What is the best measure of inference performance?
Modest performance improvements on BERT, more dramatic on ResNext.
The board delivers 1.4 TOPS/W for drones and small robots.
AI Hardware Summit 'No-Show', Groq, emerges from stealth mode.
And do you even really need one?
DeepScale argues AI chips and sensor systems should come with DNNs optimized for the application. It will fundamentally change how companies buy AI te...
Areanna claims a custom SRAM will deliver 100 TOPS/W on deep learning. Well, that's what the simulations show, anyway.
Habana Labs chip claims industry-best throughput on ResNet-50 training, includes on-chip RoCE communications for scalability