Esperanto Technologies, a San Francisco startup founded by Dave Ditzel, stepped onto the stage this year HotChips 33 presentation Show off its latest processor. The company codenamed ET-SoC-1, provided one of the first thousand-core designs to the public.
The Kilocore processor design assumes that there are a thousand processor cores. The Esperanto method involves this. Building a chip with a thousand cores requires Esperanto carefully consider its application and how the chip will operate.
There are 1088 smaller, low-power ET-Minion 64-bit RISC-V cores designed to run in parallel and easily execute workloads. In order to cooperate with the small core, Esperanto has packaged additional 4 high-performance ET-Maxion 64-bit RISC-V cores. Smaller cores are designed for sequential execution, while larger cores are designed for out-of-order execution. The entire chip is taped out using a 7nm manufacturing node in TSMC’s factory.
For smaller cores, they packaged a dedicated vector/tensor processing unit to accelerate AI workloads. These cores are designed to have a smaller number of gates/stages to achieve better frequencies at lower voltages, thereby increasing efficiency. There are two hardware threads for execution, so the ET-Minion core has SMT. This design runs in an orderly execution mode, yields the highest efficiency and does not waste instruction cycles. The operating frequency range of the ET-Minion core is 300MHz to 2GHz.
The larger ET-Maxion core is designed for an uncompromising RISC-V ISA experience. With all the functions of privileged and compressed ISA, if configured like this, these kernels can run independently. The kernel is executed out of order, and the operating frequency ranges from 500MHz to 2GHz.
ET-SoC-1 is designed with more than 160 million bytes of on-chip SRAM, which is divided into L1 cache and scratchpad memory. It supports PCIe Gen4, 256-bit wide (quad channel) LPDDR4x memory, and has a RISC-V service processor for secure boot, system management, and timers. All of these are packaged in a 570 square millimeter chip, which contains 22.5 billion transistors. In normal operation, it uses less than 20 watts of power, and it can be software configured for 10-60 watt workloads.
Esperanto pointed out that this “on-chip supercomputer” is suitable for various workloads, not just AI/ML, because it runs on a general-purpose RISC-V core. This means that every application that requires advanced parallel processing can benefit from using this chip, and its low power consumption is suitable for a variety of applications.
How it compares to the competition is the real question. Thousand-core CPUs with tensor processing capabilities begin to look a lot like custom chips for artificial intelligence. The four high-performance cores may not keep up with AMD and Intel, or even Arm’s modern CPU designs.