When Nvidia recently pulled the wraps off its next-generation data center GPU codenamed Blackwell, the tech world became excited about its potential performance. That is decidedly not the case for Jim Keller, a renowned chip designer responsible for laudable chip designs with both Apple and AMD in the past. He has taken to Twitter to state Nvidia's $10 billion R&D costs for Blackwell could have been slashed to $1 billion if it had used one weird trick.
Keller's tweet states that Nvidia shouldn't have used its propriety interconnect technology, NVLink. Instead, it should have gone with an open standard like Ethernet, which would have allowed it to save $9 billion in upfront costs. Of course, Nvidia would likely argue it will surely recoup those costs in Blackwell sales, as NVLink requires Nvidia hardware, so it'll make more in the long term despite the hefty development costs.
Tweet (opens in a new tab)
Nvidia would claim that its custom interconnect is a better solution than Ethernet in performance and scalability, especially since 5th-generation NVLINK affords 1.8TB/s interconnect bandwidth. Plus, Nvidia currently controls the data center GPU market, so it's not about to begin using open standards that will allow companies to buy hardware from its competitors.
Keller has previously criticized Nvidia's CUDA platform, according to Wccftech, so his latest salvo towards Team Green will be added to the list. Keller likely sees Nvidia as the antithesis of his own company, Tenstorrent, which promotes open-source hardware and software solutions.
What's notable about his comments is that Intel recently announced an Nvidia competitor in the AI space named Gaudi 3, which uses Ethernet for its interconnect technology. In its announcement, Intel boasted that customers choosing Gaudi 3 will avoid "vendor lock-in," referencing Nvidia's NVLinktechnology. It certainly seems like Intel and Keller see eye-to-eye on this issue.
Tagged In
Data CentersMore from Computing
