All News
ai-hardwareintelnvidiahpcinference

Intel's first D270 benchmarks claim 5.9x over Nvidia's RTX PRO 6000. The worst of the four cells says 1.92x

Intel's first D270 benchmarks claim 5.9x the FP64 application throughput of Nvidia's RTX PRO 6000 at 350W, the best of four cells, on pre-production chips.

Vlad MakarovVlad Makarovreviewed and published
3 min read
Intel's first D270 benchmarks claim 5.9x over Nvidia's RTX PRO 6000. The worst of the four cells says 1.92x

Intel published the first benchmarks for GPU D270 on 16 November, and led with its biggest number: 5.9x the FP64 application throughput of Nvidia's RTX PRO 6000 Blackwell Server Edition, at 350 watts against 600. The post is framed as a case for high-performance computing, which is a careful way of saying a card designed for AI inference now wants to be judged on double precision.

Four cells, one headline

Intel's financial table has four rows, and 5.9x is the best of them.

Financial workloadD270 vs RTX PRO 6000
European Monte Carlo5.9x
Black-Scholes (European)5.6x
Binomial4.6x
American Monte Carlo1.92x

The bottom row is twice as fast, not six, on pre-production silicon tested by Intel against a part Intel chose.

The table also moved after publication. An editor's note records that on 17 November Intel relabelled a row, correcting the workload to American Monte Carlo from Black-Scholes (American). TechPowerUp's story still carries the old description. What Intel shipped:

  • Xe 3P, 32 Xe cores, 256 XMX engines, native FP4 through FP64 at full rate, 64 FP64 FMA per core
  • 160 GB of LPDDR5X at 1.5 TB/s, partner configurations to 480 GB
  • 350 W, air-cooled, PCIe 5.0 add-in card

No ship date or price was published.

The denominator was built for something else

Nvidia's own whitepaper states that the GB202 chip's FP64 rate is one sixty-fourth of FP32, about 2 teraflops against 126. The D270's peak FP64 vector rate is roughly 10.2 teraflops. Five to one, before anyone runs an application, so 5.9x is not a miracle.

It is, though, slightly better than raw math predicts, and that surplus is where the memory argument lives: 160 GB holds a working set that 96 GB has to stream. Intel's appendix also lists eight RTX PRO 6000 cards in the comparison system with one used, so the runs are one card against one card, a detail early coverage got wrong. The objection in the TechPowerUp thread is the fair one: the RTX PRO 6000 was never an FP64 product, and nobody compared the D270 to a card that is.

What survives the framing

Intel's AI results are more modest and more relevant: up to 29% lower latency than the RTX PRO 6000 on Drex 2.02, and 1.12x output throughput on Qwen3-30B-A3B in FP8, both at 350 W. Bandwidth is where the case narrows: Intel's 1.5 TB/s sits below the roughly 1.6 to 1.8 TB/s that NAND Research attributes to Nvidia's card, so the transferable claim is capacity per watt, not speed. A 350 W card holding 160 GB is a different shape of inference box, and that is a harder argument to fake than a benchmark table. It still needs a price.

Related Articles

Scroll down

to load the next article