Source: The content is compiled from "
HPC
" by Semiconductor Industry Observer (ID: icbank), thank you.
Five months ago, when Cerebras Systems introduced its second-generation wafer-scale system-on-a-chip (CS-2), co-founder and CEO Andrew Feldman hinted at the company’s upcoming cloud plans, and now those plans have come to fruition.
Cerebras and Cirrascale Cloud Services have launched the Cerebras Cloud @ Cirrascale platform, providing access to Cerebras’ CS-2 Wafer-Scale Engine (WSE) system through Cirrascale’s cloud services.
The CS-2 machine runs 850,000 AI-optimized computing cores, weighs about 500 pounds, and is housed in a Cirrascale data center in Santa Clara, California, but the service will be available worldwide, with open access to the CS-2 available to anyone with an internet connection and $60,000 per week to train very large AI models.
“For training, we haven’t found latency to be an issue,” Cirrascale CEO PJ Go said at a media pre-launch event last week in conjunction with the AI Hardware Summit.
Feldman agreed, adding, “If you’re doing a 20-hour or longer training run, the speed of light from Cleveland to San Jose might not be a big deal.”
Cirrascale's Cerebras Cloud customers will receive full access to Cerebras software and compiler packages.
“The compiler toolset sits underneath the cloud toolset that Cirrascale has developed,” Feldman said. “So you’ll come in, you’ll get access to the compute cluster, the storage, the CS-2; you’ll run your compilation stack, you’ll do your work, you’ll be checkpointed and stored in the Cirrascale infrastructure, and it’ll be recognized so you can come back to that work later. It’s all integrated together.”
The environment supports familiar frameworks such as TensorFlow and PyTorch, and the Cerebras Graph Compiler automatically converts practitioners’ neural networks from their framework representations into CS-2 executables. According to Cerebras, this eliminates the need for cluster orchestration, synchronization, and model tuning.
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 can access an entire CS-2 system. “The sharing model doesn’t work for us,” Feldman said. The raison d’être of the wafer-scale system is “to be as big as the machine is, so you can solve your problems as quickly as you can,” he told HPCwire.
Cerebras doesn't reveal 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 cite “try before you buy” as one of the motivations for Cerebras’ Cloud product, converting renters impressed by the CS-2’s capabilities into buyers of one or more systems. But the companies also expect a large portion of users to stick with the cloud model.
OPEX preference is one reason, but it’s also a question of skills and experience. Speaking of this, Feldman said, “One of the little-known facts about our industry is how few people can actually build large GPU clusters, and how rare that is — it requires skills, not just money. The skills to spread a large model across 250+ GPUs probably exist in a few dozen organizations in the world.”
Cerebras Cloud simplifies this process by providing performance through cloud-based hardware and software infrastructure, as well as access to billing, storage and other services through the Cirrascale portal. “It was an obvious choice for us to expand our reach to different types of customers,” Feldman said.
Cerebras’ first CS system deployments were in the government lab space (the U.S. Department of Energy was a fundamental win, announced at the 2019 AI Hardware Summit) and commercial sites, primarily on-premises in pharmaceuticals (GSK is a customer). By making the CS-2 accessible as a managed service, Cerebras is courting a wider range of organizations, from startups to Fortune 500 companies.
“We’ve been working on this partnership for a while now,” Andy Hock, vice president of product at Cerebras Systems, said in a promotional video. “We’re starting with a focus on training large natural language processing models from scratch, like BERT, and we’ll expand our offering from there.”
“The workloads that Cerebras CS-2 handles are ones that we can’t do on GPUs today,” said David Driggers, founder and CTO of Cirrascale. “[This is] a very massive scaling scenario where we have a model that can’t be parallelized, but it processes a lot of data. So the largest NLP models today require a lot of data input as well as a lot of compute. That’s very hard to do on a [traditional] cluster because of the IO communication that’s required.”
The Cerebras CS-2 allows us to take advantage of a very large memory space, large built-in networks, and a large number of cores to scale NLP to heights we couldn’t do before.”
Analyst Karl Freun, who participated in the pre-briefing call, applauded the partnership. “Cerebras seems to be firing on all cylinders lately, with customer wins, second generation WSE, and most recently their bold claim that they are building a brain-scale AI that is 1,000 times greater than anything we’ve ever seen,” he told HPCwire.
“What you have is a very hot commodity (their technology) that a lot of people want to try, but who don’t want to spend a lot of money to purchase 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 access to the largest and fastest AI processors in the industry. This will undoubtedly create new opportunities for Cerebras in the future, both in the cloud and on-premises.”
Asked about the risk that today’s AI chips won’t be suitable for tomorrow’s AI models, Freund said, “If anything, Cerebras is a company whose architecture is skating in the direction the puck is heading: big AI.”
Original link:
https://www.hpcwire.com/2021/09/16/cerebras-wafer-scale-engine-ai-system-is-now-available-in-the-cloud/
*Disclaimer: This article is originally written by the author. The content of the article is the author's personal opinion. Semiconductor Industry Observer reprints it only to convey a different point of view. It does not mean that Semiconductor Industry Observer agrees or supports this point of view. If you have any objections, please contact Semiconductor Industry Observer.
Today is the 2805th content shared by "Semiconductor Industry Observer" for you, welcome to follow.
Semiconductor Industry Observation
"
The first vertical media in semiconductor industry
"
Real-time professional original depth
Scan the QR code
, reply to the keywords below, and read more
Wafers｜ICs｜Equipment
｜Automotive Chips｜Storage｜TSMC｜AI｜Packaging
Reply
Submit your article
and read "How to become a member of "Semiconductor Industry Observer""
Reply
Search
and you can easily find other articles that interest you!
京公网安备 11010802033920号