NVIDIA spent a sizeable $10 billion on R&D for its next-generation Blackwell GPU architecture, but chip legend Jim Keller said he could’ve done the same job for just $1 billion.
I first saw Keller’s tweet, noticing that it looked familiar… he had taken an image of the story I wrote about NVIDIA spending $10 billion on R&D for Blackwell. Great to see, but how does Keller fix NVIDIA’s $10 billion R&D budget for one-tenth of that cost, just $1 billion?
First, NVIDIA uses dual Blackwell B100 dies on its new B200 AI GPU, which has a whopping 208 billion transistors, each B100 GPU die featuring 104 billion transistors. NVIDIA uses two B100 dies to create the B200 in full, with NVIDIA using its in-house NV-High Bandwidth Interface (NV_HBI), which has up to 10TB/sec of bandwidth.
Keller says that ethernet could’ve solved this huge $10 billion research and development effort on Blackwell, where instead of using NV-High Bandwidth Interface, ethernet in its place would’ve worked. From there, NVIDIA uses its 5th Generation NVLINK technology, which is proprietary — in-house for NVIDIA — so while it has mammoth bandwidth, latency, and scalability, it’s a proprietary product and doesn’t work with other manufacturers’ hardware.
NVLINK might be super-fast, but it has its boundaries: it is great with NVIDIA hardware and works with nothing else. Jim Keller sways into the open-source side of the business, so this could be a hint towards using Ultra Ethernet in the future, a new standard on its way to feed the high-bandwidth and low latency requirements of ethernet-based solutions for AI and HPC systems using Ultra Ethernet in the future.