Intel D270 GPU First Look: FP64 Throughput Up to 5.9x That of RTX PRO 6000

Intel's D270 GPU claims 5.9x FP64 performance over RTX PRO 6000 and a slight AI inference lead, sparking alternative talk.
Intel released first benchmark data for its next-gen Crescent Island D270 GPU, claiming a 5.9x FP64 throughput advantage over NVIDIA RTX PRO 6000 Blackwell, plus 29% lower latency and 1.12x throughput in AI tests. The analysis sees Xe3P's FP64 push as a differentiation strategy targeting HPC scientific computing to avoid NVIDIA's AI training stronghold. Reddit sees the D270 as a potential value alternative to the RTX PRO 6000, but the article cautions that vendor self-reported data needs third-party validation, the oneAPI-CUDA gap is a longer slope than hardware specs, and real inference performance depends on undisclosed variables like memory bandwidth and model compatibility.
Intel recently published benchmark data for its next-generation Crescent Island "D270" GPU for the first time, claiming a throughput advantage of up to 5.9x over the NVIDIA RTX PRO 6000 Blackwell on FP64 workloads. The news sparked discussion in the Reddit community about hardware choices for local AI inference.

Performance Data: A Dual Narrative of FP64 and AI Inference
According to the published test results, the D270 achieved a 29% reduction in latency and a 1.12x throughput improvement over the RTX PRO 6000 Blackwell in AI benchmarks. These two data points paint different performance profiles: an order-of-magnitude lead in FP64, and a moderate but clear advantage in AI inference.
Behind this gap lies a difference in architectural positioning. The aggressive investment in FP64 by the Xe3P architecture indicates that Intel is targeting the D270 at double-precision floating-point-intensive scenarios such as scientific computing and engineering simulation—domains long dominated by NVIDIA's Tesla/Quadro product lines. In AI inference, the 1.12x throughput gap, while not dramatic, combined with lower latency, still has practical significance for latency-sensitive inference tasks.
FP64 (double-precision floating point) is the 64-bit floating-point arithmetic format in the IEEE 754 standard, providing approximately 15-17 significant decimal digits of precision. In scientific computing, applications such as climate modeling, fluid dynamics, and molecular dynamics require iterative solutions of differential equations, where rounding errors accumulate with the number of computational steps, making FP64 essential for numerical stability. In contrast, lower-precision formats such as FP32 (single precision) and FP16/BF16 are widely adopted in AI training and inference, because neural networks are insensitive to numerical precision and reducing precision can significantly improve throughput and energy efficiency. Over the past decade or so, consumer GPUs have continued to cut FP64 units (typically only 1/32 to 1/64 of FP32), while professional HPC cards maintain higher ratios. The D270's aggressive FP64 investment means its die area and power budget are tilted toward double-precision units, in sharp contrast to the design orientation of mainstream AI accelerators.
Throughput and latency are two orthogonal dimensions for evaluating inference hardware. Throughput refers to the number of requests processed or tokens generated per unit time, determining the overall service capacity and unit cost of a system; latency refers to the response time of a single request from input to output, which is critical for interactive applications. The two often involve trade-offs: increasing batch size can improve throughput but pushes up per-request latency. The D270 achieving both a 1.12x throughput improvement and a 29% latency reduction suggests that its advantage is not simply due to batching strategy but may come from architectural-level improvements. However, the test conditions for these two metrics (batch size, model scale, sequence length) were not disclosed, so cross-platform comparisons should be made with caution.
Why the Xe3P Architecture Bets on FP64
The phrase "Xe3P Goes Big On FP64" in the headline deserves unpacking. At a time when consumer GPUs are generally cutting FP64 capabilities, Intel is doing the opposite—and this is not merely technical showmanship.
FP64 performance is hard currency in the HPC market. Applications such as climate modeling, fluid dynamics, and molecular dynamics have rigid requirements for double-precision computing, and this market tolerates hardware premiums far more than general AI inference. Intel's choice to strengthen FP64 on Crescent Island is essentially about finding a differentiated entry point in the accelerator computing landscape dominated by NVIDIA—establishing a beachhead through the specialization of scientific computing, then expanding into broader AI workloads.
