Newshot

Intel's New 350W D270 Card Takes Aim, Delivering 5.9x the Financial Compute of the 英伟达 600W Accelerator

AI summary

Intel on November 16 published performance data for its Data Center GPU D270, claiming that this 350W accelerator card reaches up to 5.9x the application throughput of the 600W 英伟达 RTX PRO 6000 Blackwell Server Edition in a specific financial computing test. The D270 is built on the Xe 3P architecture with 32 Xe cores and 256 XMX engines, natively supports FP4 through FP64 with full-rate dual-precision compute, and packs 160 GB of LPDDR5X memory (1.5 TB/s bandwidth, expandable to 480 GB) in a 350W PCIe 5.0 air-cooled form factor.

Translation(model translation, not the source's own text)

Intel's New 350W D270 Card Draws Its Sword: Financial Compute Up to 5.9x Faster Than 英伟达 600W Accelerator

IT Home, November 16 news: Intel published performance data for its data center GPU D270 on November 16, claiming that this accelerator card, set at 350W power, achieves application throughput up to 5.9 times that of the 600W 英伟达 RTX PRO 6000 Blackwell server edition in specific financial computing tests.

In terms of specifications, the relevant information for the data center GPU D270 is as follows:

  • Silicon and precision: Xe 3P architecture, 32 Xe cores, 256 XMX engines; native support from FP4 to FP64, with full-rate double-precision computing capability.
  • Memory and platform: 160 GB LPDDR5X (1.5 TB/s bandwidth, expandable to 480 GB); 350W PCIe 5.0 air-cooled form factor.
  • Workloads and ecosystem: supports HPC scenarios (such as financial modeling and seismic imaging), decoupled deployment, and agentic AI inference; programming and development are done through an open software stack.

In its blog post, Intel noted that high-performance computing (HPC) is undergoing a fundamental shift. Traditional large-scale numerical modeling and simulation remain the foundation, but research and enterprise environments are increasingly integrating simulation, analysis, and AI inference into continuous workflows.

Emerging agentic AI architectures accelerate this evolution, connecting data, models, and domain tools into autonomous loops that complete input preparation, evaluate intermediate outputs, and dynamically guide the next simulation step.

This convergence trend requires heterogeneous infrastructure, with different compute engines executing different stages of the workflow. Intel's GPU D270, previously codenamed Crescent Island, is designed precisely for this environment. Although it was built from the ground up for efficient AI inference, its architecture provides the high-precision computing capability, large memory capacity, and standard enterprise form factor required by demanding HPC environments.

Scientific research, engineering, and financial simulation rely on double-precision arithmetic to maintain numerical fidelity and prevent the accumulation of rounding errors. Built on the Intel Xe 3P architecture, the Intel GPU D270 features 32 Xe cores and 256 Xe Matrix Extensions (XMX) engines, with each core capable of 64 FP64 fused multiply-add (FMA) operations.

The architecture gives developers a choice of precision: low precision (FP4 to FP8) for scenarios that need inference throughput, and full-rate native FP64 compute for scenarios where mathematical precision cannot be compromised.

The memory architecture is equally critical for handling complex working sets. This Intel-branded board carries 160 GB of high-speed LPDDR5X memory with 1.5 TB/s bandwidth, and partner-designed configurations can expand to 480 GB.

This large memory capacity allows researchers to keep massive datasets and wavefields resident on the GPU, minimizing costly data movement across the host bus. All of this compute density is integrated into a standard 350W air-cooled PCIe form factor, deployable within standard enterprise servers or scaled across racks.

Real-world application benchmarks based on pre-production silicon demonstrate the tangible advantages of D270's balanced architecture in financial analysis and scientific simulation scenarios:

  • Financial services: In options pricing benchmarks, D270 delivers up to 5.9 times the application throughput of the NVIDIA RTX PRO 6000 Blackwell server edition [1], demonstrating superior performance in compute-intensive financial modeling. This performance is especially notable given that D270 draws 350W while the NVIDIA GPU draws 600W.
WorkloadApplication / ModelBaseline HardwareIntel GPU D270 Test Result
Financial servicesEuropean Monte Carlo simulationRTX PRO 6000 (Blackwell)Up to 5.9x throughput
Financial servicesBlack-Scholes model (European)RTX PRO 6000 (Blackwell)Up to 5.6x throughput
Financial servicesBinomial tree modelRTX PRO 6000 (Blackwell)Up to 4.6x throughput
Financial servicesBlack-Scholes model (American)RTX PRO 6000 (Blackwell)Up to 1.92x throughput
  • Seismic exploration: Seismic modeling is one of the most demanding workloads in the oil and gas industry, requiring repeated acoustic wave propagation computations over massive 3D datasets using FP32 compute. D270's combination of large memory capacity and FP32 compute capability delivers strong performance in this workload. Cache-aware stencil optimization maximizes data reuse in local caches, keeping large wavefield arrays resident directly on the board. Intel will demonstrate seismic modeling capability through live demos at its booth at Supercomputing '26.
  • For scientific computing, the core logic is clear: large memory capacity handles larger-scale problems, smarter data reuse simplifies data movement, and an optimized open programming model delivers the application throughput researchers need.

In modern scientific discovery, AI inference is no longer a separate task that runs after simulation ends; it runs directly inside the simulation loop. Pretrained AI surrogate models act as physics simulators, screening hundreds of thousands of design parameters within seconds and narrowing down promising candidates before launching full-scale numerical solvers. Meanwhile, agentic decision models monitor simulation telemetry in real time, identify instabilities, dynamically refine mesh boundaries, and autonomously guide the next computation without waiting for human intervention.

