Acrab GΞLIX 1 Edge AI Chip Delivers 650 TOPS for Local 100B Models

Acrab introduced the GΞLIX 1, a new system-on-chip designed to run large language models directly on edge devices. This launch matters because it targets the growing demand for local AI processing that does not rely on cloud connectivity. Buyers interested in private, on-device inference now have a hardware option that claims to handle massive model sizes without external servers.

Acrab unveils a 5nm system-on-chip with 650 TOPS AI compute for local large language model inference

The GΞLIX 1 is built on a 5nm process node and integrates a 20-core Arm CPU. Acrab also included a GPU capable of 3 TFLOPS of graphics performance. The company released a companion software stack and a reference device called the Agent Box to demonstrate the chip's capabilities.

Specifications

• Process Node: 5nm
• CPU: 20-core Arm
• AI Compute: 650 TOPS
• GPU: 3 TFLOPS
• Memory: 256-bit LPDDR5X at 8533 MT/s

The chip’s most notable specification is its AI compute power, which Acrab states reaches 650 TOPS. This figure significantly exceeds the NPU capabilities of current mainstream platforms like the Qualcomm Snapdragon X Elite and Intel Core Ultra. The system supports a 256-bit LPDDR5X memory interface running at 8533 MT/s, providing 768 GB/s of L2 cache bandwidth.

In practical testing, the GΞLIX 1 processed the Gemma 26B A4B model at 1416.8 Token/s. This pre-fill performance is 7.5 times faster than an Apple M4 Pro Mac Mini in the same benchmark. The chip is designed to execute 100B parameter large language models locally, a feat that typically requires significant cloud resources.

Acrab positions the GΞLIX 1 as a solution for high-performance edge AI applications. The company has not yet disclosed pricing or general availability dates for the chip or the Agent Box. The focus remains on the technical specifications and the benchmark results that demonstrate its local inference potential.