AMD, the chipmaker and Nvidia's main rival in AI data centre chips, has claimed its newest AI processor outperforms Nvidia's latest system.
The company said its Instinct MI455X chip processed 21,333 tokens per second running DeepSeek-R1, a popular AI model.
Tokens are the small chunks of text that AI models read and write, so the figure measures how quickly the chip can generate answers.
AMD said this was up to 30% faster than Nvidia's latest published result for its Vera Rubin system.
The tests were carried out by AMD itself using MLPerf, an industry benchmark for measuring AI performance, rather than by an independent body.
AMD also said its Helios system, which links dozens of its chips together in a single server rack, could move data to and from memory up to 20% faster than Vera Rubin.
A cheaper option
The company has also launched the Instinct MI450X, a new chip for Helios systems.
It has the same computing power and connections as the MI455X, but less memory, with 288 gigabytes per chip.
AMD said customers could choose between the two depending on the size of the AI models they run, how much text they need to process at once and how many users they serve.
The MI450X is aimed at customers who want to keep running costs down.
The MI400 range
AMD launched its MI400 range of chips in July.
The MI455X was designed for building and running the most advanced AI models.
A sister chip, the MI430X, was aimed at scientific research, with up to 288 TFLOPS of the high-precision computing power used in areas such as climate modelling.
A full Helios rack combines 72 MI455X chips with 18 of AMD's latest Venice processors and its own networking equipment.
It runs on AMD's ROCm software, the company's alternative to Nvidia's CUDA platform.
Read more about Advanced Micro Devices
