Tenstorrent, the firm led by legendary chip architect Jim Keller, the mastermind behind AMD’s Zen architecture and Tesla’s original self-driving chip, has launched its first hardware. Grayskull is a RISC-V alternative to GPUs that is designed to be easier to program and scale, and reportedly excels at handling run-time sparsity and conditional computation.
Off the back of this, Tenstorrent has also unveiled its Grayskull-powered DevKits – the standard Grayskull e75 and the more powerful Grayskull e150. Both are inference-only hardware designed for AI development, and come with TT-Buda and TT-Metalium software. The former is for running models right away, while the latter is for users who want to customize their models or write new ones.
The Santa Clara-based tech firm’s milestone launch comes hot on the heels of a partnership with Japan’s Leading-edge Semiconductor Technology Center (LSTC). Tenstorrent’s RISC-V and Chiplet IP will be used to build a state-of-the-art 2nm AI Accelerator, with the ultimate goal of revolutionizing AI performance in Japan.
By the power of Grayskull!
The Grayskull e75 model is a low-profile, half-length PCIe Gen 4 board with a single Grayskull processor, operating at 75W. The more advanced e150 model is a standard height, 3/4 length PCIe Gen 4 board containing one Grayskull processor operating at up to 200W, and balancing power and throughput.
Tenstorrent processors comprise a grid of cores known as Tensix Cores and come with network communication hardware so they can talk with one another directly over networks, instead of through DRAM.
The Grayskull DevKits support a wide range of models, including BERT for natural language processing tasks, ResNet for image recognition, Whisper for speech recognition and translation, YOLOv5 for real-time object detection, and U-Net for image segmentation.
The Grayskull e75 and e150 DevKits are available for purchase now at $599 and $799, respectively.