5.9 times faster than Nvidia, but not in everything: Intel has revealed the capabilities of the D270 accelerator

Источник изображений: Intel

Intel has released new performance data for its dedicated D270 graphics accelerator, formerly codenamed Crescent Island. According to the company, this accelerator in the form of a PCIe expansion card with a power consumption of 350 W provides 5.9 times the throughput of the Nvidia RTX Pro 6000 Blackwell Server Edition for certain financial computing tasks.

  Image source: Intel

Image source: Intel

The D270 accelerator is equipped with a graphics processor based on Intel Xe3P architecture with 32 Xe cores and 256 XMX units. The device has 160 GB of LPDDR5X memory with a bandwidth of 1.5 TB/s. At the same time, in configurations from partners, the memory capacity can reach 480 GB. For comparison, the Nvidia RTX Pro 6000 Blackwell Server Edition accelerator offers 96 GB of GDDR7 memory and its power consumption reaches 600 W. The new product from Intel has a power limit of 350 W.

According to tests of pre-production D270 samples, the Intel accelerator shows the most significant performance advantage in financial applications that use double precision calculations (FP64). According to the company, depending on the type of workload, the D270 provides 1.92 to 5.9 times higher throughput compared to Nvidia’s solution.

Such high performance is directly related to the hardware capabilities of working with the FP64 format. The Intel Xe3P architecture supports native FP64 operations, performing 64 multiply-add (FMA) operations per Xe core. In turn, the Nvidia RTX Pro 6000 Blackwell Server Edition is focused primarily on professional graphics and artificial intelligence tasks, and not on maximum performance in FP64 computing. Thus, in tasks requiring high precision calculations, the Intel architecture is preferable.

Intel also presented comparison results in inference (inference) tasks of neural networks, where the same Nvidia accelerator acted as a competitor. When testing the Qwen3-30B-A3B (FP8 format), the D270 showed 1.12 times higher throughput than the RTX Pro 6000. In Drex 2.0 tests, the latency of the Intel solution was 29% lower. In both cases, the comparison was made using individual graphics processing units (GPUs), although the test platforms and software configurations differed.

Intel’s financials appendix mentions a system equipped with one D270 GPU and three Arc Pro B70 cards, while Nvidia’s test system included eight RTX Pro 6000 GPUs. It does not specify how the load was distributed between the GPUs when performing specific financial tasks.

Intel also notes that it is working with SambaNova on a modular platform for inference (inference) of AI models, in which D270 GPUs are used to process prompts, and SN50 accelerators are used to generate tokens. Intel plans to demonstrate this configuration at the SC26 conference. The company has not yet announced either the timing of the start of widespread sales of the D270, or its price.

If you notice an error, select it with the mouse and press CTRL+ENTER.