Nearly two years since its massive 1.2 trillion transistor Wafer Scale Engine chip debuted at Hot Chips, Cerebras Systems is announcing its second-generation technology (WSE-2), which its says packs twice the performance into the same 8″x8″ silicon footprint.

“We’re going bigger, faster and better in a more power efficient footprint,” Cerebras Founder and CTO Andrew Feldman told HPCwire ahead of today’s launch.

With 2.6 trillion transistors and 850,000 cores, the WSE-2 more than doubles the elements on the first-gen chip (1.2 trillion transistors, 400,000 cores). The new chip, made by TSMC on its 7nm node, delivers 40 GB of on-chip SRAM memory, 20 petabytes of memory bandwidth and 220 petabits of aggregate fabric bandwidth. Gen over gen, the WSE-2 provides about 2.3X on all major performance metrics, said Feldman.

Compared to the the largest GPU, which has ~54 billion transistors, the WSE-2 is 2.55 trillion transistors larger. Further, Cerebras claims its new platform has 123 times more cores, 1,000 times more on-chip memory, more than 12,000 times the memory bandwidth and more than 45,000 times the fabric bandwidth as the leading GPU.

Both Cerebras’ first and second generation chips are created by removing the largest possible square from a 300 mm wafer to create 46,000 square millimeter chips roughly the size of a dinner plate. An array of repeated identical tiles (84 of them) is built into the wafer, enabling redundancy.

To drive its new engine, Cerebras designed and built its next-generation system, the CS-2, which it bills as the industry’s fastest AI supercomputer. Like the original CS-1, CS-2 employs internal water cooling in a 15 unit rack enclosure with 12 lanes of 100 Gigabit Ethernet.

The new system has a max power draw of 23 kW, up from 20 kW max for the original chassis. “We tried to stay in the original power envelope, and made some changes in the system to take full advantage of the power envelope,” said Feldman.

Feldman said that the last two years have taught him how valuable it is for a system to be physically easy to deploy. The “CS” systems weigh approximately 500 lbs, about the same as 15 RU servers, but deploy in just 15 minutes, he said, adding that cabling projects on a typical cluster can take weeks.

Looking further back to when Cerebras was still designing its first-generation product and charting its go-to-market strategy, Feldman said that he originally underestimated the size of the market due to how quickly the space is moving.

“In my career, I’ve always misestimated on the too big side. I’ve always assumed the market was going to be bigger than it is,” he shared. “In 2015, I estimated the market will be smaller than it is, and the demand for AI and the the rate of innovation.

“We’re selling a lot of systems to do BERT. BERT didn’t exist the first half of 2018, right? That’s a quick moving market if what hadn’t yet existed until Q3 2018 is [now] the bulk of business. And things that are brand new, like graph neural networks, are piquing people’s interest and top of everybody’s mind. The market is moving unbelievably quickly.”

Making the leap to wafer-scale

As Feldman describes it, codes that have been optimized for CS-1’s 400,000 cores will scale to leverage CS-2’s 850,000 cores without any modification. Further, he attests that GPU codes are easy to port to the Cerebras platform. “We can take as input any TensorFlow or PyTorch model designed for a GPU. You define your model and write it in TensorFlow, that’s your model function. You define your parameters, that’s your input function. All you have to do is take your TensorFlow code, and type one thing: “est = CerebrasEstimator.

“That’s how you take a model that was written for a GPU and run it on our machine,” he said. “Each layer of your neural network is converted into a region of compute. Then we configure a circuit through it, begin streaming your data, and we send out the answers.” (See figure below.)

The CTO further claimed it is easier to go from a model written for one GPU to a Cerebras CS-1, than it is to go from a model written for one GPU to 20 GPUs.

Primarily dedicated to AI computing, the Cerebras engine is also being applied to HPC workloads. At last year’s Supercomputing Conference, researchers from Cerebras and NETL revealed how the CS-1 was being used perform fast stencil-code computation (a CFD code), demonstrating a speed up of 200X over the largest existing supercomputers.

How big is too big?

With a chip this size, the question arises as to whether users can fill up all that silicon real estate? Are today’s models able to take advantage of an 850,000 core machine? Feldman said there’s no question the demand is there for the larger Cerebras machine, and he sees smaller size platforms (including GPUs) as being primarily for entry-level use cases.

“If you’re happy on one GPU, then we’re not the machine for you,” he said. “But we think that is a small part of the market. As soon as you do work on one GPU, you want more, you want to write bigger models. When you use AI to further the mission of your organization – those are our customers, not the hobbyist.”

Users who need a smaller slice of the Cerebras engine may still be in luck, however. Cerebras plans to announce a cloud offering down the road, according to Feldman.

Some big name wins

Cerebras has racked up a number of key deployments over the last two years, including cornerstone wins with the U.S. Department of Energy, which has CS-1 installations at Argonne National Laboratory and Lawrence Livermore National Laboratory. CS-1 systems are also in place at Pittsburgh Supercomputer Center, EPCC and GlaxoSmithKline, and Cerebras says it has customers in the heavy manufacturing, pharma, biotech, and the military and intelligence sectors.

CS-2 systems will begin shipping in the third quarter of this year, according to Cerebras, and current customers GlaxoSmithKline and Argonne National Lab are expected to be among the the first to take delivery of the upgraded machines.

“At GSK, we are pioneering the use of AI in drug discovery and design,” said Kim Branson, executive vice president of AI, GlaxoSmithKline. “We have been early adopters of the Cerebras technology and have found extraordinary speedups over our legacy infrastructure. We are excited to receive delivery of our CS-2.”

“As an early customer of Cerebras solutions, we have experienced performance gains that have greatly accelerated our scientific and medical AI research,” said Rick Stevens, Argonne National Laboratory associate laboratory director for computing, environment and life sciences, in a statement. “The CS-1 allowed us to reduce the experiment turnaround time on our cancer prediction models by 300X over initial estimates, ultimately enabling us to explore questions that previously would have taken years, in mere months. We look forward to seeing what the CS-2 will be able to do with more than double that performance.”