:50earlier is
642. 21:51the scale so both the cerebrus chip we
643. 21:53talked about how big it is
644. 21:55uh and all the different components they
645. 21:57need to actually work together and also
646. 21:59the tpu v3 a lot of the performance
647. 22:03is actually on multi-node hundreds or
648. 22:05even thousands of nodes
649. 22:07all of those need to actually work
650. 22:09together so definitely we see
651. 22:11a lot of the um the especially the
652. 22:14emerging training
653. 22:15accelerators turning some in some way
654. 22:18similar to hpc problem where not only we
655. 22:21need to design each individual node very
656. 22:24efficiently
657. 22:24but i also need to think about how the
658. 22:26different nodes actually work together
659. 22:28in the consistent fashion so that's
660. 22:30definitely very
661. 22:31important these days in the hardware for
662. 22:33machine learning space
663. 22:35at the same time i mentioned both two
664. 22:36are mostly like training and large scale
665. 22:38we also see really exciting development
666. 22:41and like
667. 22:42in the in the edge space where like low
668. 22:45power
669. 22:46low power devices even microcontrollers
670. 22:48how could they actually potentially
671. 22:50support
672. 22:51efficient machine learning algorithm
673. 22:53mostly of course in the inference space
674. 22:56that's also very important we also see a
675. 22:59lot of
676. 22:59a lot of development in those space so
677. 23:02we definitely see
678. 23:03maybe one is the scale especially for
679. 23:06the training the large scale data center
680. 23:09scale and another one is actually for
681. 23:10the edge devices under extremely
682. 23:12conditions
683. 23:13how can we actually design more
684. 23:15efficient
685. 23:16hardware in those scenarios sounds good
686. 23:20questions let me just add something here
687. 23:22kath when kathy ella came to visit us
688. 23:24she said that in mulch
689. 23:25in most of the hpc applications high
690. 23:28performance giving applications they
691. 23:29noticed that
692. 23:30um they were not processor limited but
693. 23:32they were i o limited so i o
694. 23:34just moving data around is the most
695. 23:36costly thing they've noticed as i did
696. 23:37you know
697. 23:38an audit of where where time is being
698. 23:40spent just moving data is just
699. 23:41remarkably expensive in terms of time
700. 23:44are we seeing that as much on the tpu
701. 23:47you know a million nodes of tpu v3
702. 23:51is it really that's where the what's
703. 23:53like what's the bottleneck i guess in
704. 23:55these new domain specific architectures
705. 23:56i guess that's the question
706. 23:57that's a good question i think first it
707. 24:00it really depends on the
708. 24:01applications uh so in this particular
709. 24:03case a lot of training
710. 24:05applications is still pretty compute
711. 24:07intensive
712. 24:08in the sense that we talk about the
713. 24:09computer memory ratio that cassie and
714. 24:12patterson they use to also use the roof
715. 24:14line model to quantify that
716. 24:15so a lot of those compute terminals they
717. 24:18are indeed very compute intensive so
718. 24:19there is significant compute actually
719. 24:21going on
720. 24:22in individual node but once we
721. 24:24especially look at some of the amount
722. 24:26perf
723. 24:27results where we are basically competing
724. 24:29to see the best performance that we can
725. 24:31get what's really interesting
726. 24:33in the sense that of course the easiest
727. 24:35way is basically scale up to any
728. 24:37like as many machines as possible
729. 24:39there's no limit in terms of the number
730. 24:41of nodes that you use
731. 24:42you could use as many nodes as possible
732. 24:45but typically the reason
733. 24:46all those different hardware or the
734. 24:48software vendor stop at a particular
735. 24:50node is because it doesn't scale anymore
736. 24:53if we see the reports of say oh we
737. 24:56reached we used 2000 nodes to reach this
738. 24:58performance
739. 24:59the reason they don't use 4000 nodes is
740. 25:01not because they don't have 4000 nodes
741. 25:03it's because when they actually further
742. 25:05split things up to force out the node
743. 25:07it actually they don't get better
744. 25:08performance so for those extreme
745. 25:10conditions we definitely see
746. 25:12given the application have so much after
747. 25:14computer memory reuse
748. 25:15we already see some of the compute bund
749. 25:18scenarios showing up
750. 25:19when we look at this kind of machine
751. 25:22learning space yeah
752. 25:23very good point that's great and all
753. 25:26gets us in the end
754. 25:28yeah i was gonna say where is that
755. 25:29where's amdahl in here exactly the
756. 25:30students don't know that yet but
757. 25:32they'll see it they'll see it in a
758. 25:33couple of couple of weeks exactly
759. 25:35exactly go ahead um yeah a question
760. 25:38quick question for you
761. 25:39um you teach 151 can you tell us
762. 25:44something about that uh you know can you
763. 25:45put a little advertisement for that
764. 25:47class
765. 25:48um about i think about 10 of this class
766. 25:51are gonna
767. 25:51wind up in 151 so
768. 25:55why should they do that why should they
769. 25:58of course i heard you had more than 1
770. 26:00000 students this year
771. 26:01so we should be expecting a lot of
772. 26:03students coming to 131 that's wonderful
773. 26:06so yeah i teach 14151 which is
774. 26:10introduction to digital logic and also
775. 26:12integrated circuit
776. 26:13it is uh intro to heart well not
777. 26:16necessarily intro but definitely
778. 26:17thinking about digital design and also
779. 26:19hardware design
780. 26:20where it's as the name suggests it's an
781. 26:22ee and the cs
782. 26:24course which it actually truly brings
783. 26:25electrical engineering and also computer
