Intel’s 350W D270 Beats Nvidia’s 600W RTX PRO 6000 By 5.9x. Read The Footnote Before You Get Excited.
Let’s get the headline out of the way, because it’s a good one:
Intel’s new GPU D270 — codename Crescent Island, Xe3P architecture, 350W, PCIe 5.0, no display outputs — beats Nvidia’s RTX PRO 6000 Blackwell Server Edition by up to 5.9x in Intel’s own financial-services benchmarks.
Not 59%. Not 5.9%. 5.9 times.
That’s the number HotHardware led with, and it’s real — in Intel’s tests, on Intel’s pre-production silicon, on Nov 6, at SC26 in Chicago, HotHardware — Intel Benchmarks Show 350W GPU D270 Beating NVIDIA’s RTX PRO 6000 By 5.9x.
But 5.9x is the best case. The worst case is 1.92x. And the footnote is the story.
So let’s do what TweakTown does: Read the spec sheet, read the test methodology, read the footnote, and then ask: What does this actually mean for you?
First: What Is The D270? It’s Not A Graphics Card.
This is the most important thing to understand, and the thing most headlines skip:
The D270 is not a graphics card.
It has no display outputs. You cannot plug a monitor into it. You cannot play Cyberpunk on it. You cannot run Blender viewport on it. It is a compute accelerator. A GPU that does math, not pixels.
Think of it like this: A graphics card is a car that can go to the shops, take the kids to school, and do 0-100 in 3 seconds. The D270 is a dragster. It does one thing — double-precision math, FP64 — incredibly fast, and it does nothing else.
That’s why Intel calls it a “GPU D270”, not an “Arc Pro B270”. It’s in a new D-series, for data center, for HPC, for AI inference, for financial modeling, for seismic imaging. Not for gaming. Not for your desktop.
The Specs: 32 Xe Cores, 160GB, 350W, And The One Number That Matters
| Spec | Intel GPU D270 | Nvidia RTX PRO 6000 Blackwell Server |
|---|---|---|
| Architecture | Xe3P (Crescent Island) | Blackwell GB202 |
| Xe / CUDA Cores | 32 Xe cores, 256 XMX engines | 9,216 CUDA cores, 288 Tensor cores |
| Memory | 160GB LPDDR5X | 96GB GDDR7 |
| Memory Bandwidth | 1.5 TB/s | 1.6 TB/s (Server), 1.8 TB/s (Workstation) |
| Power | 350W | 600W |
| Form Factor | PCIe 5.0, air-cooled, dual-slot | PCIe / SXM, 600W, needs beefy cooling |
| FP64 | Full-rate, 64 FMA per Xe core | 1:64 ratio, 1.968 TFLOPS (Workstation) |
| FP8 AI | 1.3 PFLOPS dense | — |
| Display Outputs | None | Yes |
| Price | Not announced | ~$8,000-$10,000+ (varies) |
| Availability | Sampling Oct 2026, no GA date | Shipping now |
Look at that table. Two things jump out:
1. Memory: D270 has 160GB. Nvidia has 96GB. That’s 66% more memory. For financial models that need to fit huge datasets in memory, that’s huge. For AI inference, that’s huge. You can fit bigger models, bigger contexts, without swapping to system RAM.
2. FP64: This is the killer. FP64 is double-precision math — 64-bit floating point. It’s what scientists, engineers, and quants use when they need accuracy, not speed. Most GPUs are terrible at FP64, because games don’t need it. Nvidia’s RTX PRO 6000 is terrible at FP64, because it’s a workstation / AI card, not an HPC card. Its FP64 rate is 1:64 of its FP32 rate. That’s like having a Ferrari that can only go 30 km/h in reverse.
Intel’s Xe3P does FP64 at full rate. 64 FMA per Xe core. No penalty. That’s why it wins by 5.9x in FP64 workloads. It’s not that Intel is faster in general. It’s that Intel is faster at the one thing Nvidia is bad at.
