Intel claims 5.9x Nvidia on double-precision finance math on pre-production silicon
Intel published the per-workload split the day SC26 opened in Chicago, with no price, no ship date and no independent audit, the same week shareholders sued over a 28.6% one-day drop that erased about $140 billion. For investors the card adds nothing to near-term revenue; the 14A clock is what runs.
Vincent Jiangthree-minute read
Intel's−1.17% — Intel, down 1.17 percent today D270 data center GPU got its first per-workload numbers on November 16, the opening day of Supercomputing 2026 in Chicago 1. Against Nvidia's−1.33% — Nvidia, down 1.33 percent today RTX PRO 6000 Blackwell Server Edition on double-precision finance workloads, Intel claims 5.9x the throughput on European Monte Carlo pricing, 5.6x on European Black-Scholes, 4.6x on binomial pricing and 1.92x on American Black-Scholes 1. The tests ran on November 6 on pre-production silicon 1.
Intel's D270 FP64 claims run 5.9x to 1.92x on four finance workloads
Intel's lead thins from 5.9x on Monte Carlo to 1.92x on American Black-Scholes
- Estimate
Data
| Value | |
|---|---|
| European Monte Carlo (estimate) | 5.9x |
| European Black-Scholes (estimate) | 5.6x |
| Binomial pricing (estimate) | 4.6x |
| American Black-Scholes (estimate) | 1.92x |
The footnotes are where this benchmark lives
Intel's node held one Xeon 6741P, one D270 and three Arc Pro B70 cards; the Nvidia node held two Xeon 6944P processors and eight RTX PRO 6000 cards, and Intel never states whether the comparison is per card 1. The listed Nvidia software stack was CUDA 12.8.0, while Nvidia's FP64 emulation, which reaches about 37 TFLOPS on matrix multiply, shipped in CUDA Toolkit 13.0 1. And Nvidia's dedicated H100 and H200 parts, each at roughly 34 TFLOPS of FP64, appear nowhere in the table 1.
The peak FP64 rate is about a third of an H100's
The RTX PRO 6000 shares silicon with a gaming chip and runs FP64 at 1:64 of its FP32 rate, about 1.968 TFLOPS 1. Nand Research, an independent semiconductor analysis firm, estimates the D270's peak FP64 vector rate near 10.2 TFLOPS, which would put it around 30% of an H100; that ratio is TechTimes' arithmetic, not Intel's 1. Jon Peddie Research, a graphics-industry analyst firm, notes no independent D270 benchmark had been published as of November 16 1. There is no price and no ship date; customer sampling began in October, and Nand Research projects volume availability only in 2027 1.
The H100 Intel left out of its table carries about 3.3x the D270's estimated FP64 peak
- Estimate
Data
| Value | |
|---|---|
| Nvidia H100 | 34 TFLOPS |
| Nvidia H200 | 34 TFLOPS |
| Intel D270 (Nand est.) (estimate) | 10.2 TFLOPS |
| Nvidia RTX PRO 6000 | 1.97 TFLOPS |
The market is paying 43x for 11% growth
The same week, Intel shareholders filed a proposed class action on November 13 in federal court in San Jose, naming CEO Lip-Bu Tan and CFO David Zinsner, alleging the company understated a PC slowdown and excess inventory before shares fell 28.6% on October 30, erasing roughly $140 billion of market value 23. The class covers buyers between July 23 and October 29 23. Dan Niles, founder of Niles Investment Management, puts the valuation plainly: Intel trades near 43x next-twelve-months earnings at its November 13 close of $59.66, on about 11% expected revenue growth, against Nvidia at 14x 2. The lawsuit's allegations are unproven.
Where the 5.9x holds and where it thins
European Monte Carlo draws random numbers, builds price paths and averages payoffs, vector work that suits the D270's full-rate FP64 1. American Black-Scholes, the hardest test in the table, thins the lead to 1.92x 1. Intel published no TFLOPS figure, no problem sizes, and no performance-per-watt despite rating the D270 at 350W against the Nvidia card's up to 600W 1. Two of the four results exceed the roughly 5.2x peak-ratio arithmetic, which suggests software or memory capacity contributes, and Intel has not separated the factors 1.
Buyers should run their own kernels first
Teams needing true double precision for model-risk rules or Greeks have a reason to request D270 hardware and run their own kernels 2. Teams that accept single-precision Monte Carlo face an Nvidia card whose FP32 rate is not throttled 2. Until an independent audit, a price and a date arrive, the 5.9x stays a vendor-run claim on pre-production silicon 2. The number that matters for investors is elsewhere: outside design commitments for Intel's 14A process, which management expects by mid-2027 2.
Get Chips by email, free
How AI demand is reshaping chipmakers, memory and the fabs behind them. In our daily email, every weekday.






Reader comments
0 comments