The launch of two inference cards speeds up AI processing on 64-bit x86 hosts with RISC-V architecture, supporting advanced software stacks for development.
Machine learning acceleration specialist Tenstorrent has announced two new Grayskull inference cards, the Grayskull e75 and e150, for speeding up inference on 64-bit x86 hosts with PCI Express. Both cards support Tenstorrent’s software stacks, TT-Buda and TT-Metalium.
The company claims the dev kit features its first-generation AI PCIe card for inference alongside the TT-Metalium software stack.
The e75 and e150 cards are designed for inference acceleration and have suitable training technology. Each card’s processor has a grid of Tensix cores based on the RISC-V instruction set architecture, with a tensor array math unit, SIMD unit, and dedicated hardware accelerators for network operations and compression/decompression.
The e75 card has 96 Tensix cores running at 1GHz, 96MB of SRAM, and 8GB of LPDDR4 memory. It connects to a host via a 16-lane PCIe Gen. 4.0 link, with a power consumption of 75W, requiring an active cooling kit.
The e150 card features 120 Tensix cores at 1.2GHz, 120MB of SRAM, and 8GB of LPDDR4 memory, with a peak transfer rate of 118.4GB/s. It requires 200W at full load. Both cards are compatible with 64-bit x86 hosts running Ubuntu 20.04 LTS, with at least 64GB RAM and 100GB storage, though 2TB is recommended. The e75 needs a power supply with a six-pin PCIe connector, while the e150 requires a six-pin and a six-plus-two-pin connector.
The company offers two software stacks: TT-Buda for executing models on Grayskull hardware and TT-Metalium, a low-level programming framework for building models from scratch or experimenting with non-machine-learning workloads on the Tensix cores.
“Today, we are officially launching our Grayskull Dev Kit, available for purchase on our website,” Tenstorrent wrote in its announcement, which was brought to our attention by Linux Gizmos. This is our first-gen AI [Artificial Intelligence] PCIe [PCI Express] card—an inference-only hardware kit we are releasing alongside TT-Metalium, our open—source software stack.”
