en.Wedoany.com Reported - GUNNIR tested the inference performance of the Intel Arc Pro B70 GPU on the DeepSeek R1 model. Results show that under high concurrency scenarios, the GPU's token throughput surpasses that of the NVIDIA RTX 5090D and RTX 4090D.

The test used a configuration of four GPUs in parallel, comparing the Intel Arc Pro B70 32GB, NVIDIA RTX 5090D 32GB, and RTX 4090D 24GB. The test model was DeepSeek R1-Distill Qwen 32B FP16, with a fixed input/output length of 128 tokens and a concurrency range covering 1 to 512. The core metric was token/s throughput, used to measure output efficiency during large language model inference.

At concurrency levels below 32, the RTX 5090D ranked first, the RTX 4090D second, and the Arc Pro B70 performed on par with the RTX 4090D. When concurrency increased to 32 and 64, the Arc Pro B70 caught up with and surpassed the RTX 4090D. At a concurrency of 128, the Arc Pro B70's token throughput was 8.6% higher than the RTX 5090D and 34.2% higher than the RTX 4090D. When concurrency rose to 256, the Arc Pro B70's lead expanded further, being 7.5% higher than the RTX 5090D and 48.7% higher than the RTX 4090D. Similar performance advantages were observed at a concurrency of 512.

In this round of testing, the Intel Arc Pro B70 GPU achieved a maximum token throughput of 2320.76 tokens per second.