784. 26:27science
785. 26:28together so in this course you will
786. 26:30actually see
787. 26:31not only how to actually map your
788. 26:33program to
789. 26:34the different instruction set and but
790. 26:36i'll actually build
791. 26:38specific hardware following uh like
792. 26:41specific i say specifications
793. 26:43so that you can actually build your
794. 26:44hardware either in verilog or asic to
795. 26:47have
796. 26:47something actually fabricable in the
797. 26:49hardware in the end
798. 26:51it's also okay if you have a more ee
799. 26:52background i think i assume for this
800. 26:54cloud
801. 26:55mostly students have more cse background
802. 26:57but you will also learn actually a lot
803. 26:59of ee concept
804. 27:00where you how we attribute register file
805. 27:02how we actually build aou in hardware
806. 27:04how transistor actually behave
807. 27:06how can you assemble different
808. 27:07transistors together for all the
809. 27:09different digital logic so you will
810. 27:11actually see the entire stack
811. 27:13from i say to hardware implementation
812. 27:16and
813. 27:16see how the transistors behave it's
814. 27:19awesome it's this week that's this week
815. 27:21you just advertised for why the students
816. 27:23should watch this week's lectures i love
817. 27:25it
818. 27:26thank you thank you thank you so much
819. 27:31i love it so this week so it sounds like
820. 27:33this will will be very important
821. 27:35also for 151 so after you see the entire
822. 27:38stack in 151 you will actually build
823. 27:41real hardware in either fpga we
824. 27:44mentioned earlier that's a very cool
825. 27:45platform you can actually prototype
826. 27:47different functionalities
827. 27:48onto the board or using asic flow which
828. 27:51is actually pretty close
829. 27:53to the state of art commercial flow that
830. 27:55is able to actually eventually
831. 27:57take out a chip in the end so that's a
832. 28:00really cool class
833. 28:01covering a lot of hardware and also
834. 28:03hardware architecture details
835. 28:05and really useful for any hardware
836. 28:08engineering career that you're
837. 28:10interested in pursuing no matter whether
838. 28:11it's in industry or
839. 28:13academia it will enable you to build
840. 28:16actually physically built and in type of
841. 28:18hardware that you are interested
842. 28:20and it's also a very hands-on class we
843. 28:22have a very important lab component
844. 28:24you will actually go through the la go
845. 28:26through the lectures and go through the
846. 28:28lab
847. 28:28to see how the different things actually
848. 28:30mapped in hardware
849. 28:31so it's actually a direct follow-up of
850. 28:3361c you see all the different components
851. 28:36in 61c
852. 28:37and we'll go a little bit deeper to
853. 28:39actually see how we turn
854. 28:40all the different concepts that you
855. 28:41cover sounds like this week into
856. 28:43hardware
857. 28:46actually this week and next week is what
858. 28:49we
859. 28:49yeah exactly that's where 51
860. 28:52151 starts from from you know where we
861. 28:55leave it off
862. 28:57in about a week or so um
863. 28:59[Music]
864. 29:00how is 151 i mean there is a hardware
865. 29:02component how is that going online
866. 29:05this semester is it yeah how are you
867. 29:07even working with the online space
868. 29:10students used to be stuck in the lab you
869. 29:12know the digital lab for for hours how
870. 29:14do they do that at home what's what's
871. 29:15equivalent
872. 29:15that's a very good question so that was
873. 29:19also
874. 29:19our major concern especially starting
875. 29:22during the summer when we started
876. 29:23preparing this class
877. 29:24so first we have a wonderful team of
878. 29:27teaching staff
879. 29:28our gsis are wonderful they put in a lot
880. 29:31of time
881. 29:31into this one thing we did do is early
882. 29:34in the summer where like as
883. 29:36i mentioned there is a very strong lab
884. 29:38component so we do want to student still
885. 29:40experience lab
886. 29:42even we are in this virtual world so
887. 29:45we're starting the summer we already
888. 29:46reach out to students for students who
889. 29:48are interested
890. 29:49in fpga lab we basically ship the boards
891. 29:52to them so they actually have
892. 29:54a set of boards that's required for them
893. 29:56to
894. 29:57um to do the labs and our teaching staff
895. 30:00also figure out
896. 30:01basically two ways of programming the
897. 30:04fpga
898. 30:05either through a remote access with our
899. 30:07uh our
900. 30:09instructional servers or a small virtual
901. 30:12machine that they can use
902. 30:13to actually load their program onto
903. 30:15their fpga so instead of relying
904. 30:18completely
905. 30:18on the instruction servers as what we
906. 30:20did before students
907. 30:22can actually create local setup on your
908. 30:25laptop
909. 30:26actually to program the fpgas directly
910. 30:29and second for students who are
911. 30:31concerned about
912. 30:32the fpga lab and also the the physical
913. 30:35uh
914. 30:36physical interactions we also set up a
915. 30:39really smooth asic flow
916. 30:40where all the lab components and also
917. 30:43project components can actually be done
918. 30:46completely remotely
919. 30:47because as long as we provide remote
920. 30:49access so
921. 30:50we do have also a record high number of
922. 30:53students
923. 30:54actually being enrolled in the asic lab
924. 30:56so compared to fpga lab you actually
925. 30:58don't
926. 30:58need anything on your end as long as you
927. 31:01have your laptop you can actually remote
928. 31:03access to our instructional server
929. 31:06you can actually go through the same uh
930. 31:08very important design principles
931. 31:10directly on your end so i think that
932. 31:12also eliminates