Think of it like a powerlifter vs a sprinter. The sprinter is faster at 100m. The powerlifter is stronger at deadlift. If you test them on deadlift, the powerlifter wins by 5.9x. That doesn’t mean the powerlifter is faster. It means you tested deadlift.
Intel tested deadlift.
The Benchmarks: 5.9x, 5.6x, 4.6x, 1.92x — And Why The Spread Matters
Intel tested four financial-services workloads, all FP64, all options pricing — the math banks use to price options, manage risk, and run Monte Carlo simulations.
| Workload | D270 vs RTX PRO 6000 | What It Means |
|---|---|---|
| European Monte Carlo | 5.9x | Simulate thousands of price paths, average them. FP64 heavy. D270 crushes it. |
| European Black-Scholes | 5.6x | Closed-form options pricing. FP64 heavy. D270 crushes it. |
| Binomial Pricing | 4.6x | Tree model, step-by-step pricing. FP64 heavy. D270 wins big. |
| American Monte Carlo | 1.92x | American options can be exercised early, more complex, more branching. D270 still wins, but only 1.92x. |
Average: ~4.5x. That’s still massive. But the spread — 5.9x to 1.92x — tells you something important: The D270’s advantage depends on the workload. When the workload is pure FP64 math, D270 wins big. When the workload has branching, early exercise, complex logic, the advantage shrinks.
That’s normal. That’s how hardware works. No GPU wins everything by the same margin.
Intel also tested AI inference:
- Qwen3-30B-A3B, FP8: D270 is 1.12x faster than RTX PRO 6000. That’s a lead, not a rout. 12% faster.
- Drex 2.0, decision model: D270 has 29% lower latency. That’s good. That’s meaningful for agentic AI, where latency matters more than throughput.
So: In FP64, D270 wins by 4.5x average. In AI inference, D270 wins by 12% and 29% latency. That’s not a 5.9x rout. That’s a solid win in a specific area, and a modest win in AI.
Intel also showed seismic imaging: 45.91 billion grid points per second on a 1-billion-point 3D MiniMod grid, with Viridien, the Earth data company. Viridien ported its seismic imaging mini-app to D270 in “a matter of weeks”, with 90% of kernels running well without Intel-specific tuning, and complex kernels hitting 70% of theoretical peak right away, Intel Newsroom — Intel GPU D270 Benchmarks: Up To 5.9x Throughput Over RTX PRO 6000.
That’s not a benchmark. That’s a porting story. And for HPC labs, porting is the hard part. If you can port in weeks, not months, that’s a win.
The Footnote: The 8 Cards That Weren’t 8 Cards
Now, the part HotHardware glossed over, and the part every TweakTown reader should care about: The footnote.
Intel’s original appendix said:
– D270 system: 1x Xeon 6741P, 768GB RAM, 1x D270 (pre-production), 3x Arc Pro B70 cards – Nvidia system: 2x Xeon 6944P, 2.3TB RAM, 8x RTX PRO 6000 Server Edition cards
8 cards vs 1 card? If Intel compared 1 D270 against 8 RTX PRO 6000s, and D270 still won by 5.9x, that would be 47x per card. That’s impossible. So the comparison must have been 1 vs 1, per card. But Intel didn’t say that clearly. It said “8 cards” in the system, without saying “only 1 used.”
On Nov 17, after Hardware Busters and others raised the question, Intel updated the appendix: “Only 1 card used” for the Nvidia system, Hardware Busters — Intel GPU D270 Claims 5.9x FP64 Throughput Of NVIDIA’s RTX PRO 6000.
Good. That clarifies it: 1 vs 1.
But there are still two open questions:
1. Different host systems: D270 ran on a single Xeon 6741P, Nvidia ran on dual Xeon 6944P. Different CPUs, different RAM (768GB vs 2.3TB), different everything. That shouldn’t matter much for GPU-bound FP64, but it’s not a clean apples-to-apples test. TweakTown would never test like this. We’d use the same CPU, same RAM, same motherboard, same everything, and only swap the GPU.
