Update: Intel D270 up to 5.9x faster than RTX PRO 6000 in FP64 in first SC26 benchmarks

Intel published on Monday (November 16), during the SC26 supercomputing conference in Chicago, the first benchmarks of the Intel D270, its data center accelerator known until October by the codename Crescent Island. According to Intel's own numbers, the 350 W card delivers up to 5.9 times more FP64 throughput than NVIDIA's 600 W RTX PRO 6000 Blackwell Server Edition on financial workloads — but the AI gains are much smaller, and the tests carry important caveats. The D270 had already been introduced as an inference card with LPDDR5X memory, as we covered in our article on the D270 announcement.

What Intel claims

Across the four double-precision (FP64) financial tests, the Intel D270 posts the following throughput gains over a single RTX PRO 6000, according to Intel:

  • European Monte Carlo: 5.9x
  • European Black-Scholes: 5.6x
  • Binomial pricing: 4.6x
  • American Monte Carlo: 1.92x

The average across the four tests is 4.5x, according to Intel. In AI, the numbers are far more modest: up to 29% lower latency on the Drex v2.0 decision model and 1.12x more token throughput on Qwen3-30B-A3B in FP8.

The caveats — and Intel's correction

All results for the Intel D270 are Intel's own, measured between November 6 and 9 on pre-production silicon, with no independent verification yet. The test setup raised questions: the initial list cited eight RTX PRO 6000 cards against a single D270, without making clear whether the results were per card. On November 17, Intel updated the appendix to its material and clarified that only one card was used on each side — including in the Drex and Qwen tests — and corrected the label of the fourth financial test to American Monte Carlo (previously American Black-Scholes).

Two other caveats remain open: NVIDIA's inference numbers came from an AWS EC2 g7e.24xlarge virtual machine, while the D270 ran on a local server, and the host systems differed — a single Xeon 6741P with 768 GB of memory for the D270, versus two Xeon 6944P processors with 2.3 TB for the RTX PRO 6000. The chosen opponent also matters: the RTX PRO 6000 uses the same GB202 silicon as the GeForce RTX 5090, with very limited FP64, and is not a dedicated NVIDIA HPC card.

The hardware behind the numbers

The Intel D270 uses the Xe3P architecture, with 32 Xe cores and 256 XMX matrix engines, combining double-precision (FP64) compute with low-precision AI formats from FP4 to FP8. Intel's card carries 160 GB of LPDDR5X at 1.5 TB/s of bandwidth — partner boards go up to 480 GB — in an air-cooled PCIe 5.0 form factor at 350 W. Intel cites 1.3 PFLOPS in FP8 and 10.2 TFLOPS in FP64. Customer sampling began in October, and there is still no price or launch date.

At SC26, Intel also showed the Intel D270 in real applications: a Viridien seismic application ported from HIP and CUDA to SYCL in under a month, with about 90% of kernels running well without Intel-specific tuning, and an inference pipeline split between D270 (prefill) and SambaNova SN50 units (decode), orchestrated by Xeon 6 processors.

Featured image: Hardware Busters — graphic of the Intel D270 GPU at SC26, comparing board power (350 W) with the NVIDIA RTX PRO 6000 (600 W).

Sources

Transparency: this article was produced with the assistance of artificial intelligence, cross-checking public posts on X and sources on the internet. The performance figures are Intel's own claims, still without independent verification. Consult the original sources to confirm the data.

By Muse