Have to give Intel credit for trying. It has not been able to crack the GPU market despite numerous efforts over the years. And once again it is making a renewed push with its GPU technology into the high-performance computing (HPC) accelerator market.

The company believes its Intel GPU D270 deserves to be considered alongside the GPU accelerators from NVIDIA and AMD for workloads such as scientific computing, financial modeling, and AI-driven simulation.

Intel says that by pairing native FP64 double precision and high-capacity memory with open software, it makes a compelling case for itself in HPC. It states that a balanced, highly capable accelerator can deliver improved workflow outcomes as HPC and AI converge.

The pitch comes from an Intel technical overview, published earlier this month. The D270 GPU is based on the company’s Xe 3P architecture and incorporates 32 Xe cores and 256 Xe Matrix Extensions (XMX) engines. Each Xe core supports 64 FP64 fused multiply-add operations, providing native double-precision computing alongside lower-precision capabilities ranging from FP4 to FP8 for AI inference.

The combination is significant because HPC applications, including scientific simulations, engineering models and financial calculations, often depend on double-precision math to maintain numerical accuracy. Many AI-focused accelerators, meanwhile, prioritize lower-precision calculations (FP4-FP16) to maximize inference efficiency.

So what Intel is attempting to do is bridge those two markets – which are typically served by separate types of silicon — with a single accelerator architecture.

Memory is another important component of the D270’s design. The Intel-branded configuration includes 160GB of LPDDR5X memory delivering 1.5TB/s of bandwidth, while partner-designed versions can increase memory capacity to as much as 480GB.

The larger memory footprint is intended to utilize ever-growing scientific datasets and simulation workloads directly on the GPU, reducing the need to repeatedly transfer data between accelerator memory and main memory.

The accelerator is designed for a 350W power envelope and uses a standard PCIe 5.0 interface with air cooling. That form factor could make it easier for HPC operators and enterprise data center customers to integrate the hardware into existing server infrastructure without adopting specialized liquid-cooling systems.

Intel points to several early benchmarks that show impressive performance. In financial modeling benchmarks using pre-production silicon, the D270 delivered up to 5.9 times the application throughput of Nvidia’s RTX PRO 6000 Blackwell Server Edition in European Monte Carlo options pricing.

Intel also reported advantages of 5.6 times in Black-Scholes European options pricing, 4.6 times in binomial modeling and 1.92 times in American Monte Carlo calculations. Intel emphasized that its accelerator achieved these results at 350W, compared with the 600W power limit of the Nvidia RTX PRO 6000.

Beyond conventional scientific computing, Intel sees an opportunity in the growing convergence of HPC and artificial intelligence. AI inference is increasingly being incorporated directly into scientific simulation workflows. Intel argues that these workloads require different types of processing resources working together rather than relying on a single, monolithic computing architecture.

One example involves a collaboration with SambaNova Systems. Under the proposed disaggregated architecture, Intel’s D270 handles the computationally intensive prefill stage of AI inference, while SambaNova’s SN50 Reconfigurable Dataflow Units handle the memory-intensive decode stage.

The objective is to assign individual processing stages to hardware best suited for their requirements, improving overall throughput and energy efficiency.

The company says this approach can help developers migrate existing GPU workloads without extensive code modifications or dependence on proprietary programming environments.

Intel cited the Viridien seismic application as an example. The workload was ported to the D270 in less than a month, with approximately 90% of its kernels performing well without Intel-specific optimizations. Some complex kernels achieved as much as 70% of theoretical peak performance during initial testing.

Intel also reported early AI inference results, including up to 29% lower latency than the RTX PRO 6000 Blackwell Server Edition when running the Drex 2.0 decision-making model. In a separate Qwen3 30B-A3B benchmark, the D270 delivered 1.12 times the throughput at the same comparison point.

Rather than competing exclusively on traditional GPU performance, Intel is emphasizing a combination of FP64 computing, substantial memory capacity, power efficiency and software portability.

The strategy could appeal to HPC operators looking to incorporate AI inference into existing scientific and enterprise computing environments without completely replacing their established infrastructure.