2. Different software: D270 ran SYCL / OneMKL, Nvidia ran CUDA or SYCL, whichever was faster. Intel says it used the best Nvidia path. But Nvidia’s FP64 emulation — a technique that speeds up matrix multiplication by using FP32 to emulate FP64 — was not in the version Intel tested. That technique only helps matrix multiplication, not all FP64, but it could have helped Nvidia in some workloads.
3. Pre-production silicon: All D270 tests were on pre-production hardware, Nov 6-9. Final silicon could be faster, or slower, or the same. We don’t know. Intel hasn’t said.
So: Is the 5.9x real? Probably, in Intel’s tests, on Intel’s workloads, on Intel’s pre-production card. Is it a clean, independent, reproducible test? No. Not yet. No independent D270 benchmark had been published as of Nov 16, according to Jon Peddie Research, TechTimes — Intel Says D270 GPU Beats Nvidia At FP64: Nearly 6x On Monte Carlo, Under 2x On American Options.
Wait for independent testing. Always wait for independent testing. Vendor benchmarks are marketing. Independent benchmarks are truth.
Why Nvidia Is The Wrong Opponent — And Why Intel Picked It Anyway
Here’s the thing Chris would ask: Why did Intel compare against the RTX PRO 6000, and not against Nvidia’s real HPC card, the B200 / H200?
Because the RTX PRO 6000 is a workstation card. It’s for AI, for graphics, for professional visualization. It’s not for HPC. Its FP64 is 1:64. Nvidia’s H200 / B200 has full-rate FP64, or close to it, and would have been a much tougher opponent.
Intel picked the RTX PRO 6000 because it’s the card people know, because it’s the card with 96GB, because it’s the card that costs $8,000-$10,000, and because it’s the card Intel can beat by 5.9x in FP64.
If Intel had compared against the H200, the number would have been much smaller. Maybe D270 still wins on price, on power, on memory. But not by 5.9x.
That’s not cheating. That’s marketing. Every vendor picks the opponent they can beat. Nvidia does it. AMD does it. Intel does it. Your job as a reader is to notice who they *didn’t* pick.
Intel didn’t pick H200. Intel didn’t pick B200. Intel didn’t pick AMD MI325. Intel picked RTX PRO 6000.
Remember that.
The Real Story: Intel Is Back In HPC, And It’s About Power And Memory, Not Just Speed
Forget the 5.9x for a second. Here’s the real story:
Intel is back in HPC. After years of Xeon Max, Ponte Vecchio, Falcon Shores, delays, cancellations, and “we’re focusing on AI”, Intel has a card that does FP64 at full rate, has 160GB of memory, and draws 350W.
350W. That’s the key number for builders.
350W is a normal PSU. 350W is air-cooled. 350W is a dual-slot PCIe card you can put in a server, in a workstation, in a 4U chassis, without liquid cooling, without a 1200W PSU, without redesigning your data center.
600W is not normal. 600W is a 1200W PSU. 600W is liquid cooling, or a very loud blower, or a very hot server room. 600W is expensive to run, expensive to cool, expensive to deploy.
If you run 100 D270s at 350W, that’s 35kW. If you run 100 RTX PRO 6000s at 600W, that’s 60kW. That’s 25kW difference. That’s real money, real cooling, real power bill.
And 160GB vs 96GB? That’s the difference between fitting a model in memory and not fitting it. That’s the difference between running a simulation in one go and splitting it across two cards.
That’s why Intel is also showing a hybrid system with SambaNova SN50 RDUs — D270 for prompt processing, SN50 for token generation. D270 does the FP64, the prompt, the heavy math. SN50 does the fast token generation. Together, they make a system that’s faster and cheaper than a single big GPU.
That’s the future of AI + HPC: Not one big GPU that does everything, but a system of specialized chips that each do what they’re good at. D270 is Intel’s bet on that future.
What This Means For Gamers: Nothing. And That’s Okay.