933. 31:13a lot of students concern so we are
934. 31:16still halfway uh
935. 31:17less than halfway through the semester
936. 31:18we finished four labs so still two more
937. 31:20labs and the project to go
938. 31:22at least so far so good i think the team
939. 31:26of
940. 31:26gsis are very important with all of us
941. 31:29pouring a lot of time
942. 31:30to interact with the students so to get
943. 31:32questions answered and
944. 31:34holding live labs and also finding
945. 31:38a personalized way to interact with the
946. 31:40students
947. 31:41so so far so good we'll see how this
948. 31:43scales and we did
949. 31:44during the summer process we also
950. 31:47figured out a way actually to use aws
951. 31:49to actually program fpga we didn't end
952. 31:52up pursuing that route in the end but if
953. 31:54the remote
954. 31:55instruction become a new norm i think
955. 31:58that would also be
956. 31:59very useful for classes like 151
957. 32:03that's great thank you for all your hard
958. 32:05work to support our our students who are
959. 32:07all
960. 32:07all around the world taking these
961. 32:08berkeley courses that's thank you so
962. 32:09much for that that work in there
963. 32:11i see two questions in the q a board do
964. 32:13you want to jump on those or should i
965. 32:14challenge
966. 32:15you why didn't you read them yeah yeah
967. 32:16sure i thought they were very good
968. 32:18questions so simon
969. 32:19asks how do you know when to make an
970. 32:22accelerator for specialized applications
971. 32:24uh for example if you may if you make a
972. 32:25chip for speech recognition another chip
973. 32:27for other things
974. 32:28we have too many specialized chips and
975. 32:30they're performing worse than a single
976. 32:31generalized chip what
977. 32:32what's the trade-off there that's a
978. 32:34really good question i think that's also
979. 32:36very
980. 32:36important uh in in basically in the past
981. 32:40ten years when people start approaching
982. 32:42the accelerator design
983. 32:43people have started thinking about
984. 32:45hardware accelerate like hardware
985. 32:47accelerator is a not new idea right i'm
986. 32:49not sure whether your folks cover
987. 32:50floating point in it
988. 32:51yeah floating core processors six eighty
989. 32:54one
990. 32:54sixty eighty 81 i remember that one well
991. 32:57exactly that's basically the first
992. 32:58accelerator
993. 33:01yes exactly that's basically one of the
994. 33:04first accelerators that we built
995. 33:06basically a functionality that may not
996. 33:08be required for all the applications but
997. 33:10may
998. 33:10might be actually very important for a
999. 33:12subset of applications
1000. 33:14we are interested in thinking about
1001. 33:15hardware mechanism for that so
1002. 33:17it has been going on for a while it's
1003. 33:19really interesting how
1004. 33:20machine learning actually completely
1005. 33:22changed the landscape
1006. 33:24and that really actually make it very
1007. 33:27very feasible actually for a lot of
1008. 33:30companies we see all the amount of
1009. 33:31activities going on here
1010. 33:33um to to think about specialized because
1011. 33:36there is a huge market
1012. 33:37a lot of use cases for it so i think to
1013. 33:40simon's question earlier
1014. 33:41uh to how we actually decide to make an
1015. 33:43accelerator for that
1016. 33:45i think partially or unfortunately
1017. 33:48there's largely a
1018. 33:49market decision or application decision
1019. 33:52where we really need to think about what
1020. 33:54applications
1021. 33:55are really important that can actually
1022. 33:57reach to a large amount of market
1023. 33:59so this is more like we talked about
1024. 34:01hardware architecture is really thinking
1025. 34:03about
1026. 34:03architecture is something in between
1027. 34:05think about application and also think
1028. 34:06about
1029. 34:07underlying device technology so for all
1030. 34:10the students who are interested in
1031. 34:11hardware always very important
1032. 34:13to see the application trend and what
1033. 34:16kind of what patterns what kind of
1034. 34:18behaviors or applications
1035. 34:19are getting important so i think that's
1036. 34:22one of the reason
1037. 34:23one of very important factors in
1038. 34:25thinking about accelerator
1039. 34:26at the same time another important
1040. 34:28factor in thinking about what kind of
1041. 34:30component can be accelerated
1042. 34:31also depends on the application patterns
1043. 34:34if
1044. 34:34like it's really nice for machine
1045. 34:36learning in the sense not only it
1046. 34:38reaches so many different applications
1047. 34:40areas but also in thinking about
1048. 34:43accelerator patterns
1049. 34:44it's actually perfectly perfect for
1050. 34:47hardware acceleration and one of the
1051. 34:49reasons to make actually machine
1052. 34:50learning
1053. 34:51so what we used is actually weights
1054. 34:53mapped onto gpu or graphic processing
1055. 34:55units
1056. 34:56it actually reached a pretty impressive
1057. 34:58performance even gpu is actually not
1058. 35:00designed
1059. 35:01for machine learning so we also of
1060. 35:04course i mentioned earlier there's a
1061. 35:05marketing reason for like market reason
1062. 35:07for that what application has a bigger
1063. 35:09reach
1064. 35:10and second there's definitely also
1065. 35:12technology and also
1066. 35:14hardware regions in reason here in the
1067. 35:17sense kind of what kind of application
1068. 35:19behavior
1069. 35:20can be a better fit for hardware
1070. 35:22acceleration
1071. 35:23and certain applications uh especially
1072. 35:26regular applications like machine
1073. 35:28learning have very regular access
1074. 35:29patterns
1075. 35:30and not very straightforward control
1076. 35:32flows and also very regular memory
