Trillion transistor CPU chip has 400K cores and is 8 inches on a side!

[qubit]

I've previously wondered if a whole wafer could be used to build a CPU and it turns out that it can - and the efficiency gains are enormous! Check out the video.

I'd like to know how they cool this monster and its clock speed, but the video doesn't explain that.

 

[Raevenlord]

There are some articles on the Wafer Scale engine already here in TPU =)

Cerebras Updates Wafer Scale Engine on 7 nm - 2.6 Trillion Transistors, 40 GB Onboard SRAM, 850,000 Cores, 12" Wafer

Cerebras has announced the successor to their record-breaking Wafer Scale Engine. The newly re-engineered Wafer Scale Engine 2 has been redesigned for TSMC's 7 nm manufacturing process - a severe improvement over the original's 16 nm. That Cerebras has moved on to TSMC's 7 nm for this giant...

www.techpowerup.com

And the 400K cores are old news - their redesign features 850K cores

 

[qubit]

Thanks, I wasn’t aware of the TPU article. Still, the video gives a different take on it that's worth watching. And yes, at the end of the video they did mention a new chip with double the transistors.

Imagine the power of an x86 CPU built like this.

EDIT: I've just realised it's your article.

 

[oobymach]

15 kW (chip) power consumption

How they would cool it, hell how would they POWER it? It (cerebras) requires 15kw or 15,000 watts....that's more than most peoples houses consume.

 

[Keullo-e]

How they would cool it, hell how would they POWER it? It (cerebras) requires 15kw or 15,000 watts....that's more than most peoples houses consume.

Well, ENIAC in the 1940s consumed about 150kW so a high consumption in a specific usage isn't anything new.

ENIAC - Wikipedia

en.wikipedia.org

It weighed more than 30 short tons (27 t), was roughly 2.4 m × 0.9 m × 30 m (8 ft × 3 ft × 98 ft) in size, occupied 167 m2 (1,800 sq ft) and consumed 150 kW of electricity.

 

[Vya Domus]

I am afraid this is far from impressive. This is NOT a CPU and should never be compared to one or a GPU for that matter. It's a an ML inference accelerator, without getting too much into details it's basically designed to perform a handful of operations and that's it.

There's a ton of inference accelerators out there, everyone is making them because it's easy. When it comes to training performance GPUs like A100 still remain unmatched.

How they would cool it, hell how would they POWER it? It (cerebras) requires 15kw or 15,000 watts....that's more than most peoples houses consume.

If you look at it in terms of power/area there is nothing special about it.