Future-proof IP for training and inference with leading performance per watt and per dollar

Overview

Tenstorrent develops AI IP with precision, anchored in RISC-V’s open architecture, delivering specialized, silicon-proven solutions for both AI training and inference. Our platforms are optimized for high performance per watt, ensuring scalable and adaptable technology for all AI-specific workloads.

Tensix Cores allow AI developers to accelerate their specific AI networks and applications in an efficient and easy-to-use manner. Our platforms are optimized for high performance-per-watt, ensuring adaptable technology for all AI-specific workloads. Tensix Cores support a broad range of precision formats and connect using a specially designed Network-on-Chip (NoC), allowing your solution to scale in concert with an ever-expanding list of models and evolve with the industry. Tensix is silicon-proven with two generations of products - Grayskull™ and Wormhole™ - and a third generation in active development. Each Tensix Core incorporates a block of SRAM, five “baby RISC-V” cores, Matrix and Vector Engines, and dedicated hardware streams built upon ethernet protocols that facilitate rapid core-to-core and chip-to-chip communication.

Key Features

  • RISC-V IP Optimized for AI
    • Designed specifically for advanced AI/ML workloads.
  • Silicone-Tested
    • Solutions have been validated on silicon for AI workload efficiency.
  • Adaptable Technology
    • AI technology platforms that are scalable and customizable.
  • Quick to Market
    • AI IP that enables swift and flexible integration for rapid deployment.
  • Power-Efficient
    • Delivers top-tier performance with minimal power consumption.

Benefits

  • MMUL ops/cycle: 4096
  • SIMD ops/cycle: 64
  • Data types: INT, FP, BFP
  • SRAM: 1.5MB

Block Diagram

Future-proof IP for training and inference with leading performance per watt and per dollar Block Diagram

Applications

  • Automotive AI, Data Centers, HPC, Wearables, Smart Cameras, DTV, and More. Accelerating CNNs, LLMs, Transformers, and other AI workloads with high efficiency and utilization.

Technical Specifications

×
Semiconductor IP