Cerebras reveals details of the world's largest chip: 2.6 trillion transistors, 850,000...

midian182

Posts: 9,738   +121
Staff member
What just happened? After unveiling a chip last year that was both the world’s largest and the first to feature over a trillion transistors, artificial intelligence company Cerebras Systems has gone one better. The follow-up, which is built on TSMC’s 7nm process, has an incredible 850,000 cores and 2.6 trillion transistors.

At Hot Chips 2019, Cerebras showed off the Wafer Scale Engine (WSE)—the world's largest chip. At 8 inches by 9 inches, it’s bigger than a standard iPad. Built on the 16nmFF process, WSE boasts 1.2 trillion transistors, 400,000 cores, 18GB of on-chip memory, 9 PB/s total memory bandwidth, and 100Pb/s of total fabric bandwidth. It also consumes ~15kW of power.

At this year’s event, Cerebras teased the second generation WSE. While the core and transistor count is more than double that of its predecessor, the company kept quiet about the bandwidth, power consumption, and how much SRAM memory is built into the latest model—but you can expect improvements in these areas, too.

To create the Wafer Scale Engines, Cerebras uses the entire wafer as a single, massive chip, rather than the usual method of etching individual chips onto a single wafer. It stitches together the dies on the wafer with a communication fabric that allows them to work as a single unit, bypassing the reticle size limitations of a foundry.

For comparison, the world’s largest GPU, Nvidia’s A100, measures 826mm2, while the WSE is 46,225 square millimeters. And team green’s product ‘only’ has 54.2 billion transistors.

Cerebras said it would reveal more details about the second-gen WSE “in the coming months,” adding that the chip is currently "running in our labs."

Permalink to story.

 
Wait..is that 15 KILOWATTS of power consumption? Continuously?? I don't even know how something that size could handle that much juice if I'm reading it right.
 
Wait..is that 15 KILOWATTS of power consumption? Continuously?? I don't even know how something that size could handle that much juice if I'm reading it right.
Conventional dies are fed power through connections along the edge. WSE is different by supplying current through an array of pins that are perpendicular to the entire face of the die - in others, each section gets its own direct mini-supply.

The entire system actually draws up to 20 kW - the extra 5 being for the I/O and cooling systems.
 
... What is this cpu used for ? What's the difference in usage with super computer ?
It's exclusively for deep learning; solving complex neural networks. A typical super computer will be more generalised.

By having everything (processor cores, memory, and the connections between the two) all on the same die, the potential gains in latency reduction and bandwidth are huge.

It's also going to be relatively small, compared to a super compute cluster, because it's a single device - so while the cost and power consumption are equivalent, there's a clear advantage in terms of space required.
 
That is a LOT of energy to dissipate. Take a normal chip (i9) about 50x50mm square at ~300W. And look at some of the huge heatsinks. Most have heatsinks at least 4 times the footprint (100x100).
If you work out the power per cm2 (very roughly)
Cerebras : 20x20=400cm2 15kw/400 = 37.5W/cm2
Intel i9 : 5*5=25cm2 300/25 = 12w/cm2.
(I know the actual silicon in an i9 is smaller than the package, but the package is what is getting hot)

It is still a tiny total compared to a car engine with 100s of kw, but that is a huge lump of metal with the heat generated over a big volume around each piston, a lot of airflow and oil and water cooling etc... Not a flat wafer in a server cabinet! I'm almost more interested in the cooling than the processor :)
 
Yes very powerful but can it play Total War Hammer ll at 8k 10,000fps, now that would be some super lightening fast powerful chip, just kidding
 
Yes very powerful but can it play Total War Hammer ll at 8k 10,000fps, now that would be some super lightening fast powerful chip, just kidding
Can it render Total War? Not even a little. But it can probably play Total War - probably even beat you at the game, too. As long as you give it enough training time.
 
Well, if it doesn't sell as a processor it would make a very unique dinner plate ..... especially for the parents of nerds .....
 
I'm surprised that they didn't just use some freezer
The problem with refrigerant gases as coolants is that (a) they don’t have anything like the heat capacity of water and (b) the systems using them typically produce a fair bit of condensation, as the low temperatures condense water out of the surrounding air. All super computers are water cooled these days.
 
Wait..is that 15 KILOWATTS of power consumption? Continuously?? I don't even know how something that size could handle that much juice if I'm reading it right.
That is a LOT of energy to dissipate. Take a normal chip (i9) about 50x50mm square at ~300W. And look at some of the huge heatsinks. Most have heatsinks at least 4 times the footprint (100x100).
If you work out the power per cm2 (very roughly)
Cerebras : 20x20=400cm2 15kw/400 = 37.5W/cm2
Intel i9 : 5*5=25cm2 300/25 = 12w/cm2.
(I know the actual silicon in an i9 is smaller than the package, but the package is what is getting hot)

It is still a tiny total compared to a car engine with 100s of kw, but that is a huge lump of metal with the heat generated over a big volume around each piston, a lot of airflow and oil and water cooling etc... Not a flat wafer in a server cabinet! I'm almost more interested in the cooling than the processor :)
So this is actually wrong, the silicon package is what you should consider as getting hot. In fact you should mostly consider the cores as what gets hot. Think about hotspot temps on AMD gpus, they can often be 20-40c higher than edge temps. Similarly the temps on your heat spreader will be much lower (particularly the parts not directly above the cores) than those cores. The reason this thing can be cooled is exactly BECAUSE its power consumption per area is fairly low, as well as it being an extremely parallel design with power consumption being distributed very evenly when compared with f. Ex a modern CPU (think about it, your one zen chiplets can run 125w for an 8 core chiplet if you OC, thats almost 150w/cm2 with a chiplet being 88mm2. Your Io consumption is a fraction of that but the die is much bigger).
 
Back