The POSITRON chapter challenges competitors; Stronger than the H200

The US company POSITRON AI has claimed that its accelerator chip, called Atlas, performs better than the Nvidia H200 in the Inference operation and consumes 5 % less electricity.

According to Tom’s Hardware, Positron, which was established in year 5, expands artificial intelligence accelerators with a special focus on inference operations. Unlike graphics processors designed to teach artificial intelligence models, inference operations, technical calculations and a wide range of tasks, the POSITRON hardware is made from the base to perform high -performance inference tasks and very low energy consumption.

Positron AI accelerator has a higher power and efficiency than the Nvidia H200

The first generation POSITRON solution for large -scale transformer models is called Atlas. The system consists of four accelerators called Archer and aimed at defeating Nvidia’s Hopper Architecture -based systems, while consuming only a fraction of their energy.

According to reports, the POSITRON AI’s Atlas system can produce about 2 token per second in the LLAMA 3.1 in the range of 2 watts, with 2 billion parameters, using BF16 calculations. In contrast, a Nvidia DGX H200 server, with a 4 -watt power consumption, is capable of producing about 1 token per second for each user. Of course, this comparison was made by Positron AI itself.

The Atlas accelerator is about 5 times higher productivity in terms of Performance-Per-WATT performance as well as performance than the cost of the NVIDIA DGX H200. Of course, this claim must be proved by a third party.

Positron AI manufactures its ASIC hardware using N4 or N5 technology at TSMC Factory # 1 in Arizona and cards are assembled inside the United States. Of course, since these chips are combined with 2GB of HBM memory, they use advanced packaging technology, which is why some of the assembly is probably done in Taiwan.