The Grayskull Dev Kit comes in two versions: the Grayskull e75 priced and the more powerful Grayskull e150. Both kits are inference-only hardware designed for AI development, leveraging Tenstorrent’s unique software approaches: TT-Buda and TT-Metalium.

The e75 DevKit is a low-profile, half-length PCIe Gen 4 board featuring a single Grayskull processor, operating at 75W. This entry-level board, suitable for foundational AI work, includes 96 Tensix cores, 96MB SRAM, and 8GB LPDDR4 memory. It provides a diverse range of software capabilities, aiming to offer ML developers an alternative to traditional GPUs.

Grayskull e75 front view
(click image to enlarge)

On the other hand, the e150 DevKit, with its standard height and 3/4 length PCIe Gen 4 board, contains one Grayskull processor operating at up to 200W. This model is designed for balancing power and throughput, featuring 120 Tensix cores, 120MB SRAM, and the same memory configuration as the e75.

Both versions come with TT-Buda and TT-Metalium software. TT-Buda allows production customers to run models immediately without the need for programming new operations. TT-Metalium provides open access to Tenstorrent’s hardware and software, ideal for developers who wish to customize or write new models.

Tenstorrent TT-Buda & TT-Metalium
(click image to enlarge)

Grayskull Dev Kits support a wide range of models, including BERT for natural language processing, ResNet for image recognition, Whisper for speech recognition and translation, YOLOv5 for real-time object detection, and U-Net for image segmentation.

Tenstorrent indicates that o utilize these devices, certain system requirements must be met, including an x86_64 architecture CPU, a motherboard equipped with a PCIe 4.0 x16 slot, at least 64GB of memory, and a minimum of 100GB of storage, although 2TB or more is recommended. Additionally, specific power connectors are required, and the system must run on Ubuntu version 20.04 operating system.

Grayskull e150 front view
(click images to enlarge)

Tenstorrent’s FAQ section explains that their Tensix Cores, comprising five RISC-V processors and specialized units for tensor and vector operations, offer unique capabilities and up to 1.5 MB of SRAM per core. They emphasize the efficiency of their ‘Software Enabled TOPS’ over traditional ‘Silicon TOPS’, attributing higher throughput to features like dynamic sparsification.

Differentiating from GPUs, Tenstorrent’s processors are easier to program, more scalable, and excel in direct network communication and handling tensor computations efficiently.

Further Information

The Grayskull e75 and e150 models are available for purchase at $599 and $799, respectively, on Tenstorrent’s website. Additionally, the company has announced on Twitter the release of their open-source software stack, TT-Metallium, alongside the hardware launch.