Esperanto’s ET-SoC-1 Chip Integrates more than 1000 RISC-V Cores for Energy Efficient ML Recommendation

More than 1000 cores on a commercial chip – seems unbelievable!!!

However, the highly sophisticated manufacturing processes, unbelievably optimized design integration, and the RISC-V open-source hardware revolution have made this possible.

Esperanto‘s ET-SoC-1 Chip is a Supercomputer-on-Chip that integrates more than 1000 RISC-V cores on a single die of 570 mm² area. Dave Ditzel, Founder and Executive Chairman, Esperanto Technologies detailed about the ET-SoC-1 chip in the Hot Chips 33 conference and highlighted that the package could be used as a compelling energy-efficient solution for Machine Learning recommendation in large data centers.

The Highlights of the ET-SoC-1 Chip

• Integration of over a thousand RISC-V processors on a single TSMC 7nm chip
• Integration of 24 billion transistors on a single die
• 1088 energy-efficient ET-Minion 64-bit RISC-V in-order cores, each with a vector/tensor unit (operating frequency 500 MHz to 1.5 GHz expected)
• 4 high-performance ET-Maxion 64-bit RISC-V out-of-order cores (operating frequency 500 MHz to 2 GHz expected)
• 1 RISC-V service processor
• Over 160 million bytes of on-die SRAM are used for caches and scratchpad memory
• Interfaces for large external memory with low-power LPDDR4x DRAM and eMMC FLASH
• PCIe x8 Gen4 and other common I/O interfaces
• Compute at peak rates of 100 to 200 TOPS (Tera Operations per Second)
• Power typically < 20 watts

The manufacturer claims that ET-SoC-1 is the highest-performance commercial RISC-V chip to date that provides the highest TOPS driven by RISC-V cores. It delivers the best performance per watt, which makes it highly suitable for energy-efficient acceleration for data center inference workloads (ML recommendation).

The energy efficiency of ET-SoC-1 chip (<20 watts) enables the integration of up to six such chips (3 on top and 3 on bottom) onto a single Glacier Point v2 accelerator card that can operate with a 120-watt limit. The accelerator card provides up to 192 GB of DRAM and 822 GB/s DRAM bandwidth. Esperanto estimates that ET-SoC-1 card provides 123x better performance per watt for ML Recommendation and 25.7x better performance per watt for Image Classification.

The introduction of ET-SoC-1 chip to the semiconductor industry, along with the open-source hardware revolution, provides a roadmap for even more sophisticated integrations in the near future.

Note: The details about ET-SoC-1 chip are extracted from the presentation of Dave Ditzel at Hot Chips 33.