Tenstorrent: Jim Keller wants to compete with NVIDIA’s AI GPUs with RISC-V CPUs

Jim Keller may still be best known to some for his role in reviving AMD’s CPU division. He is also the CEO of AI startup Tenstorrent, which recently published its updated product roadmap as part of its RISC-V Days Tokyo 2023 Summer introduced and wants to break NVIDIA’s GPU AI dominance with the help of RISC-V CPUs.

During the Tokyo conference, Keeler explained the current situation. According to them, NVIDIA got the timing right for its AI products just as the trend was going fast. However, according to Keeler, the manufacturer’s GPUs should not be greatly optimized for AI processing. One of the main problems with using GPUs is that they become bottlenecks when accessing the CPU and memory. Keller claims that he has already come up with a solution to this problem using his latest chip – the Black Hole. Said bottlenecks should be avoided in this way and more power-saving and cost-effective AI computers can also be achieved in this way.

Presenting its hardware roadmap, Tenstorrent revealed its future plans to offer RISC-V and AI chips in both heterogeneous chip form and chip design. After GraySkull and Wormhole, which have a chassis width of 12nm, Black Hole follows at 6nm with the SiFive RISC-V X-280 architecture as an “independent machine learning chip”, which will be launched later this year.

The company’s next big step is to introduce Quasar’s “low power, low cost” ML design, which is intended to transition into the chiplet design. Thanks to the chiplet architecture, the performance can be precisely adapted to the respective customer requirements. Then in 2024, the company plans to use Grendel, a powerful and highly configurable ML chip design. With this, the company wants to integrate its CPU chips with dedicated ML/AI chips, similar to NVIDIA’s GH200 with Grace/Hopper Superchips implementation.