Nvidia to energy extra supercomputers as AI frenzy kicks in

Computex Not content material with unveiling its DGX GH200 AI supercomputer at Computex, Nvidia mentioned it’s concerned in a number of different supercomputers focusing on AI processing, together with one in Israel and two based mostly in Taiwan.

As a part of the announcement surrounding its Nvidia Spectrum-X Networking Platform, the GPU large mentioned it supposed to construct a “blueprint and testbed” system to showcase the expertise. This will likely be referred to as Israel-1, and described as a hyperscale generative AI supercomputer, which will likely be deployed into its Israeli datacenter.

Not like another programs disclosed, this one will likely be based mostly on Nvidia’s current HGX H100 expertise, BlueField-3 DPUs, and constructed utilizing Dell PowerEdge XE9680 servers.

Spectrum-X is itself based mostly on the pairing of Nvidia’s Spectrum-4 Ethernet change with BlueField-3 DPUs. The Spectrum-4 is claimed to offer 64 ports of 800Gbps Ethernet, whereas these DPUs use RDMA over Converged Ethernet (RoCE) to spice up knowledge transfers.

In accordance with The Instances of Israel, this “blueprint and testbed” system is predicted to be able to a efficiency of as much as 8 exaflops, which might make it one of many world’s quickest AI supercomputers when it comes on-line a while in direction of the top of 2023.

“Spectrum-X is a brand new class of Ethernet networking that removes limitations for next-generation AI workloads which have the potential to remodel total industries,” Nvidia’s senior vp of networking Gilad Shainer mentioned in a press release.

Taipei-1 will likewise be constructed and operated by Nvidia and based mostly round its H100 expertise. On this case it should comprise 64 DGX H100 programs, plus 64 OVX programs. Every DGX H100 packs in 8 of the H100 GPUs, based mostly on Nvidia’s Hopper structure, whereas the OVX options L40 GPUs based mostly on the Ada Lovelace structure.

Additionally based mostly in Taiwan will likely be Taiwania 4, which is about to be constructed by Asus and situated on the Nationwide Heart for Excessive-Efficiency Computing (NCHC). This will likely be based mostly on Nvidia’s Grace CPU Superchip, which mixes two Arm-based processor dies for a complete of 144 compute cores. It’s understood that Taiwania 4 will comprise 44 nodes, linked utilizing Nvidia’s Quantum-2 InfiniBand interconnect.

In the meantime, Asus can be planning to supply AI servers based mostly on Nvidia’s DGX {hardware}. In accordance with Bloomberg, the AFS Equipment will likely be provided below a subscription-based mannequin, however put in on the client’s personal premises. That is to permit organizations to benefit from generative AI modesl whereas holding management over their knowledge, moderately than constructing functions within the cloud.

Nevertheless, this comes at a worth: the AFS Equipment with Nvidia DGX will doubtless value round $10,000 a month, though Asus informed Bloomberg it goals to have 30 to 50 enterprise clients in Taiwan and develop internationally by the top of this yr.

This information follows the announcement yesterday of the DGX GH200 AI supercomputer, that includes 256 Grace-Hopper superchips stitched along with Nvidia’s NVLink expertise, plus a reference structure for servers utilizing its accelerators referred to as MGX.

All of this frenzy of curiosity round AI, particularly the most recent era of enormous language fashions (LLMs) and generative AIs has helped gasoline demand for Nvidia’s merchandise, with the consequence that the GPU-flinger has now hit a trillion-dollar market worth, in keeping with Reuters.

This turnaround follows information earlier this month that CEO Jen-Hsun Huang noticed his remuneration fall 10 p.c due to missed monetary targets for its fiscal yr 2023 that led to January. ®