The last few years has seen a glut of processors enter the market with the sole purpose of accelerating artificial intelligence and machine learning workloads. Due to the different types of machine learning algorithms possible, these processors are often focused on a few key areas, but one thing limits them all – how big you can make the processor. Two years ago Cerebras unveiled a revolution in silicon design: a processor as big as your head, using as much area on a 12-inch wafer as a rectangular design would allow, built on 16nm, focused on both AI as well as HPC workloads. Today the company is launching its second generation product, built on TSMC 7nm, with more than double the cores and more than double of everything.

Second Generation Wafer Scale Engine

The new processor from Cerebras builds on the first by moving to TSMC’s N7 process. This allows the logic to scale down, as well as to some extent the SRAMs, and now the new chip has 850,000 AI cores on board. Basically almost everything about the new chip is over 2x:

Cerebras Wafer Scale
AnandTech Wafer Scale
Engine Gen1
Wafer Scale
Engine Gen2
Increase
AI Cores 400,000 850,000 2.13x
Manufacturing TSMC 16nm TSMC 7nm -
Launch Date August 2019 Q3 2021 -
Die Size 46225 mm2 46225 mm2 -
Transistors 1200 billion 2600 billion 2.17x
(Density) 25.96 mTr/mm2 56.246 mTr/mm2 2.17x
On-board SRAM 18 GB 40 GB 2.22x
Memory Bandwidth 9 PB/s 20 PB/s 2.22x
Fabric Bandwidth 100 Pb/s 220 Pb/s 2.22x
Cost $2 million+ arm+leg

As with the original processor, known as the Wafer Scale Engine (WSE-1), the new WSE-2 features hundreds of thousands of AI cores across a massive 46225 mm2 of silicon.  In that space, Cerebras has enabled 2.6 trillion transistors for 850,000 cores - by comparison, the second biggest AI CPU on the market is ~826 mm2, with 0.054 trillion transistors. Cerebras also cites 1000x more onboard memory, with 40 GB of SRAM, compared to 40 MB on the Ampere A100.


Me with Wafer Scale Gen1 - looks the same, but with less than half the cores.

The cores are connected with a 2D Mesh with FMAC datapaths. Cerebras achieves 100% yield by designing a system in which any manufacturing defect can be bypassed – initially Cerebras had 1.5% extra cores to allow for defects, but we’ve since been told this was way too much as TSMC's process is so mature. Cerebras’ goal with WSE is to provide a single platform, designed through innovative patents, that allowed for bigger processors useful in AI calculations but has also been extended into a wider array of HPC workloads.

Building on First Gen WSE

A key to the design is the custom graph compiler, that takes pyTorch or TensorFlow and maps each layer to a physical part of the chip, allowing for asynchronous compute as the data flows through. Having such a large processor means the data never has to go off-die and wait in memory, wasting power, and can continually be moved onto the next stage of the calculation in a pipelined fashion. The compiler and processor are also designed with sparsity in mind, allowing high utilization regardless of batch size, or can enable parameter search algorithms to run simultaneously.

For Cerebras’ first generation WSE is sold as a complete system called CS-1, and the company has several dozen customers with deployed systems up and running, including a number of research laboratories, pharmaceutical companies, biotechnology research, military, and the oil and gas industries. Lawrence Livermore has a CS-1 paired to its 23 PFLOP ‘Lassen’ Supercomputer. Pittsburgh Supercomputer Center purchased two systems with a $5m grant, and these systems are attached to their Neocortex supercomputer, allowing for simultaneous AI and enhanced compute.

Products and Partnerships

Cerebras sells complete CS-1 systems today as a 15U box that contains one WSE-1 along with 12x100 GbE, twelve 4 kW power supplies (6 redundant, peak power about 23 kW), and deployments at some institutions are paired with HPE’s SuperDome Flex. The new CS-2 system shares this same configuration, albeit with more than double the cores and double the on-board memory, but still within the same power. Compared to other platforms, these processors are arranged vertically inside the 15U design in order to enable ease of access as well as built-in liquid cooling across such a large processor. It should also be noted that those front doors are machined from a single piece of aluminium.

The uniqueness of Cerebras’ design is being able to go beyond the physical manufacturing limits normally presented in manufacturing, known as the reticle limit. Processors are designed with this limit as the maximum size of a chip, as connecting two areas with a cross-reticle connection is difficult. This is part of the secret sauce that Cerebras brings to the table, and the company remains the only one offering a processor on this scale – the same patents that Cerebras developed and were awarded to build these large chips are still in play here, and the second gen WSE will be built into CS-2 systems with a similar design to CS-1 in terms of connectivity and visuals.

The same compiler and software packages with updates enable any customer that has been trialling AI workloads with the first system to use the second at the point at which they deploy one. Cerebras has been working on higher-level implementations to enable customers with standardized TensorFlow and PyTorch models very quick assimilation of their existing GPU code by adding three lines of code and using Cerebras’ graph compiler. The compiler then divides the whole 850,000 cores into segments of each layer that allow for data flow in a pipelined fashion without stalls. The silicon can also be used for multiple networks simultaneously for parameter search.

