Cerebras Brings Its Wafer-Scale Engine AI System to the Cloud
Five months ago, when Cerebras Systems debuted its second-generation wafer-scale silicon system (CS-2), co-founder and CEO Andrew Feldman hinted of the company’s coming cloud plans, and now those plans have come to fruition. Today, Cerebras and Cirrascale Cloud Services are launching the Cerebras Cloud @ Cirrascale platform, providing access to Cerebras’ CS-2 Wafer-Scale Engine (WSE) system through Cirrascale’s cloud service.
The physical CS-2 machine – sporting 850,000 AI optimized compute cores and weighing in at approximately 500 lbs – is installed in the Cirrascale datacenter in Santa Clara, Calif., but the service will be available around the world, opening up access to the CS-2 to anyone with an internet connection and $60,000 a week to spend training very large AI models.
“For training, we have not found latency to be an issue,” said Cirrascale CEO PJ Go in a media pre-briefing, held in conjunction with the AI Hardware Summit this week.
Feldman agreed, adding, “If you’re going to run your training for 20 hours or more, the speed of light to get from Cleveland to San Jose is probably not too big issue.”
Cirrascale’s Cerebras Cloud customers will gain full access to Cerebras’ software and compiler package.
“The compiler toolset sits underneath the cloud toolset that Cirrascale has developed,” said Feldman. “And so you will enter, you’ll gain access to a compute cluster, storage, a CS-2; you will run your compile stack, you will do your work, you will be checkpointed and stored in the Cirrascale infrastructure, it will be identified so you can get back to that work later. All of that has been integrated.”
The environment supports familiar frameworks such as TensorFlow and PyTorch, and the Cerebras Graph Compiler automatically translates the practitioner’s neural network from their framework representation into a CS-2 executable. This eliminates the need for cluster orchestration, synchronization and model tuning, according to Cerebras.
With a weekly minimum buy-in — pricing is set at $60,000 per week, $180,000 per month or $1,650,000 per year — Cirrascale customers get access to the entire CS-1 system. “The shareable model is not for us,” said Feldman. The raison d’etre of the wafer-scale system is “to get as big of a machine as you can to solve your problem as quickly as you can,” he told HPCwire.
Discounts are provided for multi-month or multi-year contracts. Cerebras does not disclose list prices for its CS systems, but buying a CS-2 system outright will set you back “several million dollars,” according to Feldman.
Both CEOs agreed that “try before you buy” was one of the motivations of the Cerebras Cloud offering, converting renters who are impressed by what CS-2 can do into buyers of one or more systems. But the companies also expect a good share of users to stick with the cloud model.
A preference for OPEX is one reason, but it’s also an issue of skills and experience. Driving home this point, Feldman said, “A little known fact about our industry is how few people can actually build big clusters of GPUs, how rare it is — the skills that are necessary, not just the money. The skills to spread a large model over more than 250 GPUs is probably resident in a couple of dozen organizations in the world.”
Cerebras Cloud offers to streamline this process by making the performance available via a cloud-based hardware and software infrastructure with the billing, storage and other services accessible via the Cirrascale portal. “It was an obvious choice for us in extending our reach to different types of customers,” Feldman said.
Cerebras’ first CS system deployments were on-premises in the government lab space (the U.S. DOE was a foundational win, announced at the 2019 AI Hardware Summit) and commercial sites, mainly pharma (GlaxoSmithKline is a customer). By making CS-2 accessible as a hosted service, Cerebras is going after a broader set of organizations, from startups to Fortune 500 companies.
“We’ve been working on this partnership for some time,” said Andy Hock, vice president of product at Cerebras Systems, in a promo video. “We’re beginning with a focus on training large natural natural language processing models, like BERT, from scratch and we’ll expand our offering from there.”
“The Cerebras CS-2 handles a type of workload that we cannot do on GPUs today,” said David Driggers, founder and CTO, Cirrascale. “[It’s] a very-large scale-up scenario, where we’ve got a model that just does not parallelize and yet it’s managing to deal with a very large amount of data. So the largest NLP models today require a tremendous amount of data input as well as a tremendous amount of calculation. This is very difficult to do on a [traditional] cluster due to the IO communication that is required. The Cerebras CS-2 allows us to leverage the very large memory space, the large built-in networking and the huge amount of cores to be able to scale NLP to heights that we haven’t been able to do before.”
Analyst Karl Freund (principal, Cambrian AI Research), who was on the pre-briefing call, gave the partnership his nod of approval. “Cerebras seems to be firing on all cylinders of late, with customer wins, the 2nd gen WSE, and most recently their audacious claims that they are building a brain-scale AI 1000 times larger than anything we have seen yet,” he told HPCwire.
“What you have is a very hot commodity (their technology) that a lot of people want to experiment with, but who do not want to spend the very big bucks it would take to buy and deploy a CS-2. Enter Cirrascale, and their CS-2 cloud offering, which will make it easier and at least somewhat more affordable for scientists to get their hands on the biggest, fastest AI processor in the industry. This will undoubtably create new opportunities for Cerebras going forward, both in the cloud and on-premises.”
Asked about the risk that today’s AI silicon won’t be suitable for future AI models, Freund said, “if anything, Cerebras is the company who’s architecture is skating to where the puck is going: huge AI.”
This article first appeared on sister website HPCwire.
