Singaporean start-up Acrab recently launched an AI chip called GELIX 1, aiming to challenge the status of Apple's Mac mini and other products in the field of small desktop AI computing. This chip manufactured using a 5-nanometer process is equipped with a 20-core ARM CPU and a multi-core neural processing unit (NPU), and has a unified memory bandwidth of 273GB/s similar to Apple's M4 Pro.
In the test officially announced by Acrab, when GELIX 1 ran the Gemma 26B model (40,000 Token context window), the prefill rate reached 1,416 Tokens per second. In comparison, the Mac mini equipped with the M4 Pro chip performed only 188 Tokens per second under the same conditions. However, these benchmark test data are currently still in the company's self-testing stage and still need to be independently verified by a third-party agency.
In addition to performance, Acrab also claims that the chip has the ability to run large-scale language models of up to 100 billion parameters (100B), and the equipment footprint is equivalent to that of a Mac mini. It is designed to provide localized large-model computing power solutions for users with strict requirements on privacy and latency. However, industry analysts point out that running tens of billions or even hundreds of billions of models puts forward extremely high requirements for unified memory capacity (usually requiring 128GB or higher configuration), which may mean that the price of its high-end version will be quite expensive. At present, the specific price and launch time of GELIX 1 have not been announced.