1077. 35:34accesses
1078. 35:35all those patterns are very important so
1079. 35:37we think about application also need to
1080. 35:39look for
1081. 35:40those patterns here i think there's also
1082. 35:43the domain issue you've got
1083. 35:44like think about alexa alexa has to make
1084. 35:46a determination locally without going to
1085. 35:48the cloud did you say the wake word
1086. 35:50alexa
1087. 35:50but then after you say alexa it takes
1088. 35:53what you'd say and sends it to the cloud
1089. 35:55so
1090. 35:55how much has to be done locally for any
1091. 35:57particular application versus has to
1092. 35:59could be done by you know millions of
1093. 36:01machines waking up doing something like
1094. 36:02a google search or an alexa
1095. 36:04nlp problem and coming back can you wait
1096. 36:06that little half second
1097. 36:07for it to go to the cloud and come back
1098. 36:09the cloud the cloud is amazing so
1099. 36:11how much has to be done locally you know
1100. 36:12at maybe at the edge versus being done
1101. 36:14at the system core level uh
1102. 36:16up up in the cloud second question the
1103. 36:19second question
1104. 36:20uh what aspect of the r d that you do
1105. 36:23applies only to the to the domain of
1106. 36:27this is benjamin by the way to the
1107. 36:29domain of machine learning so of the r d
1108. 36:31you do how much of it is just machine
1109. 36:32learning specific because it's such a
1110. 36:33big problem and you're trying to
1111. 36:35carve some kind of a hardware solution
1112. 36:37to that versus
1113. 36:38what you're what you're doing is
1114. 36:39generalizable for approaching the
1115. 36:41general problems of domain
1116. 36:42specialization so how much is
1117. 36:43machine learning specific and if it
1118. 36:45weren't machine learning nobody can
1119. 36:47use it but how much of it actually could
1120. 36:48be how much of your own research
1121. 36:50development
1122. 36:50can be used in other another
1123. 36:52applications right this is also really
1124. 36:54good question and benjamin
1125. 36:56so i mentioned earlier thinking about
1126. 36:57both applications how we can attribute
1127. 36:59specialized hardware for it
1128. 37:01and second really thinking about system
1129. 37:03integration how we can actually tie
1130. 37:04different accelerators
1131. 37:06together and finally thinking about from
1132. 37:08more design methodologies
1133. 37:10perspective how we can actually improve
1134. 37:12the design process
1135. 37:13i would say maybe the first one is more
1136. 37:16tied to specific algorithm but the other
1137. 37:18two are more generalizable in thinking
1138. 37:20about
1139. 37:21like things like simon mentioned earlier
1140. 37:23we actually have different accelerators
1141. 37:24how we actually use them efficiently and
1142. 37:26how we actually integrate them
1143. 37:28together and finally in terms of design
1144. 37:31process
1145. 37:32whether some of the methodologies or
1146. 37:35design process that we use
1147. 37:36for machine learning accelerators or
1148. 37:39thinking about machine learning
1149. 37:40applications
1150. 37:41can can also be useful for some other
1151. 37:44application domains so on this point i
1152. 37:46also want to i just read the full
1153. 37:48question from simon earlier in terms of
1154. 37:50how this will be actually used for other
1155. 37:52specialized
1156. 37:54node i do want to actually encourage you
1157. 37:56to think more instead of building
1158. 37:58application specific hardware
1159. 38:00or asic what we used to call which has
1160. 38:02been also around for a while
1161. 38:04i think we would we do need to emphasize
1162. 38:07on the domain specific
1163. 38:08part uh instead of application specific
1164. 38:11part in the aspect of domain specific
1165. 38:14we're not only thinking about
1166. 38:15oh we accelerate this one particular
1167. 38:18application that only work
1168. 38:20for this this particular piece of code
1169. 38:22but really thinking about what are the
1170. 38:24common factors
1171. 38:25across this domain and how that can
1172. 38:28actually be
1173. 38:29be uploaded onto hardware in some sense
1174. 38:33you will see
1175. 38:33some of the trends or patterns in the
1176. 38:36software engineering field when we are
1177. 38:38actually building
1178. 38:39a lot of domain specific platforms like
1179. 38:41uh like pytorch tensorflow
1180. 38:43and also other platforming other
1181. 38:45application spaces
1182. 38:46we do see more this kind of domain
1183. 38:48specificity
1184. 38:50coming up not only in hardware but also
1185. 38:52in in software space
1186. 38:53and that domain knowledge should be also
1187. 38:56be transferable
1188. 38:57uh depends on the use cases here
1189. 39:01all right uh thank you sophia thanks for
1190. 39:05uh visiting us over here and i i hear
1191. 39:08that sofia is relatively new so she
1192. 39:11still takes on
1193. 39:12undergraduate
1194. 39:13[Laughter]
1195. 39:22it's way too easy to say yes to
1196. 39:23everything and next thing you know you
1197. 39:25can't
1198. 39:25you can't breathe
1199. 39:29[Laughter]
1200. 39:30and but you guys while still while she's
1201. 39:33still taking students
1202. 39:39thank you thank you so much sophia great
1203. 39:41to see you thank you folks
1204. 39:43thanks again everybody thanks again
1205. 39:44thank you wonderful
1206. 39:46wonderful that's great feel free to head
1207. 39:48out if you'd like to
1208. 39:49yeah i'll tell you i'll take it over
1209. 39:50perfect perfect all right continue
1210. 39:53all right here we go so i wanted to um
1211. 39:56i wanted to show you if i yeah give me a
1212. 39:59pin and then i'll i'll jump in
1213. 40:00here we go okay now it's this spotlight
1214. 40:04actually the pinning is just for your
1215. 40:05own thing it's the spotlight it's the
1216. 40:07guy
