Intel’s Arc Pro B70 graphics card demonstrated impressive performance in AI large language model (LLM) workloads, outperforming NVIDIA’s more expensive RTX 5090D in tests using the DeepSeek R1 benchmark. Despite costing roughly a quarter of NVIDIA's price, the Arc Pro B70 delivered faster token processing speeds, especially at higher concurrency levels.

The evaluation was carried out by Gunnir, a graphics card manufacturer, which tested the Intel Arc Pro B70 32 GB model against NVIDIA’s RTX 5090D 32 GB and RTX 4090D 24 GB in quad-GPU setups. DeepSeek R1, an AI LLM benchmark utilizing Distill Qwen 32B FP16, measured token throughput across a range of concurrency from 1 to 512 tokens, focusing on tokens per second as the key metric.

Early test results confirmed NVIDIA’s RTX 5090D leading overall, with the RTX 4090D trailing closely. However, the Intel Arc Pro B70 matched NVIDIA’s 4090D performance at concurrency levels below 32, then began to exceed it starting at 32 and 64 tokens. At concurrency levels beyond 128 tokens, the B70 consistently outperformed even the RTX 5090D. Specifically, at 128 concurrency, the Arc Pro B70 exhibited an 8.6% token throughput advantage over the RTX 5090D and was 34.2% faster than the 4090D. This margin grew at higher concurrency, with a 7.5% lead over the 5090D and 48.7% over the 4090D at 256 concurrency. The GPU peaked at delivering token speeds of up to 2320.76 tokens per second.

The Arc Pro B70’s edge stems in part from its 32 GB memory capacity, compared to the 24 GB on the RTX 4090D, allowing better handling of larger AI workloads. Additionally, Intel’s XMX cores are optimized for AI acceleration, especially FP16 floating-point operations, which provide the card with a substantial advantage at higher concurrency. While NVIDIA’s RTX 5090D includes NVFP4 cores meant to enhance AI processing, Intel’s architecture currently shows superior efficiency in FP16 tasks.

It is important to note that the original RTX 5090D with 32 GB memory is no longer manufactured in China, replaced by a newer RTX 5090D V2 with reduced 24 GB memory. This change could limit NVIDIA’s ability to handle large AI language models effectively, similar to the bottlenecks observed with the RTX 4090D’s smaller memory.

Pricing also highlights Intel’s competitive positioning. The Arc Pro B70 32 GB is available at approximately $999 USD, significantly undercutting NVIDIA’s RTX 5090D and RTX 4090D models, which retail for over $4000 and $2000 USD respectively in Chinese markets. This cost disparity, coupled with the Arc Pro B70’s strong AI performance, marks a notable shift in options for AI researchers and professionals seeking effective yet affordable GPU solutions.