Radeon RX 7900 XTX overtook the GeForce RTX 4090 in working with the Deepseek R1 AI model

AMD said its flagship video card Radeon RX 7900 XTX provides higher performance compared to the GeForce RTX 4090 and RTX 4080 Super in working with Deepseek R1 II model. According to David McAfee, General Manager of the AMD Client Processors, their flagship in the RDNA 3 architecture bypass RTX 4090 to 13 %, and the RTX 4080 Super model – up to 34 % in these tasks.

Image source: AMD

AMD tested three video cards with several Deepseek R1 configurations. The Radeon RX 7900 XTX model has provided the most significant superiority over the RTX 4090 in working with the Deepseek R1 Distill QWEN 7B (7 billion parameters), where the ADA LoveLace generation by 13 %. In three other LLM configurations, the AMD card was also more effective than a competitor: in two tasks Distill Llama 8b (8 billion parameters), its superiority was 11 %, and in Distill QWEN 14B (14 billion parameters) it was 2 % faster. The GeForce RTX 4090 model was faster than the Radeon RX 7900 XTX in only one LLM configuration – Distill QWEN 32B (32 billion parameters), where its superiority was 4 %.

AMD also compared the Radeon RX 7900 XTX and GeForce RTX 4080 Super. In working with the Deepseek R1 Distill QWEN 7B (7 billion parameters), the Red Card showed 34 percent superiority over the competitor. In the tasks of Distill Llama 8b and Distill QWEN 14B, the advantage of the AMD video card decreased to 27 and 22 %, respectively.

According to the portal Tom’s Hardware, the results of the tests provided by AMD should be perceived with a certain degree of skepticism. Firstly, we are talking about internal tests of the company, and not about independent tests. Secondly, it is not known how Nvidia video cards were configured for these tests. In the end, not all workloads of AI require the use of all GPU computational capabilities. For example, Stable Diffusion may not use FP8 calculations or Tensorrt code for processing.

Radeon RX 7900 XTX cards are infrequently used as specialized AI accelerators, but their architecture allows this to do this. RDNA 3 in the Radeon RX 7900 XTX is capable of performing matrix operations, supporting the calculations of BF16 and Int8. AMD even officially uses the phrase “accelerator” in advertising materials about the architecture of RDNA 3 to demonstrate its capabilities in the tasks of artificial intelligence. The Radeon RX 7900 XTX graphics processor contains 192 AI acceleration units.

Previously, AMD also published instructions on how to use the new Deepseek R1 II model on its consumer hardware, including the Radeon RX 7900 XTX video card.

admin

Share
Published by
admin

Recent Posts

Canadians built a photon quantum computer and promised to quickly scale it to a million cubes

The Canadian startup XANADU, previously marked by joint work with NVIDIA on quantum simulators, reported…

7 minutes ago

The acute deficiency of the GeForce RTX 5080 and RTX 5090 turned into riots in Japan

The situation with the launch of the GeForce RTX 5000 video cards looks tragic or…

17 minutes ago

The marketing director of Ubisoft declassified the sales of Prince of Persia: The Lost Crown for the first year after the release

Prince of Persia metroide: The Lost Crown from the Ubisoft Montpellier studio has recently turned…

28 minutes ago

A group of investors led by MRBEAST is ready to offer more than $ 20 billion for Tiktok “significantly”

A group of American investors who expressed the willingness to buy Tiktok with the support…

2 hours ago