1217. 40:08i there we go okay so
1218. 40:11we did this we did this i wanted to show
1219. 40:13i wanted to show you this
1220. 40:15which is we didn't get to show this
1221. 40:16interesting i showed this in cs10
1222. 40:18this is this just came out i think this
1223. 40:20is like last thursday or friday
1224. 40:22a drone that flies inside your home
1225. 40:24think about privacy implications we try
1226. 40:25to do computing in the news
1227. 40:27um this is i guess the idea here
1228. 40:30just to set this up set the context if
1229. 40:33you didn't have a camera
1230. 40:35on like every window if you only had
1231. 40:36like one main camera for your home
1232. 40:38security system
1233. 40:39only facing out or facing your door but
1234. 40:41you had all the windows of things and
1235. 40:42you heard something
1236. 40:43there's something that happened you know
1237. 40:44like somebody rattles a window you
1238. 40:46didn't think about somebody would break
1239. 40:47into your side window
1240. 40:48let me just show you this all the only
1241. 40:49sound of this is like some some so
1242. 40:53there's nothing there's nobody speaks of
1243. 40:54this so it's a 30-second video let me
1244. 40:55just show you this really fast
1245. 40:58and here so here's this the drone
1246. 41:02wakes up you get an alert something's in
1247. 41:04your house
1248. 41:05and then you can control the drone the
1249. 41:07drone tries to find toward where the
1250. 41:08sound was
1251. 41:10but i think you can also control the
1252. 41:11drone as well and by the way i was told
1253. 41:13that
1254. 41:14not everything you're seeing in this
1255. 41:16video is a real
1256. 41:17shot some of it's simulated so
1257. 41:18somebody's just walking around
1258. 41:19pretending to be a drone
1259. 41:20they haven't perfected it yet but the
1260. 41:22idea is you'll have this drone that you
1261. 41:24can either control to see if
1262. 41:26well there's a sound downstairs and you
1263. 41:27get a notification for sound downstairs
1264. 41:28from the security system
1265. 41:29uh and then you'll you'll figure out
1266. 41:32you'll you'll figure out uh
1267. 41:34what to do whether you can fly yourself
1268. 41:36or not so i found it
1269. 41:38fascinating that anybody would buy this
1270. 41:41this is just me because uh they've
1271. 41:44already shown that you can hack
1272. 41:46the uh i think i forget what it
1273. 41:50what particular system was an amazon or
1274. 41:52a ring or who was driving this but
1275. 41:53they've already shown you can hack
1276. 41:54the the cameras that are at your front
1277. 41:58door
1278. 41:59uh they're so it's very easy to get
1279. 42:01access to that
1280. 42:02video data uh and so you can imagine if
1281. 42:05that's easy to do that then you can get
1282. 42:06access to this drone and fly it
1283. 42:08around the house in the middle of night
1284. 42:09including you know when you don't want
1285. 42:10to be
1286. 42:11when you're not controlling it so i
1287. 42:13would like i'm the last person so i
1288. 42:15recommend
1289. 42:16think twice before you number one put
1290. 42:19your front door lock and there's a whole
1291. 42:20digital front door that's the last thing
1292. 42:22i would put as digital things can be
1293. 42:23hacked like crazy don't put a front door
1294. 42:24lock that's just me
1295. 42:26and i wouldn't put a drone that could be
1296. 42:27floating around and watch me when i'm in
1297. 42:29the shower or something so
1298. 42:30don't do either of those things all
1299. 42:31right um
1300. 42:33yeah an interesting thing about this i
1301. 42:35bought a light bulb
1302. 42:36about uh i don't know a year and a half
1303. 42:38ago that is also
1304. 42:40the built-in speaker oh look that
1305. 42:43you know i i i tend to probe into these
1306. 42:46things before
1307. 42:47you know actually installing them and it
1308. 42:49saved a password is plain text
1309. 42:52so anybody who has
1310. 42:53[Laughter]
1311. 42:56yeah sometimes the security system isn't
1312. 42:58done so well break into that but they
1313. 42:59get
1314. 43:00password to your wi-fi uh they get
1315. 43:02access to everything in your house
1316. 43:04for that so yeah watch out for god watch
1317. 43:06out folks
1318. 43:07all right we do want to share we do want
1319. 43:09to share uh where we are on our schedule
1320. 43:11and then also shares
1321. 43:12any announcements we have so here we are
1322. 43:14looking at the schedule i just finished
1323. 43:16this whole weekend i was
1324. 43:17locked in my basement making my movies
1325. 43:19and then sitting in my computer doing
1326. 43:20all the post-processing so hopefully
1327. 43:22you uh get access to them i think we
1328. 43:24posted them throughout already
1329. 43:25so sds state logic and logic blocks are
1330. 43:28all out there all the clicker questions
1331. 43:30are out there we tried to make quicker
1332. 43:31questions that would be
1333. 43:32you know really not just pedantic did
1334. 43:33you watch minute 35 what color shirt was
1335. 43:35he wearing but
1336. 43:36you know what do you understand the
1337. 43:37principles behind it so i actually try
1338. 43:39to pull some
1339. 43:39old exam questions and try to throw them
1340. 43:41in here so actually some of these are
1341. 43:42really good questions in here so i thank
1342. 43:43steven
1343. 43:44for some of the questions you
1344. 43:44contributed and for other folks so
1345. 43:47we're deep in the digital logic world i
1346. 43:49love it there's a whole new field by the
1347. 43:51way this is a great i want to make a
1348. 43:52point people know this
1349. 43:53if you're behind in 61c like i'll never
1350. 43:56catch up because it's all continuous
1351. 43:57you can jump in if you're behind
1352. 43:59watching the lectures you can jump ahead
1353. 44:01to be caught up in this
1354. 44:02module this is really important this is
