NVIDIA's Blackwell joins two chips into one giant GPU for AI
Unveiled at GTC 2024, the B200 packs 208 billion transistors by linking two dies so tightly they act as a single chip.
In 30 seconds
- Blackwell is NVIDIA's GPU architecture after Hopper (H100).
- Each Blackwell GPU has 208 billion transistors across two dies joined by a 10 TB/s link.
- It is made on a custom TSMC 4NP process.
At its GTC conference in San Jose in March 2024, NVIDIA introduced Blackwell, the successor to the Hopper architecture behind the H100 chips that power much of today’s AI.
Two dies, one GPU
There is a physical limit to how big a single chip can be printed. NVIDIA’s answer was to build two of the largest possible dies and connect them with a link that moves 10 terabytes of data per second. Software sees the pair as one GPU. Together they hold 208 billion transistors.
Built for huge models
NVIDIA also announced the GB200, which pairs two Blackwell GPUs with its Grace CPU, and rack-scale systems that connect dozens of these into one machine. The target is training and running AI models with trillions of parameters.
Why packaging matters
The speed of the connection between dies depends on advanced packaging, the technology used to put several pieces of silicon together. This step, once seen as routine, is now one of the most important and scarce parts of the chip supply chain.
Why it matters for India
The IndiaAI Mission plans to give Indian startups and researchers access to thousands of GPUs. The chips in those data centres, and their price and power use, decide how many Indian AI models can actually be trained at home.
Our take
Joining two dies is a sign that single chips are hitting their size limits. Expect more of the industry to move to multi-die designs, which is good news for advanced packaging, an area India is now investing in.
Sources
We read these before writing this story. Go to them for the full detail.