Tenstorrent's CEO, Jim Keller, has revealed an updated RISC-V product roadmap and how he plans to tackle NVIDIA's AI dominance using CPUs and not GPUs.
Tenstorrent Is Going To Tackle NVIDIA's AI Dominance By Leveraging RISC-V & Chiplet Tech For AI CPUs
During a recent conference in Tokyo Japan as part of the "RISC-V Days Tokyo 2023 Summer", the well-known and legendary Jim Keller revealed his plans to push the AI industry forward using the RISC-V architecture and utilizing chiplet designs. Jim pointed out that while NVIDIA and AMD have been focusing on GPUs for AI advancements, he believes that CPUs designed specifically to run AI workloads can achieve higher efficiency and do a much better job than traditional GPUs.
"NVIDIA's GPU responded to the timing when the trend of AI greatly increased. However, GPUs are not optimized for AI processing. We have proven that our AI accelerator exceeds in efficiency. Also, one of the issues with AI processing by GPU is that access to CPU and memory becomes a bottleneck.The generation of Blackhole migration, which integrates RISC-V as a CPU, eliminates these bottlenecks. As a result, more power-saving and low-cost AI computers can be realized.”
Raja Koduri, the chief architect behind AMD's RTG and Intel's EVP & head of GPU division also left to join the Board of Directors at Tenstorrent.
The rationale put behind this is the fact that GPU prices and power consumption has been going up and that's not sustainable for the AI industry itself. A CPU can integrate some of the technologies used by GPUs and RISC-V can also bypass the expensive licenses that are often associated with the use of Intel, AMD, and ARM chips. It is also revealed that while CUDA remains to be dominant in the AI and HPC field, the BUDA compiler can also be used to run some of the high-level languages such as PyTorch and Java that are widely used in the AI domain while a low-level programming language is also in the works in the form of BUDA M (M for Metal) which will essentially be a replacement for CUDA.
"Many developers of AI are programmed from high-level languages such as PyTorch and Java. We believe that our current compiler “Buda” can be used to run , and the demo is already running.On the other hand, data scientists need low-level programming like CUDA. We would like to respond by providing developers with Buda M (M means metal)."
Showcasing its hardware roadmap, Tenstorrent revealed that they currently plan of offering RISC-V and AI chips in the form of heterogenous and chiplet designs. The company has already introduced GraySkull and Wormhole which featured a 12nm process node and compute capabilities of up to 328 TFLOPs (FP8) with 200 GB/s scale-out Ethernet. The company will be launching its Black Hole "Standalone ML Computer" chip which will be using a 6nm process node and feature the SiFive RISC-V X-280 architecture for heterogeneous computing. This chip is expected to launch in 2023.
Next up, the company is planning to move to a chiplet design with its Quasar Low-Power and Low-Cost ML design. This chip will utilize a full-on chiplet architecture and can be scaled up depending on the customer's needs. Blackhole will be integrating RSIC-V-based CPUs within AI accelerators and Tenstorrent plans to release its first-gen RISC-V-based CPU by the end of this year.
But the big guns will be launching in 2024 in the form of "Grendel" which is a highly configurable and performant ML chiplet design that will be combining the CPU chiplet with dedicated ML/AI chiplets in a similar fashion to NVIDIA's GH200 and the Grace/Hopper Superchips implementation. The CPU side is shown to feature 16 cores while the AI chiplets pack 40 cores each for a total of 80 cores. The AI chiplet block features four memory chiplets and the same goes for the CPU side the whole chip is connected to an IO chiplet which will be responsible for PCIe & other capabilities.
Tenstorrent also unveiled its RISC-V O-o-O processor family which includes various range of IPs. The company is also developing its second generation RISC-V-based CPU and will also deploy new features such as Vector Arithmetic units into the upcoming chips.
Tenstorrent further announced a brand new collaboration with LG which will be using RISC-V's AI technologies to power its AI needs within its future product lineup. In addition to this, the EU and the Japanese market are already developing a range of systems with RISC-V architecture and it is also seeing increased adoption within the Japanese automobile market. Besides this, Jim also presented the same projection slide featuring AMD's Zen 5 and NVIDIA Grace performance from before which you can read more on here.