The market for graphics cards made for AI has received a surprise. Gunnir has published several performance tests comparing the new 32 GB Intel Arc Pro B70 against Nvidia’s powerful RTX 5090D 32 GB and RTX 4090D 24 GB. All three were tested in systems with four GPUs.
The results show that Intel’s option is not only cheaper, but also beats its rivals in heavy-load situations.
The test used the popular DeepSeek R1 LLM model, specifically the Distill Qwen 32B FP16 version. When measuring speed in tokens per second, the Nvidia RTX 5090D was ahead at first with low concurrency levels, under 32 users or requests at the same time. However, things changed as the context load increased.
Intel Arc Pro B70 Beats RTX 5090D in AI Tests at a Much Lower Price
At 128 concurrency, the Intel Arc Pro B70 beat the RTX 5090D by 8.6% and the RTX 4090D by 34.2%. This lead continued at 256 and 512 concurrency levels, where Intel’s GPU was up to 7.5% faster than the 5090D and up to 48.7% faster than the 4090D, reaching a maximum speed of more than 2,000 tokens per second, exactly 2,320.76 tokens/s.
There are two main reasons behind this. First, the Arc Pro B70 has 32 GB of VRAM, which removes the memory limits seen on the RTX 4090D with 24 GB. Second, Intel’s design includes XMX cores made for FP16 AI workloads, giving it better efficiency than Nvidia in this specific case.
The price difference is huge. The Intel Arc Pro B70 32 GB sells for $999, while the RTX 4090D and RTX 5090D cost more than $2,000 and $4,000 in Asian markets because of low stock. This shows that Intel can lead in local AI processing at about one quarter of the cost of its rivals.
