On-Prem

HPC

Nvidia shovels $500M into Israeli boffinry supercomputer

System to feature hundreds of liquid-cooled Blackwell systems


Nvidia is constructing a 30-megawatt research-and-development supercomputer stuffed with its latest-generation Blackwell GPUs in northern Israel at an estimated cost of half a billion dollars.

The cluster's home is a 10,000-square-metre facility located in the Mevo Carmel Science and Industry Park near the city of Yokne'am Illit. And according to Nvidia, it will be pack hundreds of Nv's liquid-cooled Blackwell-based systems, BlueField-3 SuperNIC, Spectrum-X800, and Quantum-X800 switches.

We understand the big beast will be used by Nvidia employees to develop next-gen datacenter technologies.

The exact number of accelerators being deployed isn't clear — Nvidia has several liquid-cooled Blackwell reference systems with varying numbers of accelerators on board — however, Israeli media reports the full configuration will feature "several thousand" GPUs rivaling the nation's Israel-1 supercomputer.

That system is made up of 2,048 H100 accelerators stitched together using Nvidia's Ethernet-based Spectrum-X family of switches and superNICs to provide peak FP64 performance of 69 (vector) to 137 (matrix) petaFLOPS for scientific workloads or 8 exaFLOPS for AI applications (sparse FP8).

On a chip-for-chip basis, Blackwell promises up to 2.5x the floating-point performance of Hopper at most precisions, and up to 5x when dropping down to 4-bit precision. In addition to the higher computational performance, Blackwell also promises between 1.66x and 2.38x higher memory bandwidth compared to its predecessor. So even with an equal number of accelerators as Israel-1, the new facility should translate into a more powerful system.

System building is said to have begun sometime last year and is expected to start operations in the first half of 2025.

What about the new export rules?

Depending on how far along Nvidia is and just how many GPUs it plans to deploy, the chipmaker could just about potentially run afoul of the Biden administration's latest round of export controls on AI accelerators.

As we saw with xAI's Colossus AI supercomputer, Nvidia and its partners have proven more than capable of deploying massive systems in a matter of months.

Under the rules disclosed on Monday, Israel would be labeled a tier-two nation and be subject to import caps equivalent to 50,000 advanced GPUs over a two-year period between 2025 and 2027.

However, it's our understanding that these rules won't go into effect for 120 days from publication, and may not survive a Trump presidency. If the president-elect's administration does move forward with these plans, the implementation period could give Nvidia enough time to complete any remaining shipments necessary to finish the cluster.

It's also possible the super may slip under the export caps or be excluded from the rules as it's believed the computer will be used exclusively for internal research-and-development purposes.

Under the rules, chip orders with a collective computational power of up to around 1,700 advanced GPUs would not require a special license from Uncle Sam and would not count against a nation's chip caps.

Meanwhile, entities in tier-one nations that meet US security standards can place up to seven percent of their global computation in other nations around the world.

Given Israel's strong ties to the US, there's a good chance the nation would be granted "National Verified End User" status, boosting the import cap to 320,000 accelerators over a two-year period.

Even still, the rules have become a source of major concern for Israeli tech companies, which worry the rules could eliminate Israel's ability to compete in the emerging AI field. ®

Send us news
23 Comments

Where does Microsoft's NPU obsession leave Nvidia's AI PC ambitions?

While Microsoft pushes AI PC experiences, Nvidia is busy wooing developers

AI frenzy continues as Macquarie commits up to $5B for Applied Digital datacenters

Bubble? What bubble?

CoreWeave drops £1bn in UK datacenters – but don't expect the latest Nvidia magic just yet

Rent-a-GPU outfit's latest datacenters are packed to the brim with H200s

Nvidia snaps back at Biden's 'innovation-killing' AI chip export restrictions

'New rule threatens to squander America's hard-won technological advantage' says GPU supremo

With AI boom in full force, 2024 datacenter deals reach $57B record

Fewer giant contracts, but many more smaller ones, in bit barn feeding frenzy

Nvidia shrinks Grace-Blackwell Superchip to power $3K mini PC

Tuned for running chunky models on the desktop with 128GB of RAM, custom Ubuntu

Additional Microprocessors Decoded: Quick guide to what AMD is flinging out next for AI PCs, gamers, business

Plus: A peek at Nvidia's latest hype

AI hype led to an enterprise datacenter spending binge in 2024 that won't last

GPUs and generative AI systems so hot right now... yet 'long-term trend remains,' says analyst

Just as your LLM once again goes off the rails, Cisco, Nvidia are at the door smiling

Some of you have apparently already botched chatbots or allowed ‘shadow AI’ to creep in

AI datacenters putting zero emissions promises out of reach

Plus: Bit barns' demand for water, land, and power could breed 'growing opposition' from residents

UK unveils plans to mainline AI into the veins of the nation

Government adopts all 50 venture capitalist recommendations but leaves datacenter energy puzzle unsolved

Europe hopes Trump trumps Biden's plan for US to play AI gatekeeper

Export controls would limit shipments of GPUs to large swaths of EU