Intel Says D270 GPU Beats Nvidia at FP64: Nearly 6x on Monte Carlo, Under 2x on American Options
13 hour ago / Read about 27 minute
Source:TechTimes

SANTA CLARA, CALIFORNIA - JANUARY 22: The Intel logo is displayed on a sign in front of Intel headquarters on January 22, 2026 in Santa Clara, California. Intel will report fourth-quarter earnings today after the closing bell. Justin Sullivan/Getty Images

Intel's Nov. 16 newsroom post gives the first per-workload split of its Intel GPU D270 double-precision claim: up to 5.9x the application throughput of Nvidia's RTX PRO 6000 Blackwell Server Edition on European Monte Carlo pricing, falling to 1.92x on American Black-Scholes. Intel's own footnotes say the tests ran on pre-production silicon, and the Nvidia software it lists predates Nvidia's FP64 emulation, a technique that speeds up only matrix multiplication.

The split matters to anyone sizing double-precision hardware for options pricing, risk or simulation. It shows where Intel's card leads, where the lead thins, and which question the footnotes leave open. Intel published the post as Supercomputing 2026 (SC26) gets under way in Chicago, and it gave no price or release date.

Intel's Table Runs From 5.9x on Monte Carlo to 1.92x on American Black-Scholes

Intel reports four results against a single RTX PRO 6000 Blackwell Server Edition, all on FP64 financial workloads:

WorkloadD270 vs. RTX PRO 6000
European Monte Carlo5.9x
Black-Scholes (European)5.6x
Binomial4.6x
Black-Scholes (American)1.92x

The four figures average about 4.5x. Intel lists the D270 at 350 W and the Nvidia card at up to 600 W, but it does not report performance per watt. The financial tests ran Nov. 6 on pre-production hardware, and the post gives no problem sizes.

Why Does the D270 Win at FP64? Nvidia Runs Double Precision at 1:64 on This Silicon

The RTX PRO 6000 shares its GB202 silicon with the GeForce RTX 5090, a gaming chip where FP64 is a small fraction of FP32. StorageReview lists the Workstation Edition at 1.968 TFLOPS of FP64, a 1:64 ratio to its 126 TFLOPS of FP32. Nvidia's datasheet for the Server Edition that Intel tested lists 120 TFLOPS of FP32 and no FP64 figure at all.

Intel says each Xe core delivers 64 FP64 fused multiply-add operations and runs FP64 at full rate, but it publishes no TFLOPS number and no clock speed. Nand Research lists a peak FP64 vector rate of 10.2 TFLOPS, built from an approximate 2.5 GHz clock that Nand itself derived from Intel's FP8 rating. Against the Workstation Edition's 1.968 TFLOPS, that is a peak ratio of about 5.2x by TechTimes arithmetic.

Two of Intel's four results, 5.9x and 5.6x, exceed that ratio. Raw FP64 peak therefore does not explain them alone. Candidate factors include software, the D270's 160 GB of memory against the Nvidia card's 96 GB, and the test setup, and Intel has not separated them.

How Does Nvidia's FP64 Emulation Work, and Where Does It Stop?

Nvidia's cuBLAS library can emulate FP64 matrix multiplication on Tensor Cores using the Ozaki scheme, which rebuilds double-precision results from low-precision integer arithmetic. By default an automatic precision framework aims for accuracy equal to or better than native FP64 and falls back to native execution when emulation offers no benefit. On the RTX PRO 6000 Blackwell Server Edition, Nvidia's post shows speedups of up to about 13x over native FP64 on matrix multiplies, and a roughly 1.5x end-to-end gain on a Quantum Espresso benchmark. Newer Nvidia figures put emulated matrix throughput near 37 TFLOPS on this card, as relayed in TechTimes' earlier report.

The limit is scope. Nvidia describes the feature as a matrix-multiplication capability, with other routines listed as future work. AMD fellow Nicholas Malaya told The Register that emulation offers little or no benefit for roughly 60 to 70 percent of HPC workloads, which rely on vector fused multiply-add. Nvidia's Dan Ernst countered in the same Register report that more FLOPS do not mean more useful FLOPS, because memory limits often keep vector workloads far below peak. Both vendors have commercial reasons for their positions.

Intel's four tests sit on the vector side of that line by description, though Intel has not published kernels. Monte Carlo pricing draws random numbers, builds price paths, computes payoffs and averages them. European Black-Scholes is a closed-form formula, and binomial pricing works backward through a lattice of prices. None of those recipes centers on a large dense matrix multiplication, which is TechTimes' reading, not an Intel statement.

The software stack matters too. Intel's footnote lists a CUDA 12.8.0 container with oneMath and DPC++ for the Nvidia node and says the Nvidia result is the better of CUDA or SYCL, without saying which produced the number. Nvidia's post describes emulation in CUDA Toolkit 13.0 Update 2 and calls it the first release. Emulation therefore does not appear in the stack Intel lists. On matrix-heavy code, Nvidia's reported 37 TFLOPS of emulated throughput is above the 10.2 TFLOPS vector peak Nand lists for the D270, though matrix and vector rates are not a like-for-like comparison and Intel has published no FP64 matrix figure.

