Fractile Computes in Memory for AI Inference
Startup Fractile claims its data-center NPU will be 100× faster than an Nvidia H100 at 1/10th the cost. The secretive company indicates it's employing in-memory computation to improve MAC-unit utilization and power efficiency.
Other contents
Esperanto Technologies Exits AI Chips
Indian startup InCore has proven its SoC Generator in a test chip
Intel Auto Business Runs Out of Gas
Microsoft’s Maia NPU Slips
Google TPU to Host OpenAI Models
The MCUs Aren’t All Right
Intel axes auto business. In other news, Intel had an automotive business.
Huawei and SiliconFlow describe speedy LLM for CloudMatrix 384
Slash Server Costs Without Hardware or Software Changes
AMD Bares Instinct MI350 GPU, Teases MI400 and MI500