Tenstorrent's second-generation multi-purpose AI processor is expected to be launched before 2025.
According to recent estimates, Nvidia dominates the artificial intelligence (AI) processor market, controlling over 80% of sales. However, the legendary processor designer and current CEO of Tenstorrent, Jim Keller, believes that Nvidia does not serve certain markets well. Therefore, Tenstorrent and other AI processor developers have opportunities.
"Nvidia does not serve many markets well," Jim Keller said in an interview with the Japanese media.
Jim Keller has a truly outstanding history in the computer industry, having made waves at AMD, Intel, and Tesla, and is now leading the development of AI processors at Tenstorrent. From 1998 to 1999, Jim Keller worked on the K7/K8 architecture that supported the Athlon at AMD. From 2008 to 2012, he led the development of the A4 and A5 processors at Apple. From 2012 to 2015, he was in charge of the K12 ARM project and the Zen architecture project at AMD. From 2016 to 2018, he developed the FSD autonomous driving chip at Tesla, and from 2018 to 2020, he participated in a mysterious project at Intel.
Jim Keller's leadership at Tenstorrent aims to provide an affordable alternative to Nvidia's expensive GPUs, which sell for $20,000 to $30,000 each or more. Tenstorrent's business approach aims to serve markets that Nvidia has not fully addressed, especially in the edge field. Tenstorrent claims that its Galaxy system is three times more efficient than Nvidia's DGX and 33% cheaper, which may be the world's most popular AI server.
It is reported that Tenstorrent is expected to release its second-generation multi-purpose AI processor before the end of this year, but the name of the processor has not been disclosed. According to Tenstorrent's latest roadmap last fall, the company intends to release its Black Hole standalone AI processor and Quasar low-power, low-cost small chip for multi-chip AI solutions.
The company claims that its upcoming processor offers performance efficiency comparable to Nvidia's AI GPUs. The touted efficiency and lower cost are partly achieved by avoiding the use of high-bandwidth memory (HBM) and instead using GDDR6, which is logical for entry-level and mainstream AI processors designed primarily for AI inference. At the same time, Tenstorrent stated that its architecture consumes less memory bandwidth than its competitors, which is a key reason for its higher efficiency and lower cost.Although Tenstorrent has not yet captured a significant share of the AI processor market, the company is currently focused on cost-effective and scalable AI solutions that can address a variety of applications that Nvidia is currently unable to serve properly. However, these markets will not be blue oceans, as many companies will attempt to tackle these markets with their products in the coming quarters, competing against the well-established Nvidia. Instead of directly competing with Nvidia, it is better to make it easier for new entrants to address niche markets that are not directly catered to by the "green team."
In early 2023, Canadian AI chip startup Tenstorrent recently announced a change in personnel, with CTO Jim Keller taking on the role of CEO.
Interestingly, the original CEO, Ljubisa Bajic, swapped roles with Keller and will serve as the new CTO.
Keller wrote on a social platform, "Ljubisa Bajic and I have decided to swap roles to continue developing artificial intelligence, CPU products, and more, in order to tackle the next generation of computing challenges."
Tenstorrent was founded seven years ago, and Bajic is one of the co-founders. Jim Keller met Bajic while working at AMD and became the first external investor in Tenstorrent, officially joining as CTO in January 2021.
During his tenure as CTO, Keller mainly focused on two tasks: first, bringing in the RISC-V architecture, on which the company's team developed the Ascalon CPU core. Second, he introduced David Bennett as the Chief Customer Officer, who is an AMD veteran and has also served as the president of Lenovo Japan and CEO of NEC's personal computer business.
However, the commercial products of Ascalon will still need to wait for some time, as Tenstorrent's new generation of BlackHole AI chips will be developed based on SiFive's X280 RISC-V core.
Jim Keller suggests that Nvidia should abandon proprietary standards.
Jim Keller is not only a chip design guru with outstanding technical skills but also a staunch supporter of open technology, always detesting closed technologies. Naturally, Nvidia has become his "enemy."Recently, Jim Keller has suggested that NVIDIA's latest Blackwell GPU should not use the proprietary NVLink standard protocol for multi-chip interconnect and network interconnect, but should switch to the open Ethernet standard, which could save NVIDIA tens of billions of dollars. He also believes that NVIDIA should not use its own InfiniBand solution in data center networks, but should switch to Ethernet as well.
Although NVIDIA's InfiniBand network has low latency and high bandwidth, reaching up to 200GbE, Ethernet can achieve 400GbE or even 800GbE.
Industry giants such as AMD, Broadcom, Intel, Meta, Microsoft, and Oracle are also collaborating to develop the next generation of ultra-high-speed Ethernet (Ultra Ethernet), which offers higher throughput and is more suitable for AI and HPC applications.
In addition, Jim Keller has always been dissatisfied with NVIDIA's closed CUDA ecosystem, once criticizing it as a swamp rather than a moat.
Arm, Intel, Qualcomm, and Samsung have also partnered to establish the Unified Acceleration Foundation (UXL), one of whose goals is to replace NVIDIA's solutions. For Huang Renxun, both NVLink and CUDA are the results of his company's investment of billions of dollars over the years and are the treasures for maintaining its own interests.
*Statement: This article is the original creation of the author. The content of the article is the author's personal opinion. We republish it only for sharing and discussion, and do not represent our approval or agreement. If you have any objections, please contact the backend.
Comment