Intel Takes On Nvidia: the 350W D270 Beats the 600W RTX PRO 6000 by Up to 5.9x
A 350W Intel card posts 5.9x the throughput of Nvidia's 600W RTX PRO 6000 on Monte Carlo and Black-Scholes tests. Every number from Intel's own lab, on engineering samples.
📜 Table of Contents
When Crescent Island launched a few months ago, I closed that piece with one line: Intel has to publish the benchmarks. Without numbers it is all slides. This time the homework is in.
On November 16, at the SC26 supercomputing show, Intel published the first performance data for the data-center card now officially named GPU D270. Intel says that on financial workloads this 350W card reaches up to 5.9 times the application throughput of Nvidia’s RTX PRO 6000 Blackwell server edition. That card draws 600W.
Four financial tests, as published:
| Financial test (FP64) | D270 throughput vs RTX PRO 6000 |
|---|---|
| European Monte Carlo | 5.9× |
| Black-Scholes (European) | 5.6× |
| Binomial | 4.6× |
| Black-Scholes (American) | 1.92× |
What the 5.9x is worth
The gap has a reason. D270’s Xe3P architecture runs full-rate FP64 — 64 floating-point multiply-adds per Xe core — so one architecture choice covers low-precision inference and high-precision compute at once. The RTX PRO 6000 uses the GB202 core it shares with the RTX 5090, which is built for graphics and low-precision AI throughput; double precision was never its home turf. On FP64-heavy financial code, a gap of several times is what a card that did its assigned job should show.
On AI inference, head to head, D270 also holds up: on Qwen3-30B it reaches 1.12 times the throughput, and on the Drex 2.0 decision models it runs 29% lower latency while drawing a little over half the power. One caveat, stated plainly: every one of these numbers comes from Intel’s own lab, on pre-production engineering samples, with the Nvidia side measured on cloud virtual machines. There is room to recheck once retail cards exist.
There is one number in the announcement more interesting than the 5.9x.
$2 billion in orders
Lip-Bu Tan said at the SGLang Summit that orders for the D-series inference cards have passed $2 billion, making it the best-selling GPU in Intel’s history. Two conditions apply. First, this is a pure AI card: no display output, the 3D graphics hardware cut, so you cannot even play Minesweeper on it. Second, it is still in customer sampling; volume production is scheduled for 2027 and no price has been announced.
A card that cannot ship yet has $2 billion of orders queued. That is the most direct footnote on how tight compute cards are.
You cannot buy it, but it reaches you
The chain runs like this. Tech Spot’s tracker puts the global average price of the RTX 5090 up about 262%, with US listings at $9,400 at one point; Chinese channel quotes sit at 55,000 to 60,000 yuan. The 5080 starts at 17,000 yuan, and the 5070, once the value pick, is 8,000 yuan at spot. In November Colorful ran another factory price round, adding 2,200 to 3,300 yuan to a single 5090 DV2; Galax and MSI paused shipments; AMD is reported to raise prices this month too. All of the above is channel and media reporting, not official pricing.
The cause of the increases is that compute cards cannot be bought. HBM capacity is taken by AI servers, and proper AI accelerators are expensive and back-ordered, so companies turn to high-end gaming and professional cards for inference. The RTX PRO 6000 and the 5090 use the same GB202 die and compete for the same capacity. Every premium in a parts list has some AI buying behind it.
Cards like D270 exist to take that demand elsewhere. It does not use HBM; it uses LPDDR5X, the cheaper, more available memory from phones and laptops, 160GB per card to start and up to 480GB with partners. The pitch is performance per dollar and tokens per watt, aimed at inference buyers who cannot afford or cannot queue for HBM parts.
The logic rhymes with the last mining boom: miners took gaming cards, gaming cards vanished, then purpose-built miners arrived and gaming card prices fell. The more plentiful dedicated compute cards are, the harder it is for AI buyers to reach into the gaming shelf.
Timing, and the next card
The competition has barely started. D270 does not reach volume production until 2027, its price is unset, and LPDDR5X is itself getting more expensive. Memory capacity being squeezed by AI is an industry-wide problem; GDDR7 is tight as well, and several research houses now put the point where graphics cards get cheaper at 2028.
At the same show Nvidia announced a Rubin HPC GPU for 2028, aimed squarely at HPC and double precision. By the time D270 ships, its opponent will probably not be this generation; both sides will have better cards.
The direction, though, is clear. Every inference card sold at 350W, with large memory and no HBM queue, is one more place for compute demand to go, and one inch of slack for the gaming card market. For anyone waiting for cards to get cheaper, this is the category of news worth watching over the next year.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.