Start-up AlphaICs yesterday announced that it has begun sampling its Gluon coprocessor for edge AI inference to customers. AlphaICs claims the chip delivers competitive performance compared to Nvidia in vision workloads such as object detection.
Gluon is based on AlphaICs’ proprietary architecture that has an instruction set architecture (ISA) optimized for AI, and has been in development for two years. The ISA refers to the instructions that a chip can process. AlphaICs calls it Real AI Processor (RAP) architecture. AlphaICs further has a SDK with support for TensorFlow. The company reports that other frameworks are in development.
A theoretical disadvantage isn’t holding Gluon back
As the company’s first chip based on the RAP architecture, Gluon has 8 TOPS — a metric that means trillions of operations per second, or “tera operations per second” — of compute capabilities for deep learning neural networks. Although this isn’t particularly high, and comparable to (or even lower than) other edge AI and smartphone chips, AlphaICs claims it has higher utilization (higher performance per TOPS), resulting in higher performance. Gluon further supports PCIe and LPDDR4 interfaces. The chip itself is manufactured on TSMC’s 16nm process, which puts it at a theoretical disadvantage compared to other chips that may have already moved to more advanced nodes such as 7nm or even 5nm.
Nevertheless, AlphaICs provided some benchmarks to prove the chip’s capabilities, as it claims Gluon has higher performance per watt and performance per TOPS than Nvidia’s edge chips. In Yolo-V2, Gluon reportedly achieves 153 frames per second while consuming just below 5W, and 308 frames per second in ResNet-50 at 3.5W.
For comparison, according to Nvidia, the Jetson TX2, which has a 7.5W TDP, achieves just 112fps in both Yolo-V3 and ResNet-50. The Jetson Xavier NX, which is rated at 10-20W and 21 TOPS, achieves 1100fps in ResNet-50.
As a coprocessor targeted at the edge segment, where power is often a constraint, Gluon is aimed at vision workloads such as classification and detection. Segments where these capabilities are applied they include surveillance, industrial, retail, and smart cities.
AlphaIC is based in Milpitas and raised $8 million a year ago. The company was founded in 2016 by Vinod Dham, a former Intel executive known as the “father of the Pentium.”
0 Comments
Enregistrer un commentaire