The HPC (high-performance computing) market has a fundamentally different business logic from the general AI inference market. HPC users—national laboratories, meteorological agencies, pharmaceutical companies, and engineering simulation firms—have long procurement cycles and high requirements for software stack compatibility, but they also tolerate hardware premiums better and often deploy at cluster scale, with considerable value per order. This market has long been dominated by NVIDIA's Tesla (later renamed data center GPU) and Quadro (later merged into RTX PRO) product lines, with AMD's Instinct series as secondary competition. Intel previously attempted to enter through Xeon Phi without success, and is now repositioning with data center GPUs such as Ponte Vecchio and Falcon Shores. Choosing FP64 as the differentiation angle is essentially about avoiding a head-on confrontation with NVIDIA in the AI training market and instead establishing a foothold in the double-precision domain where NVIDIA's investment has relatively weakened as its product line focus shifts to AI. Whether this "flanking breakthrough" strategy will work depends on whether Intel can simultaneously deliver a usable HPC software ecosystem.
oneAPI is Intel's cross-architecture unified programming model, designed to support CPUs, GPUs, FPGAs, and other accelerators with a single codebase. Its core components include DPC++ (a SYCL-based C++ extension), the oneMKL math library, and the oneDNN deep neural network library. Its counterpart CUDA is a proprietary ecosystem that NVIDIA has built over more than a decade, with mature compilers, debugging tools, profilers, and a vast array of optimized third-party libraries and community code. The actual performance of AI and HPC workloads depends heavily on the maturity of the software stack: the same hardware can perform several times differently under different compilers and libraries. For Intel GPUs to deliver on their hardware specifications, mainstream frameworks (PyTorch, TensorFlow) and scientific computing libraries must achieve near-native efficiency on its platform. This catch-up process is typically measured in years and is a barrier harder to cross than transistor counts.
The Possibility of a Local Inference Alternative
One user's observation in the Reddit discussion is quite representative: "Maybe we're getting a decent RTX PRO 6000 alternative." The basis for this judgment is that the RTX PRO 6000 Blackwell is positioned as a workstation-class AI accelerator card at a high price, while the competitiveness the D270 demonstrates in AI inference metrics hints at room for imagination on the price-performance front.
But it is worth keeping a cool head: the 1.12x throughput advantage was achieved under specific benchmarks, and actual inference performance also depends on undisclosed variables such as software stack maturity, memory bandwidth, and model compatibility. The gap between Intel's oneAPI ecosystem and CUDA is a longer slope than hardware specifications.
A Cautious Reading Behind the Data
The 5.9x FP64 throughput advantage is eye-catching enough, but vendor self-reported data from a single source needs cross-validation. The Reddit community discussion also shows two voices: one side sees this as an effective breakthrough by Intel in a niche market, while the other cautions waiting for independent third-party testing.
For users concerned about local AI deployment, the signal the D270 sends is positive—competition emerging in the high-end inference card market means more reasonable prices and more choices. But it is still too early to regard it as a mature alternative to the RTX PRO 6000; the catch-up of the software ecosystem and validation on real workloads are the keys to whether this GPU can deliver on its promises.
Related articles

Why Microsoft CEO Nadella Still Backs Nvidia After Jensen Huang's Departure
Microsoft CEO Nadella publicly expresses confidence in Nvidia interim CEO Colette Kress. This article analyzes Nvidia's power transition after Jensen Huang's departure, Microsoft's compute dependency, and the trust logic of the AI supply chain.

Nadella: Critical Infrastructure Is 'Unpatchable' as AI Cybersecurity Arms Race Escalates
Microsoft CEO Nadella addresses AI cyberattack fears, saying much critical infrastructure is unpatchable, downplaying the Outlook outage, and analyzing AI's dual role in cybersecurity.

Musk Backs Jensen Huang: Firing Him Is Like Cutting Jordan
Musk says Jensen Huang is loyal to the U.S. and likens firing him to cutting Michael Jordan—highlighting politicization anxieties in the AI chip industry and defending talent value.