//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
SUNNYVALE, California—G42 has commissioned a major AI supercomputer, Condor Galaxy 1, to be constructed this 12 months in close by Santa Clara, based mostly on {hardware} from AI chip startup Cerebras. Condor Galaxy 1, which is able to include 64 linked Cerebras CS-2 wafer-scale AI accelerators, represents a deal value “in extra of $100 million,” Cerebras CEO Andrew Feldman instructed EE Instances.
Following Condor Galaxy 1, Cerebras and G42, an Abu Dhabi–based mostly know-how group, plan eight additional 64-CS-2 clusters inside and out of doors america, that are set to succeed in a complete of 576 linked CS-2 programs by the tip of subsequent 12 months.
The entire price is anticipated to be in extra of $900 million.
“G42 was searching for a companion that had experience in constructing, managing and working supercomputers, and there will not be so many people anymore who’re implementing large generative AI fashions and manipulating, cleansing and managing massive datasets, who may ship infrastructure with quick compute that’s simple to make use of, versatile and had a monitor file of remarkable collaboration,” Feldman stated. “These have been among the causes they selected us.”
Condor Galaxy 1 represents the primary supercomputer-scale set up of AI accelerator {hardware} from any startup, anyplace, and validates the marketplace for specialised AI {hardware} outdoors of GPUs, a sector that has attracted billions of {dollars} of enterprise capital during the last 5 years.
Describing the deal as “a problem on a number of fronts,” Feldman stated, “It’s a problem to Nvidia. It’s a problem to hyperscalers. It’s saying, ‘Look, you don’t must be one among these [big] guys to face up extraordinary compute.’”
Whereas hyperscalers would appear an apparent potential buyer for AI compute at scale, they’ve thus far chosen Nvidia GPU chips and/or constructed in-house accelerators. This deal illustrates the market potential outdoors of hyperscalers like Google, AWS, Meta and Microsoft.
“AI is consuming the world,” Feldman stated. “That is taking place all over the place, and that’s creating demand in locations that haven’t been publicized.”
Half of the deliberate Condor Galaxy 1 cluster, 32 CS-2 programs, is already up and operating in Santa Clara. It was stood up in simply 10 days.
“There’s a great alternative proper now, and there’s urgency,” Feldman stated. “There’s urgency created by the chance to make hay when the solar shines and be early within the creation of among the strongest AIs to unravel buyer issues and get up a franchise on this new area.”
G42 is a conglomerate based mostly within the UAE. It’s made up of 9 firms with 22,000 staff, and it already has the most important regional cloud within the Center East.
Among the many G42 firms that can use the Condor Galaxy 1 are the Inception Institute of Synthetic Intelligence, which makes fashions and helps enterprises deploy them all through the Center East, and M42, G42’s three way partnership with Mubadala that’s targeted on gaining new insights into healthcare information, together with genomic sequencing.
The primary part of deployment of Condor Galaxy 1—32 CS-2s—is already operating G42’s workloads, and any extra capability is bought through Cerebras’s cloud providing.
Whereas G42 owns the cluster, Cerebras will handle and function it. Inside 12 weeks, Cerebras and G42 will double their measurement to 64 CS-2s. When totally commissioned, Condor Galaxy 1 will supply a complete of 54 million AI cores and 84 TB of reminiscence, for a complete of 4 exaFLOPS of AI compute (FP16).
A part of the rationale for putting in Condor Galaxy 1 in america, in accordance with Feldman, was pace.
“We needed to face it up unbelievably rapidly, and we will transfer actually rapidly right here” within the U.S., Feldman stated. “We’re a water-cooled system, and we have now a facility right here that had megawatts for us, and water.”
Every CS-2’s peak sustained system energy is 23 kW, so 64 will draw an virtually 1.5-MW peak. The identical Santa Clara facility, Colovore, additionally homes Cerebras’s present 16-CS-2 AI supercomputer, Andromeda.
By the center of subsequent 12 months, Cerebras and G42 plan to construct one other two 64-CS-2 AI supercomputers, Condor Galaxy 2 and Condor Galaxy 3, in Austin, Texas, and Asheville, North Carolina. These machines will likely be linked to Condor Galaxy 1, for a complete of 192 CS-2 nodes at 12 exaFLOPS (FP16).
By the tip of subsequent 12 months, an extra six machines the identical measurement as Condor Galaxy 1 will likely be commissioned outdoors america and linked to the others.
As with the U.S. installations, G42 will personal the clusters, however they are going to be designed, constructed and operated by Cerebras. The businesses will resell any extra capability collectively.
Areas outdoors of america are nonetheless being thought of, however Feldman is conscious of the geopolitical influence of such a major cluster.
“We’re the quickest AI processor, the quickest AI system and we’re engaged with U.S. regulators about the place we will ship,” Feldman stated.
Cerebras will nonetheless handle and function abroad clusters, seeing each job from each IP tackle and having management of the machines, Feldman stated, arguing that whereas AI is highly effective and can be utilized nefariously, there may be much less probability of AI-specialized machines getting used for navy purposes and cryptography, which usually require 64-bit compute (CS-2 doesn’t help double precision). Safety threats are already effectively understood as a part of common cloud know-how administration, he stated.
General, the G42 and Cerebras contract will add as much as 9 clusters (576 CS-2 nodes for 36 exaFLOPS at FP16). Assuming the identical price construction, all the deal may very well be value in extra of $900 million.
