Intel D270 claims 5.9x FP64 throughput lead over Nvidia RTX PRO 6000
Intel’s 350-watt D270 delivered up to 5.9x the FP64 throughput of Nvidia’s RTX PRO 6000 in Intel tests, but the results used pre-production silicon and an incompletely described setup.
- By
- Rebecca Stone
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
By Rebecca Stone
Intel reported on Nov. 16 that one 350-watt D270 data-center GPU, code-named Crescent Island, delivered up to 5.9 times the throughput of a 600-watt Nvidia RTX PRO 6000 Blackwell Server Edition in the company’s European Monte Carlo options test.
Intel presented the result at the SC26 supercomputing conference in Chicago. The company tested pre-production silicon in its own laboratories. Customer sampling began in October, but Intel has announced neither pricing nor general availability.
What did Intel benchmark?
Intel’s first published table breaks down as follows:
- European Monte Carlo (FP64): 5.9x throughput.
- European Black-Scholes (FP64): 5.6x throughput.
- Binomial pricing (FP64): 4.6x throughput.
- American Black-Scholes (FP64): 1.92x throughput.
- Qwen3-30B-A3B (FP8): 1.12x output token throughput.
- Drex v2.0 (FP8): up to 29% lower latency.
The results remain Intel’s own measurements. Its largest reported advantage came in double-precision financial workloads, while the AI-inference gains were considerably narrower.
What does the test setup change?
The four financial tests priced options and compared Intel’s SYCL code with whichever Nvidia implementation performed better, CUDA or SYCL. Intel’s footnotes place one D270 in a single-socket Xeon 6741P server and eight RTX PRO 6000 cards in the Nvidia server.
Intel does not say how many of those eight cards ran or whether it normalized the results per card. It tested both systems on Nov. 6.
That is not a clean one-card-versus-one-card comparison.
The choice of rival also narrows the claim. The RTX PRO 6000 uses the same GB202 chip as the GeForce RTX 5090, whose FP64 rate runs at a small fraction of its single-precision rate. Intel did not test Nvidia’s dedicated HPC and data-center accelerators.
For the two inference tests, Intel ran the Nvidia card on an Amazon Web Services EC2 g7e virtual machine, not bare metal. That is not a like-for-like bare-metal comparison.
What hardware sits behind the claims?
The D270 combines 32 Xe cores and 256 XMX matrix engines on Intel’s Xe3P architecture. It carries 160GB of LPDDR5X memory with 1.5 TB/s of bandwidth; partner boards can hold as much as 480GB.
The accelerator is an air-cooled PCIe 5.0 card rated at 350 watts. Intel says each Xe core performs 64 FP64 fused multiply-add operations, a rate it calls “full-rate double precision.”
Intel named the D270 at October’s OCP Global Summit, where it also listed 1.3 petaflops of peak FP8 compute. That figure represents Intel’s stated peak capability, not a measured benchmark result.
Intel has marketed the card mainly for AI inference. Its new post argues that it also belongs in high-performance computing.
What is Intel showing beyond the benchmarks?
At SC26, Intel showed a seismic-imaging mini-application with Viridien. Intel said developers ported it from HIP and CUDA to SYCL in less than a month, 90% of its kernels ran well without Intel-specific tuning, and complex kernels reached up to 70% of theoretical peak on the first port.
Intel supplied no speed comparison for that demonstration.
The company is also demonstrating D270 cards handling the prefill stage of inference while SambaNova SN50 chips perform decoding. Last week, Intel formed a 23-company Korean alliance to pitch the D-Series as an inference card that avoids scarce HBM, the stacked memory used by most AI accelerators.
The D270’s use of LPDDR5X rather than stacked HBM makes memory availability central to that pitch.
What will determine the commercial result?
Intel’s strongest claims concern double-precision finance; its reported inference advantage narrows to 1.12x output token throughput and up to 29% lower latency.
The commercial test is customer adoption at a price and schedule Intel has not disclosed. Until Intel sets price and availability and customers report results, the D270’s case against Nvidia’s dedicated HPC and data-center parts will rest on Intel’s own 5.9x claim against an incompletely described configuration.
Source: mlq.ai
More from Rebecca Stone
Show full bio
Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,037 articles
Related articles
intel-18a-claims-18-efficiency-gain-as-18a-p-enters-risk-production
Intel 18A Claims 18% Efficiency Gain as 18A-P Enters Risk Production
intel-s-18a-yield-gains-put-tsmc-s-2nm-scale-to-the-test
Intel's 18A Yield Gains Put TSMC's 2nm Scale to the Test
amd-mi430x-targets-288-tflops-of-fp64-as-ai-reshapes-hpc
AMD MI430X targets 288 TFLOPs of FP64 as AI reshapes HPC
test-data-arrives-without-findings-rtx-spark-superchip-review-unverifiable-48a88da4
Test Data Arrives Without Findings: RTX Spark Superchip Review Unverifiable


