Singaporean Startup Unveils GELIX 1 AI Chip, Asserting Local Processing Capability for 100 Billion Parameter Models
7 hour ago / Read about 0 minute
Author:小编   

On July 23, 2026, Acrab, a startup based in Singapore, launched its inaugural edge AI terminal system-on-chip (SoC), the GELIX 1. Leveraging cutting-edge 5nm process technology, this chip integrates a 20-core ARM CPU, a GPU boasting 3 teraflops (TFLOPS) of computing power, an independent neural processing unit (NPU), and an audio-video processing engine. Together, these components deliver a formidable total AI computing power of 650 tera-operations per second (TOPS), enabling the local operation of large AI models with up to 100 billion parameters.

The GELIX 1 chip is further enhanced with 8 megabytes (MB) of L1 cache, a shared L2 cache bandwidth of 768 gigabytes per second (GB/s), and external low-power double data rate 5X (LPDDR5X) memory featuring a 256-bit width and a speed of 8533 mega-transfers per second (MT/s). In prefill performance testing, the GELIX 1 demonstrated a remarkable processing speed of 1416.8 tokens per second (Tokens/s) under the Gemma 26 billion (26B) A4B configuration, with a 40,000 (40k) key-value (KV) cache and a 10,000 (10k) token input. This performance outpaces the Apple M4 Pro Mac Mini by a factor of 7.5.

Additionally, Acrab has introduced a comprehensive software stack, an agent reference design, and a proprietary terminal device known as the Agent Box, further solidifying its position in the AI hardware market.