Synthetic intelligence is on a tear. Machines can converse, write, play video games, and generate authentic pictures, video, and music. However as AI’s capabilities have grown, so too have its algorithms.
A decade in the past, machine studying algorithms relied on tens of hundreds of thousands of inner connections, or parameters. Right now’s algorithms repeatedly attain into the lots of of billions and even trillions of parameters. Researchers say scaling up nonetheless yields efficiency features, and fashions with tens of trillions of parameters might arrive in brief order.
To coach fashions that huge, you want highly effective computer systems. Whereas AI within the early 2010s ran on a handful of graphics processing items—pc chips that excel on the parallel processing essential to AI—computing wants have grown exponentially, and prime fashions now require lots of or 1000’s. OpenAI, Microsoft, Meta, and others are constructing devoted supercomputers to deal with the duty, and so they say these AI machines rank among the many quickest on the planet.
However at the same time as GPUs have been essential to AI scaling—Nvidia’s A100, for instance, remains to be one of many quickest, mostly used chips in AI clusters—weirder alternate options designed particularly for AI have popped up lately.
Cerebras presents one such various.
Making a Meal of AI
The dimensions of a dinner plate—about 8.5 inches to a aspect—the corporate’s Wafer Scale Engine is the greatest silicon chip on the earth, boasting 2.6 trillion transistors and 850,000 cores etched onto a single silicon wafer. Every Wafer Scale Engine serves as the center of the corporate’s CS-2 pc.
Alone, the CS-2 is a beast, however final yr Cerebras unveiled a plan to hyperlink CS-2s along with an exterior reminiscence system known as MemoryX and a system to attach CS-2s known as SwarmX. The corporate mentioned the brand new tech might hyperlink as much as 192 chips and practice fashions two orders of magnitude bigger than right this moment’s greatest, most superior AIs.
“The business is shifting previous 1-trillion-parameter fashions, and we’re extending that boundary by two orders of magnitude, enabling brain-scale neural networks with 120 trillion parameters,” Cerebras CEO and cofounder Andrew Feldman mentioned.
On the time, all this was theoretical. However final week, the firm introduced they’d linked 16 CS-2s collectively right into a world-class AI supercomputer.
Meet Andromeda
The brand new machine, known as Andromeda, has 13.5 million cores able to speeds over an exaflop (one quintillion operations per second) at 16-bit half precision. Because of the distinctive chip at its core, Andromeda isn’t simply in comparison with supercomputers operating on extra conventional CPUs and GPUs, however Feldman instructed HPC Wire Andromeda is roughly equal to Argonne Nationwide Laboratory’s Polaris supercomputer, which ranks seventeenth quickest on the earth, in line with the newest Top500 listing.
Along with efficiency, Andromeda’s speedy construct time, value, and footprint are notable. Argonne started putting in Polaris in the summertime of 2021, and the supercomputer went stay a few yr later. It takes up 40 racks, the filing-cabinet-like enclosures housing supercomputer elements. By comparability, Andromeda value $35 million—a modest value for a machine of its energy—took simply three days to assemble, and makes use of a mere 16 racks.
Cerebras examined the system by coaching 5 variations of OpenAI’s giant language mannequin GPT-3 in addition to Eleuther AI’s open supply GPT-J and GPT-NeoX. And in line with Cerebras, maybe an important discovering is that Andromeda demonstrated what they name “near-perfect linear scaling” of AI workloads for big language fashions. In brief, which means as extra CS-2s are added, coaching instances lower proportionately.
Sometimes, the corporate mentioned, as you add extra chips, efficiency features diminish. Cerebras’s WSE chip, alternatively, might show to scale extra effectively as a result of its 850,000 cores are related to one another on the identical piece of silicon. What’s extra, every core has a reminiscence module proper subsequent door. Taken collectively, the chip slashes the period of time spent shuttling knowledge between cores and reminiscence.
“Linear scaling means whenever you go from one to 2 programs, it takes half as lengthy to your work to be accomplished. That may be a very uncommon property in computing,” Feldman instructed HPC Wire. And, he mentioned, it will probably scale past 16 related programs.
Past Cerebras’s personal testing, the linear scaling outcomes had been additionally demonstrated throughout work at Argonne Nationwide Laboratory the place researchers used Andromeda to coach the GPT-3-XL giant language algorithm on lengthy sequences of the Covid-19 genome.
After all, although the system might scale past 16 CS-2s, to what diploma linear scaling persists stays to be seen. Additionally, we don’t but know the way Cerebras performs head-to-head towards different AI chips. AI chipmakers like Nvidia and Intel have begun collaborating in common third-party benchmarking by the likes of MLperf. Cerebras has but to participate.
Area to Spare
Nonetheless, the method does look like carving out its personal area of interest on the earth of supercomputing, and continued scaling in giant language AI is a first-rate use case. Certainly, Feldman instructed Wired final yr that the corporate was already speaking to engineers at OpenAI, a pacesetter in giant language fashions. (OpenAI founder, Sam Altman, can be an investor in Cerebras.)
On its launch in 2020, OpenAI’s giant language mannequin GPT-3, modified the sport each when it comes to efficiency and measurement. Weighing in at 175 billion parameters, it was the most important AI mannequin on the time and stunned researchers with its talents. Since then, language fashions have reached into the trillions of parameters, and bigger fashions could also be forthcoming. There are rumors—simply that, up to now—that OpenAI will launch GPT-4 within the not-too-distant future and will probably be one other leap from GPT-3. (We’ll have to attend and see on that rely.)
That mentioned, regardless of their capabilities, giant language fashions are neither good nor universally adored. Their flaws embody output that may be false, biased, and offensive. Meta’s Galactica, educated on scientific texts, is a latest instance. Regardless of a dataset one may assume is much less susceptible to toxicity than coaching on the open web, the mannequin was simply provoked into producing dangerous and inaccurate textual content and pulled down in simply three days. Whether or not researchers can resolve language AI’s shortcomings stays unsure.
But it surely appears seemingly that scaling up will proceed till diminishing returns kick in. The following leap may very well be simply across the nook—and we might have already got the {hardware} to make it occur.
Picture Credit score: Cerebras
