跳到正文
The Elec· JEONG IL JOO·· 5 天前AI 評分43

Intel 稱 D270 GPU 吞吐量最高達 Nvidia 的 5.9 倍

Intel Says D270 GPU Delivers Up to 5.9x Nvidia's Throughput

AI 導讀

Intel 稱,為 AI 推理而設的數據中心 GPU D270,在金融服務基準測試中,吞吐量最高達 Nvidia RTX PRO 6000 Blackwell 的 5.9 倍。

正文

Intel GPU D270. (Photo: Intel)

Intel GPU D270. (Photo: Intel)
Intel GPU D270. (Photo: Intel)

Intel said on Nov. 16 that its data center graphics processing unit (GPU), the Intel GPU D270, delivers up to 5.9 times the performance of a competing Nvidia system.

​The D270, formerly known by the codename Crescent Island, is designed for artificial intelligence (AI) inference. Intel compared a system based on a D270 prototype with one powered by Nvidia's RTX PRO 6000 Blackwell GPU.

​In financial services benchmark tests, the Intel D270 recorded up to 5.9 times the throughput of the RTX PRO 6000. The result was achieved in a European Monte Carlo model. Performance gains reached 5.6 times in the European Black-Scholes model, 4.6 times in the binomial model and 1.92 times in the American Black-Scholes model.

​During the tests, the Intel D270 consumed about 350 watts of power, compared with 600 watts for Nvidia's product.

​The Intel D270 is the first product to use Intel's latest GPU architecture, Xe3P. It features 32 Xe3P cores and 256 Xe Matrix Extensions (XMX) engines. The GPU supports computing formats ranging from four-bit floating point (FP4) low-precision operations to 64-bit floating point (FP64) calculations. Its FP8 computing performance reaches 1.3 petaflops.

​FP64 is a computing format that performs calculations with extremely high numerical precision, including digits after the decimal point. It is essential for scientific, engineering and financial simulations that require high accuracy, as accumulated rounding errors can distort overall results. Each core can perform 64 FP64 fused multiply-add (FMA) operations.

​Intel's own-branded GPU D270 card comes with 160GB of LPDDR5X memory and bandwidth of 1.5 terabytes per second (TB/s). Products designed by partners can support up to 480GB. The large memory capacity allows extensive datasets to remain resident on the GPU, reducing the time and cost associated with transferring data back and forth between the GPU and central processing unit (CPU), thereby improving computing efficiency.

​At its booth at Supercomputing 2026 (SC26) in Chicago, Intel will demonstrate a seismic analysis environment developed by Viridien, a company whose seismic imaging workloads require large memory capacity and FP32 computing performance.

​Intel will offer a fully optimized system combining Xeon CPUs, the D270 GPU and SambaNova's Reconfigurable Dataflow Unit (RDU) chips. Customers will not need to develop their own custom orchestration software or low-level scheduling layers.

來源:The Elec · thelec.net