“These are intriguing times,” he said. The announcement of Blackwell makes for interesting conversation around scale. Blackwell and the DGX generation 2 system probably won’t be generally available for quite some time. So it makes for an interesting acquisition strategy: ‘Do I acquire the Hopper base system – the H1o0, Bluefield etc – today and get use out of those systems today and then rely on the fact that these dual die systems are seen as a single Hopper system and I can develop or design and build my infrastructure and products around these systems?’
There is some debate around the performance delta between the systems, he continued. “Nvidia was unclear on the numbers but most people are selling on the 30x performance from one system and another, what I have been able to practically see are that some of the performance tests are proving that in some cases there is a 5x improvement in training times for systems using these Blackwell processors.”
Nvidia is moving towards both FP8 (8-bit quantisation or vectorisation) to FP4 which uses a different technique, explained Townsend, which makes it difficult to quantify and validate the 30x performance. “It is subjective . . . both the models and software running the models are constantly changing,” he added.
The total addressable market for AI chips is estimated to be $205bn. “That’s an incredibly large market,” said Townsend, “However, keep in mind that AI chips and the servers they run in are only the pick and shovels of the AI gold rush. The value is the productivity and new business models AI enables.”