1355. 44:04why borah i think here has shown
1356. 44:05different colors in the in the graph
1357. 44:07that's really important because if
1358. 44:09you're all let's say you're two weeks
1359. 44:10behind this
1360. 44:11by the way happened two years ago and a
1361. 44:13year ago it's always been the case that
1362. 44:14some students get behind and
1363. 44:16can catch up and it's hard the workload
1364. 44:17is hard we'll talk about workload in a
1365. 44:18second
1366. 44:19but one of the things that we want to
1367. 44:20make sure we let you know
1368. 44:22is that what we've got is
1369. 44:26and let's let's make sure we are clear
1370. 44:27about this these modules are independent
1371. 44:29so yes they are
1372. 44:31softly relevant and it's great if you
1373. 44:32saw the stuff before but you can restart
1374. 44:34on a new module and become comfortable
1375. 44:37with this so
1376. 44:389 28 uh this is monday sds
1377. 44:41if you were behind just just catch up
1378. 44:43for this section and then you can you
1379. 44:45know
1380. 44:45just one correction i didn't update the
1381. 44:47date oh yeah sorry
1382. 44:49yeah yeah so the date okay thank you
1383. 44:50thank you for that yeah these are the
1384. 44:51older dates from
1385. 44:52from before yeah don't be added a couple
1386. 44:55of days
1387. 44:56don't panic we'll fix those um uh so
1388. 44:59let's just make sure you guys know what
1389. 45:00the section is and then right after this
1390. 45:02this is critical
1391. 45:03in these two weeks you will know the
1392. 45:04entire bottom side but all the hardware
1393. 45:06level that you're going to
1394. 45:07need to know below that which is great
1395. 45:08we get into caching after that so that's
1396. 45:10really great
1397. 45:11um this all misinformation is all needed
1398. 45:14for and you see the colors board did a
1399. 45:16nice job of showing the color connection
1400. 45:17here so
1401. 45:18the yellow stuff is needed for homework
1402. 45:19for 2a the lab this week some time to
1403. 45:22catch up which is great
1404. 45:23and work on your project obviously um
1405. 45:25homework five and lab fiber this week
1406. 45:27um this is for sorry next week next week
1407. 45:29orange is lab five
1408. 45:31homework 5 and lab 5 next week and the
1409. 45:33red stuff is for project 3 which is
1410. 45:35going to be
1411. 45:35really fun and many people by the way
1412. 45:37think that project 3
1413. 45:38the the processor design is the most fun
1414. 45:40project they've taken in this course and
1415. 45:43in many other courses i still hear from
1416. 45:44people years later i still remember the
1417. 45:4561c processor design
1418. 45:47so make sure uh you you focus on that if
1419. 45:49you're thinking about ever designing
1420. 45:50your processor and hardware
1421. 45:52that's for some people that's the most
1422. 45:53fun part of this material so
1423. 45:55jump on that when it comes next week
1424. 45:56excited about that so so
1425. 45:58just a few few notes about that exactly
1426. 46:01dan said
1427. 46:02this is a good week to try to catch up
1428. 46:04um we you know there are many things
1429. 46:06that are independent here
1430. 46:07but the red part is tied to the orange
1431. 46:12and the yellow
1432. 46:13um so the the yellow the the orange part
1433. 46:17is relatively independent
1434. 46:18you you you can just come from outer
1435. 46:21space
1436. 46:22and listen to that and you will pick up
1437. 46:23the the digital systems
1438. 46:25over uh you know over the next week
1439. 46:28that's true even if you didn't know any
1440. 46:30risk five or c you can just jump right
1441. 46:32into today's lecture that's exactly
1442. 46:33right
1443. 46:34now when we get to the red part and
1444. 46:36that's gonna go for five lectures
1445. 46:39in a row these five lectures are going
1446. 46:41to
1447. 46:42um uh
1448. 46:46uh they're basically a complex digital
1449. 46:48system a really complex digital system
1450. 46:50that we're going to build based on the
1451. 46:52principles that we have
1452. 46:53um in in orange in uh sds
1453. 46:56module based on the specification that
1454. 47:00we have uh
1455. 47:01developed in the yellow module over
1456. 47:04there over these
1457. 47:05uh seven lectures or so so we will
1458. 47:07basically take
1459. 47:08all those instructions and implement
1460. 47:10them using
1461. 47:12the principles that we have learned it's
1462. 47:14really fun
1463. 47:15uh uh interesting thing uh when we
1464. 47:18uh uh with uh sofia we
1465. 47:21you guys those that are going 151 will
1466. 47:24actually do it
1467. 47:26um again but now in the language core
1468. 47:28verilog and realize it in something that
1469. 47:30will look like a real
1470. 47:32asic or or put it into an fpga
1471. 47:35and then pass risk 5 compliance test
1472. 47:39this year we have changed the project a
1473. 47:41bit
1474. 47:42so you pass some of those tests so there
1475. 47:45are
1476. 47:45not very many hidden tests but we can
1477. 47:47add more hidden tests
1478. 47:49if you want to um so
1479. 47:52it's quite exciting i mean you do get
1480. 47:54the functional core
1481. 47:56this time around that can run assembly
1482. 47:59you can even run c compile c
1483. 48:01next time around you can actually build
1484. 48:03it if you take 151
1485. 48:06back to them no it's great i mean but
1486. 48:08what's really fun about this this this
1487. 48:09series of lectures i think
1488. 48:11you know the orange to set it up the red
1489. 48:12to actually take it home is we're going
1490. 48:14to build a working risk five machine
1491. 48:15this is amazing i mean we're not we're
1492. 48:17not physically building it but we're
1493. 48:18going to build a machine that if you
1494. 48:19could send it to somebody to do this
1495. 48:20this will actually
1496. 48:21run the machine code and i mentioned
