Tenstorrent, the company led by legendary chip architect Jim Keller, the brains 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 runtime sparsity and conditional calculations.
Following this, Tenstorrent 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 on the fly, while the latter is for users who want to customize their models or write new ones.
The Santa Clara-based technology company’s milestone launch follows 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, running at 75W. The more advanced e150 model is a standard height 3/4-length PCIe Gen 4 board and includes one Grayskull processor that runs at up to 200W and provides a balance between power and throughput.
Tenstorrent processors consist of a grid of cores known as Tensix Cores and come with network communications hardware so they can talk to each other directly over networks, rather than 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 to purchase now for $599 and $799 respectively.