Jim Keller, the CEO of Tenstorrent, has recently made waves with his critique on NVIDIA’s spending for the development of their Blackwell GPUs. Keller alleges that NVIDIA’s research and development expenditures, which amount to a staggering $10 billion, could have been drastically lowered to around $1 billion with a strategic change in their interconnect approach.
NVIDIA has implemented its 5th Generation NVLINK technology for the Blackwell GPU interconnects, praised for its superior bandwidth, low latency, and scalability. However, criticism arises from NVLINK’s proprietary nature. As it is only compatible with NVIDIA’s hardware, it limits broader industry adoption and stands as a costly choice in a competitive market.
Keller is an advocate for open-source technology and challenges the use of proprietary solutions like NVLINK. He suggests that Ethernet could serve as a more cost-effective and open interconnect method that fosters an expansive ecosystem, though it may presently lag in performance. Yet, with advancements such as the Ultra Ethernet or UE in development, which is expected to boost AI and High-Performance Computing (HPC) capabilities, Ethernet could emerge as a strong contender, possibly shifting industry preferences.
While Keller is known for his outspoken critiques of tech companies, particularly NVIDIA, his perspective on cost efficiency raises a valid point. For an industry that is rapidly evolving, where innovation and cost management are crucial, exploring alternative, open-source technologies could lead to substantial savings and foster a more collaborative tech environment.
Despite Keller’s comments, the popularity of NVIDIA’s Blackwell AI GPUs is undeniable. Demand for these chips is high, and even major players like Intel and AMD are contending to match NVIDIA’s market stronghold. Their dominance has been further solidified with the introduction of Blackwell, maintaining their competitive edge in the tech sphere.
