Intel said at its 2026 Intel AI Summit Seoul on Nov. 12 that its GPU D-Series inference card, codenamed Crescent Island, reached up to 5.9 times the throughput of Nvidia's RTX PRO 6000 in some financial simulations, even though the card's published memory bandwidth of 1.5 TB/s sits below the 1,597 GB/s Nvidia lists for the RTX PRO 6000 Blackwell Server Edition. Because memory speed cannot explain the gap, buyers budgeting inference hardware should read the multiplier as an Intel-run result on unfinished silicon, reported as 64-bit finance math, not as proof the card serves chatbots faster. Intel has named no price and no ship date.

Intel's 5.9x Covers Four FP64 Finance Workloads on Pre-Release Hardware

Anil Nanduri, who leads AI product and go-to-market strategy in Intel's Data Center AI Group, outlined Intel's strategy in Seoul at the summit, and Intel presented its own test results there. Asia Economy reported four financial simulations using 64-bit floating-point math ran at 1.92 to 5.9 times the throughput of an Nvidia product on pre-production hardware, and that the report did not name the Nvidia product or the individual workloads.

Newstap supplied more conditions. It reported that Intel cited Black-Scholes, binomial and Monte Carlo workloads, that Intel ran the tests itself against the RTX PRO 6000, and that the software was pre-alpha. The same report said the comparison card is rated at 600 W against 350 W for Crescent Island, and that the multiplier measures raw throughput rather than performance per watt.

Intel's product page footnote says performance projections against the RTX Pro 6000 rest on public information and Intel internal analysis as of February 2026, and the page prints no multiplier. The two Korean reports also differ on detail: only Asia Economy mentions FP64, and only Newstap ties the 5.9x figure to the RTX PRO 6000 by name.

Which Hardware Differences Could Explain a Win When Bandwidth Is Lower?

Intel's 1.5 TB/s is about 94% of the Server Edition's 1,597 GB/s and about 84% of the roughly 1.79 TB/s listed for the Workstation Edition, by TechTimes arithmetic. A workload limited by memory traffic should therefore land near parity or behind. A speedup of several times has to come from somewhere else.

Double-precision math, which stores each number in 64 bits, is the most visible candidate. The Workstation Edition review lists the RTX PRO 6000 at 1.968 TFLOPS of FP64, a 1:64 ratio to its 126 TFLOPS of FP32, a low ratio. Nvidia's Server Edition datasheet does not list an FP64 figure at all. Intel says the D-Series supports precision from low-precision inference up to FP64 scientific computing, but its page gives no FP64 peak. If the four Seoul workloads were limited by FP64 arithmetic, a large multiplier would follow from the arithmetic rate and say nothing about memory.

Two other explanations stay open, and Intel has ruled out neither: problem sizes that overflow the Nvidia card's 96 GB, and kernels Intel tuned for its own part while using pre-alpha software. Which one applies cannot be settled without Intel's test configuration.

What Does 1.5 TB/s Mean for Chatbot Speed on a 480 GB Card?

Intel's headline inference claim is capacity: up to 300 billion parameters on one card. Intel's announcement does not state the numeric precision behind the 300 billion figure. Single-stream text generation reads the active weights once per token, so bandwidth sets the ceiling. At FP8, a dense 300 billion parameter model occupies about 300 GB, which exceeds the 160 GB Intel-branded card, fits only a 480 GB partner board, and caps one stream near 5 tokens per second at 1.5 TB/s; a mixture-of-experts model reading 30 GB per token would cap near 50. All of these are TechTimes arithmetic, not Intel figures.

Sweeping a full 480 GB at 1.5 TB/s takes about 0.32 seconds, against about 0.06 seconds to sweep the RTX PRO 6000's 96 GB at 1,597 GB/s, a 5.3-fold gap that is the price of capacity. The card's strength is the opposite phase: prefill, which is compute-bound. With 1.3 PFLOPS of FP8 compute and 1.5 TB/s, NAND Research computes about 870 FP8 operations per byte, a ratio that favors prompt processing and batched serving over single-stream decode.

Intel's argument is that speculative decoding offsets the decode limit. A small draft model proposes tokens and the large model verifies several in one pass, spreading the cost of loading the weights across more output tokens. Per NAND Research, Intel's measurements on frontier mixture-of-experts models showed 2.9 to 4.9 tokens retired per verification pass. Intel also pairs D-Series racks for prefill with SambaNova SN50 racks for decode, a stated direction, not a shipped product, according to Futurum Group.

Capacity Is the Differentiator: 160 GB or 480 GB Against 96 GB

Intel's own card carries 160 GB of LPDDR5X, and partner boards can reach 480 GB, up to five times the RTX PRO 6000's 96 GB of GDDR7. Power is 350 W on a passive, air-cooled PCIe 5.0 card, against up to 600 W for the Nvidia card.

Nanduri's argument in Seoul was that memory capacity limits how many tasks and agents a card can run at once: when a model and its working data do not fit, operators need more cards even if compute is sufficient. That claim is about memory capacity, and it is independent of the 5.9x finance result.

No Price, No Ship Date, and Evaluation Samples Only

Intel is shipping evaluation samples to customers, according to Seoul Economic Daily's translation of Nanduri, and plans to sell Crescent Island as a standalone chip so server makers can build their own boards. BusinessKorea reported a launch goal of next year, while Intel has announced no price.

NAND Research expects volume availability in 2027 as an inference of its own, not a confirmed date, and lists software as a risk because Intel's vLLM and SGLang kernel coverage lags CUDA. It also notes that LPDDR5X pricing is volatile, so the 480 GB configuration's economics are unknown.

What Should Buyers Ask Intel Before Trusting the 5.9x Figure?

Five questions would turn the claim into something checkable: which RTX PRO 6000 edition was the baseline (the Server Edition lists 1,597 GB/s, the Workstation Edition about 1.79 TB/s); which numeric precision each of the four workloads used; the workload names and problem sizes relative to 96 GB; the host and software versions on both sides; and what the 1.92x low end of the range represents. Until Intel publishes those answers, the evidence supports a narrow conclusion: Intel reports that a pre-release card with a bandwidth deficit beat Nvidia's part by 1.92 to 5.9 times in four FP64 finance workloads, with no per-workload breakdown. It does not yet support a claim about AI inference speed, price, or availability.

Frequently Asked Questions

Is Intel's GPU D-Series faster than Nvidia's RTX PRO 6000?

Only in Intel's own tests of four FP64 finance simulations, at up to 5.9 times on pre-release hardware and pre-alpha software. No independent benchmark is cited in the reporting, and Intel's bandwidth is lower than the Nvidia card's.

Why does a card with lower bandwidth matter for chatbot speed?

Single-stream text generation is limited by how fast weights stream from memory, so a 480 GB card at 1.5 TB/s needs about 0.32 seconds per full sweep, against about 0.06 seconds for a 96 GB card at 1,597 GB/s, by TechTimes arithmetic. Batching and speculative decoding reduce the penalty.

What is FP64, and why does it matter here?

FP64 is 64-bit double-precision floating-point math used in finance and scientific simulation, not in typical AI inference, which uses FP8 or FP4. StorageReview lists the Workstation Edition's FP64 at a 1:64 ratio to FP32, so a gain in FP64 workloads does not predict an inference gain.

When can buyers get the Intel GPU D270, and at what price?

Intel is shipping evaluation samples and has named no price or date; BusinessKorea reported a launch goal of next year.