Acrab unveils 5nm edge AI chip GΞLIX 1 for local 100B model inference

Startup Acrab just launched its first chip, GΞLIX 1, built for edge AI devices. It's made with a 5nm process and can run huge AI models with up to 100 billion parameters right on the device, no cloud needed. The chip uses a mix of parts: 20 Arm CPU cores, a 3TFLOPS GPU, and an NPU, giving it 650TOPS of AI power. It has 8MB of fast L1 cache and a shared L2 cache with 768GB/s bandwidth, plus 256-bit LPDDR5X memory. Acrab sped up the attention mechanism with special hardware and made the first-token response faster. In tests with a Gemma 26B model, GΞLIX 1 hit 1416.8 tokens per second for prefill, which is 7.5 times faster than Apple's M4 Pro Mac Mini. The company also offers software tools, reference designs for AI agents, and its own device called Agent Box.