facebook rss twitter

Cerebras teases 850k core 2nd-generation wafer scale engine

by Mark Tyson on 19 August 2020, 12:10

Tags: TSMC

Quick Link: HEXUS.net/qaenqr

Add to My Vault: x

Last year at the Hot Chips Symposium, Cerebras caused quite a stir by taking the wraps off the "world's largest computer chip". The so called Wafer Scale Engine (WSE) packed 400,000 cores in its 21.5cm sq (8.5inch sq), making it about 56x larger than Nvidia's Tesla V100 GPU. It isn't surprising, as the first gen WSE consisted of 1.2 trillion transistors. Furthermore, the WSE had 3,000 times more on-chip memory and 10,000 times the bandwidth of Nvidia's Tesla V100.

This year at Hot Chips, Cerebras mainly talked about software for its AI processor and the CS-1 15U system based around the processor. Due to some time-scale slippage it couldn't launch the second gen WSE. However, it confirmed the new processor is running in its labs and at the end of the slide deck the firm teased the upcoming behemoth.

As you can see above, the WSE gen 2 packs in a huge amount of transistors and cores in the same wafer size thanks to the use of the advanced 7nm processor from TSMC. I've made a little comparison table below to show the differences:

 

WSE gen 1

WSE gen 2

Transistors

1.2 trillion

2.6 trillion

AI cores

400,000

850,000

Chip area

46,225mm²

Not confirmed

TSMC node

16nm

7nm

SRAM built-in

18GB

Not known

 

Since last year Cerebras has sold two systems based upon the WSE gen 1 to the Pittsburgh Supercomputing Center, netting US$5 million. Pictures of these systems in isolation might make you think they are similar in scale to an ATX desktop PC tower, but Anandtech's pic from June shows a CS-1 system next to an ordinary looking fellow at a trade show (below right).

More details about the WSE gen 2 are going to arrive "in coming months."

HEXUS Forums :: 6 Comments

Login with Forum Account

Don't have an account? Register today!
Posted by Tabbykatze - Wed 19 Aug 2020 12:34
This is crazy because it's a testament to TSMCs 7nm process that its defect rate is so low that Wafer-Scale engines can be build off its process. Amazing
Posted by BigBANGerZ - Wed 19 Aug 2020 17:09
Kind of puts to bed the claim you hear from the pro-Nvidia crowd about yields and die sizes being a justifiable reason to charge consumers £1.2k for a GPU. Sure its not easy but WSE is the cutting edge and limits not a 600mm2 die
Posted by nobodyspecial - Thu 20 Aug 2020 00:38
BigBANGerZ
Kind of puts to bed the claim you hear from the pro-Nvidia crowd about yields and die sizes being a justifiable reason to charge consumers £1.2k for a GPU. Sure its not easy but WSE is the cutting edge and limits not a 600mm2 die
ROFLMAO. $1200 vs. 2.5m each unit? No, sort of says EXACTLY what NV says about die sizes. A huge die sells in a machine for 2.5M dollars.

If you divide 46000 mm^2 by 800 (roughly TSMC reticle) you get about 58. 58x1200=69,600, so charging around 2.5M shows as you grow that chip size, you massively charge more money. The math just doesn't lie here, it is what it is. IF this was NV chopping it up into 58 800mm^2 dies, they'd charge around 70K for them, and that is the ENTIRE CARD, not just 58 chips. So is NV screwing you at 70k for the size of the entire wafer basically, or is this company at 2.5M for them all combined into ONE chip/wafer (whatever)? Probably NEITHER are screwing you, it is what is required to make profit. I doubt they made money here. It sounds like rev1 sold 2 units for 5mil. With all the design etc into making this box work, probably a mass loss to prove it could be done, but gen2 likely makes money after lessons learned, and people knowing it works. More sales will come, volume will help, die shrink, etc. From 16nm to 7nm means massive perf gains, power drops, etc, so again, I think they will make money this round if they didn't before.

That said, NV isn't screwing anyone or products would be left on shelf. If your price is too high, products don't move. So YOU are the problem here, not them. Improve your income so a $1000 gpu means nothing to you. I can buy one every year if desired easily. I choose not too (not willing to pay rich tax for 10% perf, I buy a few down), but because it just isn't worth it to me. I'm not rich enough to not care at all, and have better things to do with my money until a die shrink etc gets what I want at a price I'll pay. I could improve my income too, until I don't care at all (trying..LOL). That is what you should be doing your whole life. Add value to YOU, by constantly upgrading your skills, or make money ON your money (the best way). Then you do nothing but monitor it, and research to make more on it. Stocks! They've paid for my parts for decades. I don't pay for PC parts out of checks from work, I buy a part when stocks etc make it free.

You are trying to make the world fit in your income, when you should be trying to get your income to fit the world you want to provide for yourself. IE, if a you want a lambo, don't expect to get it working at walmart. You will have to improve your life enough to afford the car, the dang insurance, repairs, etc. You can wait forever for something to come to your budget, or grow your income so you can finally afford it. The choice is yours, or just keep complaining about prices I guess.. ;)
Posted by BigBANGerZ - Thu 20 Aug 2020 13:35
nobodyspecial
So YOU are the problem here, not them. Improve your income so a $1000 gpu means nothing to you. I can buy one every year if desired easily.

Wow, just wow.
Posted by Iota - Thu 20 Aug 2020 19:29
Tabbykatze
This is crazy because it's a testament to TSMCs 7nm process that its defect rate is so low that Wafer-Scale engines can be build off its process. Amazing

I have to wonder what the actual defect rate is though, it can't be cheap to produce an entire WSE, unless they designed it to include potential redundancies due to manufacturing defects.

SEE NEWER »