1497. 48:23this in the lectures run the machine
1498. 48:25code you've compiled to someone to link
1499. 48:26down to so all that
1500. 48:27ones and zeros and the machine code
1501. 48:29we're going to build a machine to do
1502. 48:30that and that's what's going to happen
1503. 48:31in the next
1504. 48:32several lectures after this it's very
1505. 48:33exciting and by the way that borah i
1506. 48:34think it's
1507. 48:35six lectures so it's it's a lot of time
1508. 48:37in your in your basement studio
1509. 48:39not five lectures at six and then we're
1510. 48:41not gonna build it's like three lectures
1511. 48:42to build the machine here's the thing
1512. 48:44and then three likes just to make it
1513. 48:45faster
1514. 48:46and so using some principles that i'll
1515. 48:47teach you in this orange section called
1516. 48:49pipelining so that's the exciting piece
1517. 48:50of that
1518. 48:50so think about that but even if you
1519. 48:52didn't make it fast if you skipped the
1520. 48:53last three lectures you'd serve a
1521. 48:54working machine just a little slower
1522. 48:55so you know what you can make this
1523. 48:56faster and your three lectures to make
1524. 48:58it faster because
1525. 48:59this course is also about performance
1526. 49:01which is exciting all right
1527. 49:03here we go now we got some i got third
1528. 49:04perfect three minutes for two
1529. 49:05announcements announcements
1530. 49:07announcements uh we just released a
1531. 49:10workload and wellness survey
1532. 49:12just went we mentioned it last week but
1533. 49:13we finally finished it today and we
1534. 49:15launched it today and
1535. 49:16already the numbers are pouring in
1536. 49:18already telling us many things we knew
1537. 49:20uh in terms of how many so please do
1538. 49:22fill this out it's on piazza
1539. 49:24um and i'm already seeing numbers of i
1540. 49:26can just give you a mostly i'm
1541. 49:28seeing i don't know you shouldn't you
1542. 49:30shouldn't uh
1543. 49:31sully a survey by giving some results
1544. 49:33but i'm seeing that
1545. 49:34it's great to see these numbers um and
1546. 49:37it's
1547. 49:38not i'm not pleased with the numbers i'm
1548. 49:39hearing people are really struggling out
1549. 49:41there
1550. 49:41in terms of emotionally uh and
1551. 49:43exhaustion wise
1552. 49:44and the number of hours people putting
1553. 49:45in we've got to also take a really
1554. 49:46deeper dive
1555. 49:48um on on our workload in 621c it does by
1556. 49:50the way
1557. 49:52the goal of 61c is to not be more than
1558. 49:5512 hours a week times 15 weeks and
1559. 49:57sometimes what happens is
1560. 49:58it's more average your average per week
1561. 50:02is higher in the first bit and then
1562. 50:03lower later so if it if at all
1563. 50:05we tell our tases sometimes around
1564. 50:06midterm week it's going to be a crazy
1565. 50:07amount of hours you put in
1566. 50:08but as long as it's average to what
1567. 50:10you're being paid for that's reasonable
1568. 50:11so
1569. 50:12these numbers are going to be high but
1570. 50:13hopefully when we do their midterm
1571. 50:14survey and then our final survey
1572. 50:15it'll all end up showing uh showing that
1573. 50:18we kind of
1574. 50:19averages out to 12 hours a week if not
1575. 50:20we really need to go back to the drawing
1576. 50:21board and figure out
1577. 50:22how we can make this lighter a lighter
1578. 50:24weight thing and thank you so much for
1579. 50:26all the suggestions of how to do that so
1580. 50:27there's a there's a point to ask the in
1581. 50:29the survey we ask you any policy changes
1582. 50:31we've
1583. 50:31you know you remember that last week we
1584. 50:32made a ton of new policy changes
1585. 50:34hopefully try to try to address this
1586. 50:35more slip days more more flexibility
1587. 50:37here more brace for project two
1588. 50:39etc we'll continue to look at policies
1589. 50:41we can do to help with the workload and
1590. 50:42wellness issues we can
1591. 50:44that's part one part two uh quest retake
1592. 50:47in fact some folks have said
1593. 50:48i don't have time for a quest we take
1594. 50:50i'm just being barraged with workload i
1595. 50:52don't have time to take more
1596. 50:54hours to study for it and to take more
1597. 50:55hours through the retake the idea you're
1598. 50:56not supposed to be studying for it
1599. 50:58you're supposed to just do it again
1600. 50:59so don't count that into what i was
1601. 51:00receiving just jump you back in there
1602. 51:03we're going to try to reduce the number
1603. 51:04i mentioned last week that people we're
1604. 51:05going to need to take that
1605. 51:07and what we need to do first of all my
1606. 51:09taste tell me there's a there's an order
1607. 51:10to this
1608. 51:11there's logic to this which is process
1609. 51:13all the regrades
1610. 51:14now the students know what their final
1611. 51:15grade is and now you know do they need a
1612. 51:17retake or not
1613. 51:18um so we can know who not to who not to
1614. 51:19bother again with all this so we're
1615. 51:21going to try to slog
1616. 51:22through i think borah the number i heard
1617. 51:23was 350
1618. 51:25regrade requests and we've got two tas
1619. 51:28processing
1620. 51:29i mean two star tas but still i mean we
1621. 51:31you know we didn't we didn't
1622. 51:33budget for 350 regrade requests usually
1623. 51:35that's not what we get we usually get uh
1624. 51:3650 100. not 350 any which in which each
1625. 51:39of the 350 requires us to look at their
1626. 51:41code figure out
1627. 51:42why does it not do something and how can
1628. 51:43we give you points back so
1629. 51:45the perspective that we have from the
1630. 51:46top is how can we get you points back so
1631. 51:49we're going to spend some time
1632. 51:50thinking about this we're not going to
