NVIDIA announces Blackwell Ultra GB300 at GTC 2025: 288 GB HBM3e and up to 15 PFLOPS dense FP4

Picture a supercomputer packed into a single chip the size of a serving tray. That is what NVIDIA announced on 18 March 2025 at its GTC conference with the GB300 Blackwell Ultra: not a graphics card for gaming, but a processor built to run large AI models.

The headline number is memory: 288 gigabytes sitting directly on the chip, up from 192 GB in the previous generation. For scale, most smartphones have 8-12 GB of RAM in total. NVIDIA got there by stacking the memory modules taller — twelve layers instead of eight — gaining 50% more capacity in the same footprint.

It is worth saying what did not change, because that part usually gets lost in the headlines. The speed at which the chip reads its own memory stays the same as before, 8 terabytes per second. And math in the formats most used for training is not any faster: it is identical to the previous generation. The roughly 50% speed gain applies only to a very compact number format called FP4, the one used to serve model responses cheaply. In short: this is a GPU with more memory, not a GPU that is faster at everything.

The link between cards, NVLink at 1.8 terabytes per second, is also not new here: it was already on the previous Blackwell generation. The real jump, up from 900 gigabytes per second, came with the move from Hopper to Blackwell.

Why does it matter outside the datacenter? Because more memory per chip means a large model fits on fewer cards, and that lowers the cost of every answer. First systems are expected in the second half of 2025. On the competitive side, AMD's MI355X claims the same 288 GB and the same 8 TB/s of bandwidth: on memory the two vendors are level, and NVIDIA's advantage has to be looked for elsewhere — in FP4 compute and in the maturity of its software stack.