For most of the last decade the question at supercomputing conferences has not been whether Nvidia would supply the accelerators but how many of them a lab could afford to power and cool. Intel arrived at SC26 in Chicago with a different framing, and it did so on November 16, 2026, with a card it now calls the Intel GPU D270. Intel Newsroom reported on November 16, 2026 that the company published its first benchmarks for the part, formerly code-named Crescent Island, and set them directly against Nvidia's RTX PRO 6000 Blackwell Server Edition.
The comparison Intel chose is not the one that dominates AI headlines. It is double precision floating point, the FP64 arithmetic that underpins simulation, seismic processing and financial modeling, where answers must be reproducible and numerically stable rather than merely fast in low precision. Intel's claim is that the D270 delivers up to 5.9 times the application throughput of the Nvidia card on a European Monte Carlo options-pricing test, a workload that European banks and risk desks run constantly. The power figures attached to that claim are 350W for the Intel card and 600W for the Nvidia card.
That combination, a large performance multiple paired with a lower power envelope, is the sort of assertion that HPC buyers tend to treat with polite skepticism until they can run the tests themselves. The benchmarks were produced on pre-production silicon, tested between November 6 and November 9, 2026, and Intel has not published a price or an availability date. Customer sampling began in October 2026. Those qualifiers matter as much as the headline number, and the reporting that followed the announcement made that clear.
Intel's pitch also extends beyond a single card. The company used SC26 to describe a mixed system in which Xeon processors orchestrate work while accelerators handle the heavy math, and it demonstrated the D270 alongside partners in seismic imaging and inference. The D270 is therefore best understood as one piece of a broader argument about open software and heterogeneous machines rather than as a standalone product launch.
Key Facts
According to Intel Newsroom, the D270 is built on the Xe 3P architecture with 32 Xe cores and 256 Xe Matrix Extensions engines, delivering 64 FP64 fused multiply-add operations per core and native support for datatypes from FP4 through FP64. The Intel-branded card carries 160GB of LPDDR5X memory at 1.5TB/s of bandwidth, with partner-designed configurations scaling to 480GB, all inside a 350W air-cooled PCIe 5.0 form factor.
The published multiples against the RTX PRO 6000 Blackwell Server Edition are 5.9x on European Monte Carlo, 5.6x on European Black-Scholes, 4.6x on Binomial and 1.92x on American Monte Carlo. Intel Newsroom reported on November 16, 2026 that the D270 runs at 350W while the Nvidia part runs at 600W, and it stated plainly that the figures come from pre-production silicon. Jon Peddie Research lists the card at 1.3 PFLOPS peak FP8, a reminder that the D270 is not a double-precision-only design.
Hardware Busters reported on November 16, 2026 that Intel used SC26 to publish its first head-to-head numbers and argued that the footnotes deserve as much attention as the headline figure. The outlet attributed the outline to Jeff McVeigh, who runs Intel's Data Center Strategic Programs, and noted that the card keeps real double-precision compute alongside low-precision AI formats.
The inference results are far less dramatic. Hardware Busters reported on November 16, 2026 that Intel claims up to 29 percent lower latency on Drex v2.0 and 1.12x the output token throughput on Qwen3-30B-A3B at FP8. On a card pitched partly as an AI accelerator, a 12 percent token throughput advantage is a different order of claim from a 5.9x FP64 multiple.
StorageReview reported on November 16, 2026 that Intel's SC26 briefing also covered Diamond Rapids, the next-generation Xeon architecture, which Intel says will scale to 256 performance cores with up to 1.6TB/s of memory bandwidth and up to 1.28GB of last-level cache. Intel gave no launch date for Diamond Rapids, an 18A-P design that supports APX, AVX 10.2 and enhanced AMX.
There is also a correction to record. Hardware Busters reported on November 17, 2026 that Intel updated its appendix after questions, clarifying that the financial test system listing eight Nvidia RTX PRO 6000 cards carries the note that only one card was used, so the comparison is one D270 against one RTX PRO 6000. Intel also corrected its page on November 17, 2026 to label the fourth benchmark the American Monte Carlo model rather than American Black-Scholes.
