Skip to main content

Acrab's GΞLIX 1 Chip Brings 100B AI Models to Your Device

A startup called Acrab just dropped its first edge AI chip, the GΞLIX 1, and it's making some bold promises. Built on a 5nm process, this little powerhouse is designed to run AI models with up to 100 billion parameters right on your device—no cloud required.

What's Under the Hood?

The GΞLIX 1 uses a heterogeneous computing architecture, combining a 20-core Arm CPU, a 3TFLOPS GPU, a dedicated NPU, and audio/video processing engines all on one chip. The result? A total AI computing power of 650TOPS. To put that in perspective, Qualcomm's Snapdragon X Elite NPU hits about 45TOPS, and Intel's Core Ultra series sits around 40TOPS. That means GΞLIX 1's AI muscle is nearly ten times stronger than those competitors.

Memory and Cache: Built for Big Models

Running a 100-billion-parameter model isn't easy—it demands massive memory bandwidth. Acrab equipped the chip with an 8MB L1 cache and a shared L2 cache that delivers 768GB/s of bandwidth. It connects to a 256-bit wide LPDDR5X memory clocked at 8533MT/s. This setup is crucial for the weight reading and KV Cache access that large models rely on, making local inference feasible.

Prefilling Performance That Stands Out

Acrab focused on the prefilling phase—the stage where the model processes the entire input before generating the first token. This directly impacts how quickly you see a response after typing a long query. By using vector hardware units to accelerate Transformer attention calculations, the GΞLIX 1 achieved 1416.8 tokens per second in tests with a Gemma 26B A4B model (40k KV cache, 10k token input). That's 7.5 times faster than Apple's M4 Pro Mac Mini.

Software and Ecosystem

Acrab also released a complete software stack and an agent reference design, covering everything from drivers to application frameworks. They even showed off a first-party device called Agent Box, which serves as a reference implementation for developers to test and prototype.

Key Points

  • Chip: GΞLIX 1, 5nm process, 650TOPS AI compute
  • Capability: Runs AI models up to 100 billion parameters locally
  • Performance: 7.5x faster prefilling than Apple M4 Pro (Gemma 26B test)
  • Memory: 8MB L1 cache, 768GB/s L2 bandwidth, 256-bit LPDDR5X
  • Software: Full stack and Agent Box reference design available