Test report DSG-2974 · Rev C · tested November 17, 2026
AI Datacenter InfrastructureDevice under test
Intel GPU D270 Claims 5.9x Lead Over RTX PRO 6000 at 350 W
Intel's D270 accelerator claims up to 5.9x the FP64 throughput of Nvidia's RTX PRO 6000 at 350 W vs 600 W, on pre-production silicon with no price or ship date.
- Read
- 3 min
- Words
- 606
- Node
- 45nm
- Operator
- Amara Osei
Spec summary
- Intel reports up to 5.9x Monte Carlo throughput vs RTX PRO 6000 Blackwell Server Edition
- D270 draws 350 W vs 600 W for the Nvidia card — a 1.71x rated power gap
- Xe3P design: 32 Xe cores, 256 XMX engines, full-rate FP64 from FP4 through FP64
- Results used pre-production silicon; price, date and availability unconfirmed
- AI claim limited to up to 29% lower latency on Drex 2.0
Intel says its next data-center accelerator, the GPU D270, delivers up to 5.9 times the application throughput of the NVIDIA RTX PRO 6000 Blackwell Server Edition in European Monte Carlo options-pricing tests — while drawing 350 W against the Nvidia card's 600 W. Wccftech and SSBCrack carried the first benchmark figures this week; every number below comes from Intel's published material or from outlets that reported it.
What did Intel report?
The claims center on financial-computing kernels, a deliberate choice aimed at quant desks that run these models nightly in double precision:
- European Monte Carlo options pricing: up to 5.9x higher throughput
- European Black-Scholes: 5.6x
- Binomial options pricing: 4.6x
- American Black-Scholes: 1.92x
- Drex 2.0 (AI inference): up to 29% lower latency
Intel states the results used pre-production silicon and may vary. Until third parties test shipping hardware, the ratios remain vendor claims.
How does the Xe3P design work?
The D270, formerly codenamed Crescent Island, uses the Xe3P architecture with 32 Xe cores and 256 Xe Matrix Extensions (XMX) engines. The XMX units handle matrix math for AI inference; the Xe cores carry the FP64 story. Each Xe core delivers 64 FP64 fused multiply-add operations, and Intel lists native support from FP4 through FP64, with double precision running at full rate rather than throttled.
That combination targets a buyer who wants one accelerator for both simulation and inference, in a market that has spent years chasing FP8 and FP4 for AI at the expense of scientific and financial code. Intel has not published a price or full retail specification sheet; memory configuration and clock behavior appear in some coverage but remain unsettle in Intel's own material.
Why does the workload mix matter?
The spread in results tells its own story. Monte Carlo and European Black-Scholes, the most embarrassingly parallel kernels, land at 5.9x and 5.6x. Binomial pricing drops to 4.6x, and American Black-Scholes falls to 1.92x. The more a workload depends on sequential decisions and memory access patterns, the less raw FP64 throughput decides the outcome. Aroged framed the caveat directly: the D270 is 5.9 times faster than Nvidia, but not in everything. Buyers running mostly American-style books should anchor on 1.92x.
What does the per-watt math imply?
The D270 is reported at 350 W and the RTX PRO 6000 at 600 W — a rated ratio of 1.71x. Multiplying Intel's throughput claims by that ratio yields roughly 10.1x per rated watt on Monte Carlo, 9.6x on European Black-Scholes, 7.9x on binomial and 3.3x on American Black-Scholes. These are derived calculations, not Intel's, and real systems rarely sit at rated power under FP64 load.
What remains unconfirmed?
- Confirmed as Intel claims: the throughput and latency ratios above, plus the pre-production caveat
- Reported: 350 W and 600 W power ratings
- Unconfirmed: independent verification, price, release date, retail specifications and availability
The timing sharpens the story: Nvidia reports quarterly results on Nov. 17, with analysts expecting about $109 billion. A single FP64 benchmark does not threaten Nvidia's data-center revenue, which rests on software and networking as much as silicon — Nvidia holds roughly 85.2% of the AI chip market. But procurement teams reading "5.9x at 350 W" will ask their Nvidia account managers to justify the gap, and that pressure tends to surface in discounts and roadmap commitments before market share shifts.
The takeaway: the specifications are design facts, the ratios are vendor measurements, and everything about price and availability is blank.
via shattered.io (Source)
More from Amara Osei
Show full bio
Staff writer covering money, funding and market shifts around hardware layer of artificial intelligence at Die Signal.
698 articles
Same lot · LOT-362D
- DSG-13695nmIntel Names Crescent Island: GPU D-Series Targets AI Inference
- DSG-453865nmNVIDIA N1X Rumor Claims iGPU Parity With RTX 5070 Core Count
- DSG-790045nmIntel's D-Series Datacenter GPU Brings 480GB Memory to AI
- DSG-600365nmIntel Announces GPU D-Series And Xeon Rack Designs
- DSG-71947nmCerebras stock slides on report Nvidia powers OpenAI 'Ultrafast'