Cerebras states that with having such a large single chip solution means that the barrier to distributed training methods across 100s of AI chips is now so much further away that this excess complication is not needed in most scenarios – to that, we’re seeing CS-1 deployments of single systems attached to supercomputers. However, Cerebras is keen to point out that two CS-2 systems will deliver 1.7 million AI cores in a standard 42U rack, or three systems for 2.55 million in a larger 46U rack (assuming there’s sufficient power for all at once!), replacing a dozen racks of alternative compute hardware. At Hot Chips 2020, Chief Hardware Architect Sean Lie stated that one of Cerebras' key benefits to customers was the ability to enable workload simplification that previously required racks of GPU/TPU but instead can run on a single WSE in a computationally relevant fashion.

As a company, Cerebras has ~300 staff across Toronto, San Diego, Tokyo, and San Francisco. They have dozens of customers already with CS-1 deployed and a number more already trialling CS-2 remotely as they bring up the commercial systems. Beyond AI, Cerebras is getting a lot of interest from typical commercial high performance compute markets, such as oil-and-gas and genomics, due to the flexibility of the chip is enabling fluid dynamics and other compute simulations. Deployments of CS-2 will occur later this year in Q3, and the price has risen from ~$2-3 million to ‘several’ million.


With Godzilla for a size reference

Related Reading

Comments Locked

136 Comments

View All Comments

  • mode_13h - Friday, April 30, 2021 - link

    > What I doubt, however, is that there'll be a set of moral rules it'll discover and go by.

    If humans are any example of such "natural morality", we have a lot to fear. I'm of the belief that morality is a social construct. You can certainly formalize it, but that doesn't make it fundamental.

    Anyway, we know about lots of immoral humans, and morality in practice has a lot to do with empathy -- whether the subject feels it and for whom. Would an AI feel empathy? I doubt it. I think empathy is another one of the social tools that evolved to enable us to live and work in groups.

    Without empathy, morality is basically a cold rule-following exercise. And when something like that ceases to have a net benefit for the subject, then it tends to fall by the wayside.
  • GeoffreyA - Friday, April 30, 2021 - link

    In full agreement with you. Trust me, I'm quite pessimistic concerning man's morality; it's cause for much, much lament. If we can depend on something, it's man's certainty of doing wrong.

    But I feel there's something inside us---a whisper indeed, easily drowned out---that seems to tell us, what we're doing or seeing isn't right. Could it be some sort of coding, of right and wrong, buried deep within us? I'd like to believe that; but perhaps it is a mode of empathy after all, tied to our bringing up and the conscience.
  • mode_13h - Saturday, May 1, 2021 - link

    Several different types of social animals have demonstrated an ability to recognize when they or one of their peers is being treated unfairly. I think that's a form of morality, but I'm not sure if it's been observed in non-social animals.

    Seems to me like it's probably a necessary or highly-advantageous capability for higher-order animals to form and maintain social groups.

    Sorry to so utilitarian. You probably don't even want to get me started on love.
  • GeoffreyA - Saturday, May 1, 2021 - link

    It's all right. I generally agree, but differ on the cause. As for love, I can't resist saying that my view is both romantic and utilitarian. I think it was Nature's flower-wreathed way of bringing two people together, to (ahem) produce young, bring them up, and send them off into the world. Instead of an empty CopyGenetics(A, B), it put in some drama, not to mention heartache (the rose is not without thorns)!

    Concerning the animals' recognition that they're being treated unfairly, and empathy, etc., the insular cortex appears to be the site where all of this takes place. Even tied to mirroring.
  • mode_13h - Wednesday, April 21, 2021 - link

    I'm surprised no one is talking about using these for crypto mining! I wonder how quickly it would pay for itself, at current rates.

    I'd imagine a few of these are finding their way into financial exchanges, as well.
  • ichaya - Thursday, April 22, 2021 - link

    Any indication of how many FLOPS? Each AI core would need to have ~1000 registers to fit the smallest GPT-3 Ada model into 3xWSE2s, 2.7 billion weights. 3000 registers, and you might be able to train GPT-3 Ada on a single WSE2 in possibly hours. Very cool.
  • mode_13h - Friday, April 23, 2021 - link

    > Any indication of how many FLOPS?

    Talk of FLOPS is decidedly absent. That can't have been an oversight.

    > Each AI core would need to have ~1000 registers ...

    Do the math: 40 GiB across 850 k cores per WSE-2 = ~48 kiB of SRAM per core
  • ichaya - Friday, April 23, 2021 - link

    I wanted to see how many registers it would take for GPT-3, but yes, ~48kib means almost 25,000 BF16 registers per core, or half as many FP registers, so more than enough for GPT-3's smallest model. TPUv3 is 90 TOPS and 32GB HBM2 in 250W, Intel PV is .5-1 PFLOP in 600W? the eetimes article says 10-40 of kw for WSE2, so a lot of FLOPs.
  • mode_13h - Saturday, April 24, 2021 - link

    Uh, this article said 23 kW. I don't see how it can be more than 24 kW, when half of those power supplies are supposed to be for redundancy.
  • occidental - Saturday, April 24, 2021 - link

    Why did the woman have a mask on?

Log in

Don't have an account? Sign up now