Acrab Launches GΞLIX 1 Edge AI Chip
Startup Acrab has launched the GΞLIX 1 System-on-Chip (SoC), designed for edge AI devices. The chip supports local execution of models with up to 100 billion parameters.
Startup firm Acrab formally announced the GΞLIX 1 System-on-Chip (SoC), its inaugural product targeting edge AI devices. Manufactured using a 5nm process, the chip is engineered to run artificial intelligence models containing up to 100 billion parameters directly on the device.
The GΞLIX 1 incorporates a heterogeneous architecture, featuring a 20-core Arm CPU, a 3TFLOPS GPU, and a Neural Processing Unit (NPU), delivering a total AI computing power of 650TOPS. It utilizes an 8MB L1 cache and boasts a shared L2 cache bandwidth of 768GB/s, coupled with 256-bit LPDDR5X memory operating at 8533T/s.
Acrab highlights specific performance gains, stating that GΞLIX 1 accelerates the attention mechanism through vector hardware and reduces initial token latency via optimizations. In a Gemma 26B A4B configuration with a 40k KV cache and 10k token input, the chip reportedly achieves a pre-fill rate of 1416.8 tokens per second, which Acrab claims is 7.5 times faster than an Apple M4 Pro Mac Mini.
Beyond the hardware, Acrab is providing a comprehensive software stack and agent reference design. The company has also developed its first first-party terminal device, the Agent Box, based on the GΞLIX 1 chip. This integrated approach aims to simplify and accelerate the deployment of AI capabilities on edge devices.