I know. You’re a gamer. You’re reading TweakTown. You want to know: Can I buy a D270? Will it play games? Will it be cheaper than a 5090?
No. No. And no.
D270 has no display outputs. It will not play games. It is not for sale to consumers. Intel has not announced a price, or a general availability date. Sampling started in October. General availability is “next year”, according to BusinessKorea, but Intel hasn’t confirmed.
What D270 *does* mean for gamers is this: Intel is still in GPUs. Intel is still investing in Xe. Intel’s Xe3P architecture will trickle down to gaming GPUs, eventually. The FP64, the XMX engines, the LPDDR5X — those technologies will show up in future Arc gaming cards, in future integrated GPUs, in future laptops.
And competition is good for gamers. If Intel can beat Nvidia in HPC, Nvidia has to respond. If Intel can beat Nvidia in AI, AMD has to respond. If all three are competing, prices go down, performance goes up, and you win.
So: D270 won’t play games. But D270 existing is good for gamers, because it means Intel is still fighting.
What This Means For Builders: PSU, Thermals, And The 350W Sweet Spot
If you’re building a workstation for AI, for HPC, for financial modeling, for seismic imaging, here’s what to take away:
PSU: 350W per card means you can run 2x D270 on a 1000W PSU, with headroom for CPU, RAM, SSDs. 600W per card means you need 1500W+ for 2 cards. That’s a different PSU, a different price, a different noise level.
Thermals: 350W air-cooled is doable in a good case with good airflow. 600W air-cooled is a jet engine. If you’re building a small-form-factor system — and I love SFF, you know I do — 350W is the sweet spot. 600W is not SFF-friendly.
Memory: 160GB means you can run bigger models, bigger simulations, without going to a multi-GPU setup. For a solo researcher, for a small lab, for a startup, that’s huge. You can do in one card what used to take two.
Software: This is the catch. D270 runs SYCL, OneAPI, OneMKL. Nvidia runs CUDA. CUDA has 15 years of libraries, tools, tutorials, Stack Overflow answers. SYCL is newer, smaller, less mature. Intel says 90% of kernels ported in weeks, with no tuning. That’s promising. But for production, for your code, for your models, you’ll need to test. Don’t buy on Intel’s benchmarks. Test your own workload.
The Bottom Line: 5.9x Is Real, But It’s Not The Whole Story
HotHardware gave you a headline: “350W GPU D270 Beating Nvidia’s RTX Pro 6000 By 5.9x.”
Intel’s D270, 350W, 32 Xe3P cores, 160GB LPDDR5X, full-rate FP64, beats Nvidia’s 600W RTX PRO 6000 by 5.9x in European Monte Carlo, 5.6x in European Black-Scholes, 4.6x in Binomial, 1.92x in American Monte Carlo — average 4.5x — in Intel’s own tests on pre-production silicon, 1 vs 1, Nov 6-9, at SC26. In AI inference, it’s 1.12x on Qwen3-30B, 29% lower latency on Drex 2.0.
It has no display outputs, no price, no ship date. It beats Nvidia’s workstation card in FP64 because Nvidia’s workstation card is bad at FP64. It does not beat Nvidia’s H200. It does not play games. It does mean Intel is back in HPC, with a 350W, 160GB, air-cooled card that could change how small labs, banks, and seismic companies do FP64 work. Wait for independent testing. Always wait for independent testing.
That’s it. That’s the D270.
It’s not a gaming GPU. It’s not a 5090 killer. It’s not a 5.9x-everything card.
It’s a dragster. It does one thing incredibly fast. And for the people who need that one thing — quants, scientists, engineers, seismic imagers — it could be a game-changer.
For the rest of us? It’s a sign that Intel is still in the fight. And in a world where Nvidia has 90% market share, that’s good news for everyone.
Would you buy a 350W, 160GB, no-display-output card for FP64 work, if it was cheaper than a 600W RTX PRO 6000? Or is CUDA lock-in too strong to switch? And be honest — did you think D270 was a gaming card when you first saw the headline?