Analysis
The most important detail in the whole disclosure is the one Intel did not lead with: Nvidia's RTX PRO 6000 uses the same GB202 silicon as the GeForce RTX 5090, where FP64 runs at a small fraction of the FP32 rate. Hardware Busters reported on November 16, 2026 that beating that card at FP64 says as much about Nvidia's product segmentation as about the D270. Nvidia did not design the RTX PRO 6000 to be a double-precision monster; it designed it to be a workstation and inference part with server credentials. Winning a race the other runner was not entered in is a real result, but it is not proof of a general purpose accelerator lead.
What this really means is that Intel has found a defensible niche rather than a decisive victory. The workloads where the D270 shows 5.9x and 5.6x multiples are exactly the ones that punish Nvidia's consumer-derived silicon, and the workloads where the gap collapses to 1.12x are the ones that matter most to AI buyers. Intel's own numbers draw that line for the market, which is unusual candor and also a strategic admission. Anyone shopping for prefill and decode throughput has little reason to switch on the strength of these slides.
The methodology caveats deserve equal weight. StorageReview reported on November 16, 2026 that the results are vendor figures on pre-production silicon that it did not run or audit, and Hardware Busters reported on November 16, 2026 that the Nvidia inference results came from an AWS EC2 g7e.24xlarge virtual machine while the D270 ran on a local server. The hosts differed as well: a single Xeon 6741P with 768GB of memory for the D270 against dual Xeon 6944P processors with 2.3TB for the Nvidia card. Cross-platform comparisons with different hosts and different virtualization layers are useful for orientation and weak as procurement evidence.
The bigger picture here is that the D270's real competition may not be Nvidia's workstation card at all but the cost of porting code. Intel argues that a single open software stack, spanning Intel MPI, Fortran, oneAPI and SYCL, lets existing HPC code run without vendor lock-in. The Viridien seismic imaging demo is the evidence for that argument: an application ported from HIP and CUDA to SYCL in under a month, with 90 percent of kernels running well without Intel-specific tuning. If that ratio holds on other codes, the switching cost story becomes more persuasive than any single benchmark multiple.
Why It Matters
Double precision still decides who buys what in large parts of the HPC market. Weather models, structural simulation, seismic inversion and derivative pricing all depend on FP64, and the vendors that walked away from it created a gap. Intel is now selling into that gap with a 350W air-cooled card that fits existing PCIe 5.0 servers, which means a lab can evaluate it without rebuilding a datacenter around liquid cooling.
Power is not a footnote in that calculation. A 350W part against a 600W part changes rack density, cooling requirements and electricity contracts over a multi-year deployment, and those costs often outweigh the sticker price of the hardware. Intel's argument only holds if independent testing confirms the multiples, but the direction of the claim is the one procurement teams care about.
The disaggregated inference demo adds a second dimension. Intel paired the D270 handling prefill with SambaNova SN50 reconfigurable dataflow units handling decode, orchestrated by Xeon 6 and Xeon 6+ processors. That is an explicitly multi-vendor architecture, and it signals that Intel is willing to sell a component rather than insist on owning the whole rack.
Next Up
The open questions are price, availability and independent verification. Customer sampling began in October 2026, and there is still no price or availability date, which means the 5.9x figure cannot yet be converted into a cost per unit of work. Until third parties run the Monte Carlo and Black-Scholes tests on shipping silicon with matched hosts, the numbers remain a vendor claim.
Intel also has to ship Diamond Rapids and prove that the software story scales beyond one carefully chosen seismic code. The company has given no launch date for the 256-core Xeon platform, and the memory channel count and speed behind the 1.6TB/s figure remain undisclosed. Those details, more than any SC26 slide, will determine whether the D270 becomes the first credible alternative in FP64 accelerators or a useful bargaining chip.
Comments (0)
Log in or sign up to leave a comment.
No comments yet. Be the first to share your thoughts.