Chip guru Jim Keller has expressed confidence that more companies will emerge to fill gaps in the AI market, as even Nvidia cannot meet all the demand.
Keller, CEO of the Canada-based AI chip design firm Tenstorrent, has observed that due to the high cost of High Bandwidth Memory (HBM), many companies, including his own, are seeking alternative solutions.
According to Nikkei Asia, Keller highlighted that as AI applications expand into smartphones, electric vehicles, and cloud services, more firms are looking for more affordable solutions. He believes the market leaves ample room for new competitors even with Nvidia supplying a big chunk of the market.
Alternatives to HBM
Keller emphasized that predicting the best application for AI, or its scale, is challenging. Thus, Tenstorrent's strategy is to develop technology adaptable to a wide range of products.
The company's Galaxy system claims to operate at three times the efficiency of Nvidia's DGX while costing 33% less. This is partly achieved by the factor of not using HBM.
He noted that HBM designs are time-consuming and expensive, and many companies are looking for more suitable memory solutions. Despite the existing scale of the HBM industry, he anticipates that any market impact from firms seeking alternatives will take years to materialize.
In typical AI chip setups, GPUs require high-speed data transfer capabilities of HBM when executing processes. Tenstorrent's chip design significantly reduces this transfer demand.
Keller believes that in certain AI development areas, Tenstorrent's chips could replace the roles traditionally played by GPUs and HBM.
Less is more
Keller has consistently addressed the high cost of AI hardware, seeing it as an entry point for startups like Tenstorrent to challenge giants like Nvidia. He previously suggested that Nvidia could have saved US$1 billion in developing its Blackwell GPU by using Ethernet interconnect technology.
Wei-han Lien, Tenstorrent's Chief CPU Architect, shared similar views with DIGITIMES Asia previously, emphasizing the firm's more pragmatic and economical ethos leads to more cost-effective system designs with superior computing power than Nvidia's.
Lien stated, "Customers do not need to pay Porsche or Ferrari prices to run their Generative AI models; they simply need the most cost-effective cars that run as fast." Lien anticipates that the current trend for expensive hardware will fade and the market will eventually stabilize.
