NVIDIA has unveiled its newest Blackwell GB200 NVL4 resolution with 4 GPUs & two complete GPUs in a single tough HPC & AI resolution. NVIDIA Speeds up AI Ahead With Robust {Hardware} Answers: Blackwell GB200 NVL4 & Hopper H200 NVL As a part of its SC24 bulletins, NVIDIA is unveiling two new {hardware} platforms, one in accordance with its current Hopper stack and the opposite powered through its newest Blackwell stack . Those two answers are designed for undertaking servers, powering a rising choice of HPC and AI programs.
Beginning with the NVIDIA H200 NVL, the corporate now confirms the provision of PCIe-based Hopper playing cards that may attach 4 GPUs during the NVLINK area, offering seven instances extra bandwidth than the PCIe resolution. The corporate says that the H200 NVL answers can are compatible into any knowledge middle and be offering quite a lot of versatile server configurations which might be optimized for hybrid HPC and AI purposes.
In relation to specs, the Hopper H200 NVL resolution gives 1.5x HBM reminiscence, 1.7x LLM inference efficiency, and 1.3x HPC efficiency. You get 114 SMs with 14,592 CUDA cores, 456 tensor cores, as much as 3 TFLOPs of FP8 (FP16 assembled). The GPU has 80 Gb HBM2e reminiscence configured at 5120-bit layout and has a TDP of 350 Watts.
Shifting directly to Blackwell, we’ve our first demonstration of the NVIDIA GB200 NVL4, which is a brand new unit that could be a greater model of the unique GB200 Grace Blackwell Superchip AI resolution. The GB200 NVL4 module will increase CPU and GPU efficiency and will increase reminiscence capability. You might be getting two Blackwell GB200 GPUs fixed on a mainboard with two Grace CPUs. This unit is designed as a unmarried server resolution with 4-GPU NVLINK area and 1.3T of Coherent reminiscence. In relation to efficiency, this module will supply a 2.2x development in estimation, and a 1.8x development in Coaching and Inference efficiency. NVIDIA’s rising lineup will be offering an NVL4 resolution within the coming months.
As for the TDP, for the reason that Superchip section is ready 2700W, we will be able to be expecting the primary GB200 NVL4 option to cross round 6KW of energy. One can see the efforts NVIDIA is making to advance the sector of AI computing. The corporate has not too long ago launched the arena report in MLPerf v4.1, each in coaching and interpretation, and no longer best is Blackwell superb, however Hopper is getting higher with steady optimization. The corporate may be accelerating its AI technique through a 12 months with a number of new units deliberate for the approaching years, such because the Blackwell Extremely and Rubin. Proportion this newsletter on Fb Twitter
