Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
InfrastructureAPI / Product · Actionable3DNews AI→ original

Intel D270 Beats Nvidia RTX Pro 6000 by 5.9x in Financial Calculations and by 12% in AI

Intel says its D270 accelerator (a 350 W PCIe card) processes 1.92–5.9 times more data per second than Nvidia RTX Pro 6000 Blackwell in double-precision FP64 financial calculations. That precision matters for tasks such as risk assessment and derivatives pricing, while Nvidia's card is built for graphics and AI. In running the Qwen3-30B-A3B model (FP8), Intel's lead is just 12%. The test platforms differed, though, so independent measurements are needed.

AI-processed from 3DNews AI; edited by Hamidun News
Intel D270 Beats Nvidia RTX Pro 6000 by 5.9x in Financial Calculations and by 12% in AI
Source: 3DNews AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary source3DNews AI ↗
Editor-in-Chief
Memory160 GB LPDDR5X (up to 480 GB from partners) versus 96 GB GDDR7
Key metric #2Latency in Intel's Drex 2.0 test is 29% lower than the RTX Pro 6000
Key metric #3Intel has not yet announced the price or availability date for the D270

Headlines like "5.9 times faster than Nvidia" should be read with caution. Intel has published data on its D270 accelerator (previously called Crescent Island). It is a PCIe card with a 350 W power limit. In double-precision (FP64) financial calculations, the company says it processes 1.92–5.9 times more data per second than the Nvidia RTX Pro 6000 Blackwell Server Edition. But the advantage in running neural networks is much more modest: on the Qwen3-30B-A3B model, throughput is only 12% higher. Here is what lies behind the numbers and who they matter to.

What FP64 is and why the gap is so large

FP64 refers to double-precision numbers, meaning numbers with a very long fractional part. Neural networks usually do not need that level of precision: they run well on compressed formats such as FP8. In finance, though, it matters. If you are assessing portfolio risk or pricing derivatives, millions of small rounding errors add up to a noticeable error, and in this field that error costs money.

According to Intel, the Xe3P architecture supports FP64 natively: each Xe core executes 64 fused multiply-add (FMA) operations. The D270 has 32 Xe cores and 256 XMX blocks (matrix units for neural networks) in total. The RTX Pro 6000 Blackwell is designed primarily for professional graphics and AI, not for maximum FP64 speed. That is why Intel wins in precision-critical workloads. The 1.92–5.9x range depends on the specific workload: 5.9x is the best case, not the average result.

Memory, power consumption and AI tests

The second strength of the D270 is memory. The base card gets 160 GB of LPDDR5X with 1.5 TB/s of bandwidth, and partner configurations can reach 480 GB. The RTX Pro 6000 Blackwell Server Edition has 96 GB of GDDR7, and its power consumption reaches 600 W versus 350 W for Intel's new card. If a large model or dataset does not fit into 96 GB, the extra memory may matter more than speed.

The AI picture is more modest. For inference (running an already trained model), the D270 delivered 1.12 times the throughput of the RTX Pro 6000 on Qwen3-30B-A3B in FP8, roughly 12% more. In Intel's Drex 2.0 test, its latency was 29% lower. In both cases, individual GPUs were compared. For a card that is not yet on the market, this is a respectable result, but not a breakthrough.

Caveats about the tests

The comparison cannot be considered clean:

  • the results were obtained on pre-production D270 samples;
  • the test platforms and software settings differed between Intel and Nvidia;
  • the appendix to the financial tests mentions a system with one D270 and three Arc Pro B70 cards, while Nvidia used eight RTX Pro 6000s;
  • Intel does not specify how the workload was distributed across GPUs.

All figures come from Intel itself, and there are no independent measurements yet.

What comes next and what practitioners should do

Intel is also working with SambaNova on a modular inference platform. In it, the D270 processes prompts (reads the request and prepares the context), while SN50 accelerators generate tokens, meaning the response itself. These two phases stress hardware differently, so it makes sense to split them across separate chips. Intel plans to show this configuration at the SC26 conference. The company has not announced a broad sales date or a price for the D270.

The practical takeaways are simple:

  • if you calculate financial models in FP64, keep the D270 on your radar and ask for benchmarks on your own workloads;
  • if you need LLM inference, wait for independent measurements: a 12% gain on a single model is not a reason to change your fleet yet;
  • if you run large models on a single card, look at memory capacity, not just speed.

For now, this is Intel's claim to a niche where Nvidia is weakest, not a direct challenge to the leader in AI accelerators.

Story

Key Questions Answered

What FP64 is and why the gap is so large?

FP64 refers to double-precision numbers, meaning numbers with a very long fractional part. Neural networks usually do not need that level of precision: they run well on compressed formats such as FP8. In finance, though, it matters. If you are assessing portfolio risk or pricing derivatives, millions of small rounding errors add up to a noticeable error, and in this field that error costs money.

What are the key details on Memory, power consumption and AI tests?

The second strength of the D270 is memory. The base card gets 160 GB of LPDDR5X with 1.5 TB/s of bandwidth, and partner configurations can reach 480 GB. The RTX Pro 6000 Blackwell Server Edition has 96 GB of GDDR7, and its power consumption reaches 600 W versus 350 W for Intel's new card. If a large model or dataset does not fit into 96 GB, the extra memory may matter more than speed. The AI picture is more modest.

What are the key details on Caveats about the tests?

The comparison cannot be considered clean: the results were obtained on pre-production D270 samples; the test platforms and software settings differed between Intel and Nvidia; the appendix to the financial tests mentions a system with one D270 and three Arc Pro B70 cards, while Nvidia used eight RTX Pro 6000s; Intel does not specify how the workload was distributed across GPUs. All figures come from Intel itself, and there are no independent measurements yet.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…