Jim Keller, the legendary chip engineer and current CEO of Tenstorrent, recently commented on NVIDIA’s US$10 billion (~RM47.8 billion) R&D cost of Blackwell. He thinks that the company could have saved itself US$9 billion (~RM43 billion), had it used a more “conventional” method.
By that, Keller is essentially saying that instead of using its own proprietary NVLink interconnect technology to develop Blackwell, it could have reach the achieved the same goals by just using Ethernet. Of course, NVIDIA would diligently argue that it should be able to recoup that cost in Blackwell sales, given that NVLink requires…well, it’s own hardware.
— Jim Keller (@jimkxa) April 11, 2024
But there’s a reason why Keller even made the comment about Ethernet in the first place. Unlike NVLink, Ethernet is ubiquitous, both on a hardware and software level. Specifically the next-generation of it can achieve speeds of up to 800 GbE, while the NVIDIA’s current NVLink interconnection has a peak bandwidth of 200 GbE.
However, as pointed out by Tom’s Hardware, the NVLink technology, InfiniBand, does have some advantages, especially when it comes to AI, HPC, and some near God-tier tail latencies. That said, other players in the industry, including AMD, Broadcom, Intel, Meta, Microsoft, and Oracle, are also developing the Ultra Ethernet interconnection technology, which as its name suggests, should offer a much higher throughput and features for AI and HPC communications.
(Source: Tom’s Hardware, Tweaktown, ExtremeTech)
