Axelera AI's Compact M.2 Accelerator Promises 214 TOPS of Compute at Just 15 TOPS/W

High-performance machine learning specialist Axelera AI has announced a gumstick-style M.2 accelerator based on its Metis artificial intelligence processing unit (AIPU), which it claims can add 214 tera operations per second (TOPS) to Next Generation Form Factor (NGFF)-compatible devices — along with a beefier PCI Express card delivering a claimed 856 TOPS of compute.

"Axelera’s Metis AIPU is equipped with four homogeneous AI cores built for complete neural network inference acceleration," explains Axelera's chief technical officer Evangelos Eleftheriou. "Each AI core is self-sufficient and can execute all layers of a standard neural network without external interactions. The four AI cores can either collaborate on a workload to boost throughput or operate on the same neural network in parallel to reduce latency or process different neural networks required by the application concurrently."

Each "AI core," which use the free and open RISC-V instruction set architecture (ISA) in a control core designed to handle data flow to and from the accelerator, can offer a claimed 53.5 TOPS of compute — and with four per AIPU, that's a total of 214 TOPS at a claimed 15 TOPS/W power efficiency at full utilization.

To prove the concept, Axelera is preparing to launch two add-in accelerators for existing devices, as brought to our attention by CNX Software. The first is an NGFF M.2 gumstick-style card, which uses a single Metis AIPU for 214 TOPS of compute performance — and a claimed 3,200 frames per second (FPS) in the ResNet-50 model. Should that prove too little, the company is also launching a full-length PCI Express (PCIe) card with four Metis AIPUs — meaning 856 TOPS and 12,800 frames per second in ResNet-50.

The company has also teased an "AI vision gateway," which includes a single-board computer with dual Ethernet ports and a choice of Intel or AMD processors along with the Metis M.2 accelerator card — though, oddly, Axelera only claims a performance of "over 100 TOPS," rather than the full 214 TOPS of the accelerator card on its own. This, however, is still enough by the company's reckoning to drive "small to medium sized smart camera deployments in robotics and retail" — defined as being up to eight cameras.

All three devices are due to go up for pre-order "in early 2023," with the M.2 accelerator priced at $149 and the PCIe version "from $499"; pricing for the gateway has not yet been confirmed. More information, and a sign-up form to be notified when orders open, is available on the Axelera website.