However, integrating high-throughput inference into supercomputing loops places demands on speed and extreme operational efficiency. Monolithic compute nodes would have different compute stages contending for the same processor resources, compressing throughput and driving up latency.

Decoupled architectures solve this bottleneck by splitting inference workloads across processors suited to different compute stages, significantly accelerating overall execution while optimizing energy consumption and resource utilization.

In this decoupled architecture approach, the Intel GPU D270 provides compute for the compute-intensive prefill stage, rapidly processing high-dimensional simulation datasets, long prompts, and scientific context windows. The SambaNova SN50 Reconfigurable Dataflow Unit (RDU) accelerates the memory-bandwidth-bound decode stage for fast inference; Intel Xeon 6 and Xeon 6+ processors handle end-to-end workflow orchestration, host communication, and tool execution.

By orchestrating workloads between Intel GPUs in the D series and SambaNova RDUs, this heterogeneous approach accelerates the entire agentic loop, allowing autonomous decision models to guide simulations in real time without idling expensive numerical compute resources.

At the same time, all of these capabilities are delivered within the safe power limits of existing air-cooled racks. Intel and SambaNova are building this solution to help service providers scale compute capacity to meet AI token demand without waiting for new data center facilities and using the most common software stacks.

At SC26 in Chicago, Intel will demonstrate this decoupled heterogeneous inference architecture live at its booth, showing the Intel GPU D270 and SambaNova SN50 RDU working together in real time.

For HPC centers and research institutions, the core value of this architecture is that it can be deployed directly: customers can deploy advanced AI inference and agentic pipelines out of the box, leveraging engineered decoupled infrastructure without bearing the burden of developing custom software orchestration or low-level scheduling layers.

If hardware innovation comes with software barriers that force teams into proprietary walled gardens that limit code portability and require costly rewrites, its practical value is greatly diminished. A viable software strategy cannot demand a complete break from past technology stacks; it must protect existing software investments while unlocking new acceleration capabilities.

Intel directly addresses this problem by providing a unified open software stack that supports both traditional numerical simulation and modern AI workloads. Rather than requiring developers to rewrite algorithms in proprietary dialects, Intel adapts to teams' existing development environments. Intel has long invested in building HPC foundational standards, including Intel MPI and optimized Fortran and high-performance C++ compilers, ensuring that existing parallel codebases can run efficiently on modern heterogeneous architectures without vendor lock-in.

By focusing engineering efforts on upstream integration with standard AI frameworks, inference runtimes, and the open SYCL programming model, research teams can deploy the Intel GPU D270 without modifying their existing toolchains and get mature performance on day one of product delivery.

The early performance data Intel has already obtained on the Intel GPU D270 is proof of these capabilities. In emerging decision-class model scenarios (models that can make decisions autonomously rather than merely generating answers), the Intel GPU D270 running Drex 2.0 delivers up to 29% lower latency than the NVIDIA RTX PRO 6000 Blackwell server edition.

In terms of large language model throughput, D270 running Qwen 3 30B – A3B³ achieves 1.12 times the throughput of comparable competitor GPUs, while drawing only 350W. These two results, obtained on pre-production silicon, already demonstrate the Intel GPU D270's strong suitability for agentic applications.

In the aforementioned Veridien seismic demo in the HPC scenario, porting the workload to the Intel GPU D270 requires only minimal effort, because the open SYCL programming model allows teams to reuse existing development work from HIP / CUDA workloads. Porting to SYCL requires no changes to host-side code, and with ahead-of-time (AOT) compilation and kernel bundling, existing compute kernels can be reused directly. High performance is achieved without extensive retuning: 90% of kernels run well without Intel-specific optimizations, and even complex kernels reached 70% of theoretical peak performance on the first unoptimized port. The end result: a seismic imaging mini-app was ported in under a month, with both numerical accuracy and performance verified.

Article title: Intel's New 350W Card D270 Draws Its Sword: Financial Computing Power Soars 5.9x Over the 英伟达 600W Accelerator Card

Body: The Intel GPU D270 was previously regarded mostly as a GPU optimized for inference, but with the combination of native FP64 precision, large memory capacity, and open software, it has demonstrated highly convincing adaptability in the high-performance computing field: amid the trend of HPC and AI convergence, it is a balanced, well-rounded accelerator capable of effectively optimizing overall workflow output.

Advertising statement: External redirect links contained in the article (including but not limited to hyperlinks, QR codes, passwords, and other forms) are used to convey more information and save selection time; the results are for reference only, and all articles on IT Home containing external links include this statement.

Related Articles

  • 英伟达 releases Rubin HPC GPU: a single card with over 50,000 CUDA cores sets a record, expected to launch in 2028
  • Samsung expects utilization of its 8nm and below process lines to potentially exceed 100% in 2027, with HPC chips becoming the dominant force
  • OPPO Find X10 Pro Max phone rumored to globally debut three 200MP all-large-sensor imaging, with its main camera debuting Samsung HPC
  • 35 英伟达 AI HPC systems benefit over 3 million researchers, as Europe launches the largest AI supercomputing buildout in history
  • Fujitsu unveils its chip development roadmap: MONAKA-XX to achieve CPU + NPU integration in 2031
  • WeRide × Lenovo globally debut the 100% automotive-grade HPC 3.0 computing platform equipped with the NVIDIA DRIVE Thor chip

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