Intel's Footnotes Pair One D270 With Eight RTX PRO 6000 Cards and Different Software

The D270 result comes from a node with one Xeon 6741P, one pre-production D270 and three Arc Pro B70 cards. The Nvidia node uses two Xeon 6944P processors and eight RTX PRO 6000 Server Edition cards. Intel's main text compares against the RTX PRO 6000 and does not specify per-card normalization. If the Nvidia figure were an eight-card total, the per-card D270 advantage would be near 47x, far above the roughly 5.2x peak ratio, so a per-card reading is more plausible. Intel has not confirmed it.

The inference tests carry their own mismatches. On Qwen3-30B-A3B in FP8, where Intel claims 1.12x the Nvidia card's output-token throughput, the D270 ran vLLM with torch.compile off, while the Nvidia card used torch.compile, CUDA graphs and FlashInfer attention. On the DREX decision model, where Intel claims up to 29% lower latency at concurrency 1, the Nvidia card ran in an AWS g7e.24xlarge virtual machine. The Qwen footnote cites concurrency up to 32, while the appendix lists up to 64.

Do Finance Teams Need FP64 for Monte Carlo? Single Precision Often Suffices Except for Greeks

Oxford mathematician Mike Giles argues that single precision is adequate for Monte Carlo finance work except when sensitivities, known as Greeks, are computed by bumping inputs, and he recommends a double-precision accumulator or binary tree summation when averaging many paths. A NAG technical report found a GPU Monte Carlo option pricer stable and accurate in single precision, with roughly twice the speed.

Nvidia's datasheet lists 120 TFLOPS of FP32 for the Server Edition. A team that accepts single-precision Monte Carlo is therefore not forced onto the card's 1:64 double-precision path, and Intel's table does not test FP32 finance workloads. The D270 case is strongest where model-risk rules or Greeks require true double precision.

D270 Has No Price, No Ship Date and No Independent Benchmark Yet

Intel states no price or availability. Customer sampling began in October, and Jon Peddie Research notes that Intel has announced no independent performance results. Nand Research projects volume availability in 2027 as its own expectation, and it flags that Intel's vLLM and SGLang kernel coverage lags CUDA.

The comparison also skips Nvidia's dedicated HPC parts. Nvidia's H100 and H200 each deliver about 34 TFLOPS of FP64. If Nand's 10.2 TFLOPS holds and the rates are comparable, the D270 would sit near 30% of that figure by TechTimes arithmetic.

Intel's post cites no audited benchmark. Independent STAC-A2 audit results for derivatives risk analytics have been published for Intel GPUs before, including four Intel Data Center GPU Max 1550 units.

What Will Intel Show at SC26? Live Demos Are Planned but Come Without Numbers

Intel plans two live demonstrations at its SC26 booth: seismic modeling on the D270, and a disaggregated inference system in which the D270 handles prefill, SambaNova SN50 chips handle decode, and Xeon 6 and Xeon 6+ processors handle orchestration and tool execution. Intel gives no booth number, no demo times and no latency or throughput figures for the combined system. SC26 lists its Grand Opening Gala for 7 to 9 p.m. Chicago time Monday, and exhibits open at 10 a.m. Tuesday.

The seismic workload is FP32, not FP64, and Intel gives no seismic performance numbers. Intel says Viridien, which Intel's post spells Veridien, ported a seismic imaging mini-application from HIP and CUDA to SYCL in under a month, that about 90% of kernels performed well without Intel-specific tuning, and that complex kernels reached up to 70% of theoretical peak on the unoptimized first port. Those are Intel's statements about a partner's work. SambaNova's earlier disaggregated demos paired its decode chips with Nvidia GPUs for prefill.

Buyers Should Test Matrix Share, Precision Needs and Per-Card Results Before Trusting 5.9x

For double-precision code built on vector math, Intel's table is a reason to request D270 test hardware and run your own kernels. For matrix-heavy code, Nvidia's emulation changes the baseline Intel measured against. For Monte Carlo teams that accept single precision, the comparison shifts again, because Nvidia's FP32 rate is not throttled. Until independent audits, a price and a ship date arrive, the 5.9x figure is a vendor-run, pre-production result against a part whose FP64 rate is deliberately low.

Frequently Asked Questions

Is the Intel GPU D270 faster than the RTX PRO 6000 at FP64?

Only in Intel's own tests of four financial workloads, on pre-production hardware, where results run from 1.92x to 5.9x. The Nvidia card's FP64 rate is 1:64 of its FP32 rate, which makes it an easy double-precision baseline. No independent D270 benchmark had been published as of Nov. 16, according to Jon Peddie Research.

Does Nvidia's FP64 emulation close the gap?

Only for matrix multiplication. Nvidia reports speedups of up to about 13x over native FP64 on that routine for this card, but AMD's Nicholas Malaya says the feature offers little benefit to vector-heavy HPC codes. Intel has not said whether its option-pricing kernels use matrix multiplication.

Does quantitative finance need FP64?

Not always. Oxford's Mike Giles considers single precision adequate for most Monte Carlo pricing, with the main exception being Greeks computed by bumping. Teams bound by model-risk rules that require double precision are the clearest fit for FP64 hardware.

When can buyers get the D270, and what will it cost?

Intel has announced no price or release date, and customer sampling began in October. Nand Research expects volume availability in 2027, which is its projection, not an Intel date.