1633. 51:51both slam you on a quest and then not
1634. 51:52give you points back
1635. 51:53but we're going to see how we can do
1636. 51:55this and how we can you know maybe fix a
1637. 51:56bug and then does it
1638. 51:57does it pass some tests okay get those
1639. 51:59points maybe get ding for the bug we had
1640. 52:01to fix that kind of thing
1641. 52:02so but hopefully you gave us code that
1642. 52:03worked i mean hopefully they're just
1643. 52:04giving garbage code
1644. 52:06and it would you know it didn't work at
1645. 52:07all hopefully it worked on your system
1646. 52:09if it worked on your system but it
1647. 52:10doesn't work on holes and that's
1648. 52:11there's some mismatch we're going to try
1649. 52:13to get down get down to the details of
1650. 52:15that
1651. 52:16and again we're going to try to endeavor
1652. 52:18so because they were working on that
1653. 52:19that's probably
1654. 52:20all this week is working through the
1655. 52:21regrades slogging through that then
1656. 52:23doing an analysis of who
1657. 52:24didn't have a video that worked and
1658. 52:25figure out who needs the exam blah blah
1659. 52:27blah that's probably
1660. 52:28so all this is probably going to happen
1661. 52:29next week so again we're just pushing
1662. 52:31this off
1663. 52:31so don't worry about it for now survive
1664. 52:33your project two work on that for now
1665. 52:35and then we'll probably do this so again
1666. 52:36i hope to have better a better answer
1667. 52:38for you next next week once our ptas
1668. 52:39have told me that all the regrades are
1669. 52:41done but
1670. 52:41we want to give a dude to do justice to
1671. 52:43all the regrets but it does take time
1672. 52:44and so that
1673. 52:45please please do uh please do recognize
1674. 52:47that we're working as hard as we can
1675. 52:51and they're very informative yeah very
1676. 52:54thank you and please pat tell your tell
1677. 52:55other classmates i'm only looking here
1678. 52:57at 230
1679. 52:58there's more than 1200 in this class so
1680. 53:00please please do
1681. 53:02make sure all your classmates and by the
1682. 53:03way i believe this is due in two days
1683. 53:05it's due on wednesday
1684. 53:06evening so please do serve that we're
1685. 53:07going to read the whole survey and we'll
1686. 53:09even share
1687. 53:09we'll share the results next week
1688. 53:11actually to have a conversation about
1689. 53:13what policies will change
1690. 53:14uh but thank you all for that
1691. 53:15information this survey isn't much you
1692. 53:17just a couple clicks a couple clicks and
1693. 53:18feel free to
1694. 53:19you can type a lot if you wanted to
1695. 53:20we'll have some open fields for you to
1696. 53:22type some stuff but if you don't want if
1697. 53:23you don't have a lot of time just click
1698. 53:24the buttons and then
1699. 53:25and then give it to us but please do
1700. 53:26give us some feedback if we can for that
1701. 53:27thank you for that
1702. 53:29i think i'm done borah i'm done with
1703. 53:30this we're already over a minute for
1704. 53:31about a minute and a half so
1705. 53:32i want to thank everybody let's thank
1706. 53:34sophia for her guest uh appearance that
1707. 53:36was wonderful to have her here
1708. 53:38again thank you borah for for the for
1709. 53:40the week for the for all the good work
1710. 53:41you've done and
1711. 53:42good luck with seven seven straight
1712. 53:44lectures in your basement
1713. 53:45i know i know how long it took me to set
1714. 53:47up for those four lectures there it took
1715. 53:48me
1716. 53:49a full week of just editing slides so i
1717. 53:51i wish you good luck
1718. 53:52i'm editing so i'm still editing slides
1719. 53:55i didn't like
1720. 53:56the the technical quality of the or the
1721. 53:59quality of the last time yeah
1722. 54:03yeah it takes a long time appreciate
1723. 54:05this takes forever so we'll try to
1724. 54:07together
1725. 54:08yeah what am i seeing here is what what
1726. 54:10am i learning from the survey is that
1727. 54:12the people on the average
1728. 54:13um are spending uh maybe two and a half
1729. 54:16hours more than what we expected which
1730. 54:18yeah that's something too yeah more yeah
1731. 54:21but the tail is
1732. 54:22huge we have a really really long tail
1733. 54:24and we need to figure out how to
1734. 54:26get this uh um get some support
1735. 54:29guys how do you get support for those
1736. 54:30people who i mean a lot of times you
1737. 54:32could be stuck for 10 hours on the same
1738. 54:33bug i mean that's just that's a wasted
1739. 54:3410 hour that's not
1740. 54:35that's not a useful 10 hours so that's
1741. 54:37it there is a question about whether we
1742. 54:38have office hours this weekend i think
1743. 54:40both of us have office hours
1744. 54:41mine is on on wednesday and yours i
1745. 54:43think board is on
1746. 54:44thursday or friday friday yeah friday in
1747. 54:47this in this lecture and we may
1748. 54:49already turn this this time the way how
1749. 54:51we do this we may start running
1750. 54:53more of like reviews and so on but this
1751. 54:55week are just regular drop-in
1752. 54:57for the office hour yeah and i've been
1753. 54:59mostly i've been
1754. 55:00taping mine and i've been mostly doing
1755. 55:02kind of exam reviews so every week
1756. 55:03so this week we'll actually take a look
1757. 55:05at some sds stuff uh and maybe even some
1758. 55:07from risk five we didn't finish last
1759. 55:08week so i've been trying to go through
1760. 55:09old exams through that so
1761. 55:11come to my office hours if not i'll try
1762. 55:12to upload those videos so people have
1763. 55:13those as reviews
1764. 55:14wonderful thank you borah for a good
1765. 55:16week thanks everybody for coming in
1766. 55:17thanks joining folks we'll see you later
1767. 55:18folks
1768. 55:19take care bye okay bye everybody