NVIDIA CEO says the company needs to 'create more resilience in our supply chain' for AI GPUs

NVIDIA has finally unveiled its next-gen Blackwell B200 AI GPU and new DGX systems for the future of AI, with CEO Jensen Huang saying that his company needs to focus on both compliance and resilience when it comes to serving the Chinese market.

3

Jensen noted that many parts of NVIDIA's systems are made in China, saying "that's just the truth," and that NVIDIA needs to do everything in its power to create more resilience in the supply chain. Jensen said: "There are two things we have to do right away. One is to make sure we understand the policies and comply. And the other is to do everything we can to create more resilience in our supply chain".

Jensen noted that many of NVIDIA's systems use parts made in China, where the NVIDIA CEO added: "That's just the truth," adding that the same thing is true for both the auto and defense industries. Jensen continued: "The supply chain of the world is quite complicated. We have every confidence that the goal of the nations is not adversarial... the doomsday scenario is not likely to happen".

• Read more: NVIDIA spent $10 billion on developing its next-generation Blackwell GPU
• Read more: NVIDIA CEO: our next-gen Blackwell GPU will cost $30,000 to $40,000 each
• Read more: NVIDIA GB200 Grace Blackwell Superchip: 864GB HBM3E, 16TB/sec bandwidth
• Read more: NVIDIA Blackwell AI GPU: multi-chip die, 208B transistors, 192GB HBM3E
• Read more: NVIDIA's new Blackwell-based DGX SuperPOD: ready for trillion-parameter scale for generative AI

Now, as for NVIDIA's new Blackwell B200 AI GPU: the new NVIDIA B200 AI GPU features a whopping 208 billion transistors made on TSMC's new N4P process node. It also has 192GB of ultra-fast HBM3E memory with 8TB/sec of memory bandwidth. NVIDIA is not using a single GPU die here, but a multi-GPU die with a small line between the dies differentiating the two dies, a first for NVIDIA.

The two chips think they're a single chip, with 10TB/sec of bandwidth between the GPU dies, which have no idea they're separate. The two B100 GPU dies think they're a single chip, with no memory locality issues and no cache issues... it just thinks it's a single GPU and does its (AI) thing at blistering speeds, which is thanks to NV-HBI (NVIDIA High Bandwidth Interface).

3

NVIDIA's new B200 AI GPU has 20 petaflops of AI performance from a single GPU, compared to just 4 petaflops of AI performance from the current H100 AI GPU. Impressive. Note: NVIDIA is using a new FP4 number format for these numbers, with H100 using the FP8 format, which means that B200 has 2.5x theoretical FP8 compute than H100. Still, very impressive.

Each of the B200 GPU dies two full reticle-size chips, with 4 x HBM3E stacks of 24GB each, along with 1TB/sec of memory bandwidth on a 1024-bit memory interface. The total of 192GB of HBM3E memory, with 8TB/sec of memory bandwidth, is a huge upgrade over the H100 AI GPU, which had 6 x HBM3 stacks of 16GB each (at first, H200 kicked that up to 24GB per stack).

NVIDIA is using an all-new NVLink chip design that has 1.8TB/sec of bi-directional bandwidth and packing support for a 576 GPU NVLink domain. This NVLink chip itself features 50 billion transistors, manufactured by TSMC on the same N4P process node.