Intel D270 GPU claims 5.9x FP64 throughput over Nvidia RTX PRO 6000
★ Add us as a preferred source on GoogleTL;DR
Intel shared its first vendor benchmarks for the 350W D270 accelerator at the Supercomputing conference in November, pointing to heavy double-precision gains over Nvidia hardware.
Intel presented its first head-to-head benchmarks for the D270 accelerator at the Supercomputing conference in November. Jeff McVeigh outlined the numbers for the air-cooled, 350W PCIe card, which uses an Xe3P architecture and a 1280-bit LPDDR5X interface delivering 1.5TB/s of memory bandwidth.
Intel says the D270 reaches up to 5.9 times the FP64 throughput of Nvidia's 600W RTX PRO 6000 Blackwell Server Edition on financial workloads. Inference claims are smaller, including up to 29 percent lower latency on Drex v2.0 and 1.12 times the token throughput on Qwen3-30B-A3B at FP8. Jon Peddie Research pegs peak FP8 performance at 1.3 PFLOPS.
The base card includes 160GB of memory, while partner boards scale up to 480GB. Intel says running 300-billion-parameter models on a single card requires either a 480GB partner board or FP4 weights. Customer sampling started in October, but Intel has not announced pricing or a release date.
The benchmarks use pre-production silicon. The financial comparison matches a single D270 against eight Nvidia cards without stating if results are normalized per card, and the Nvidia inference figures use an AWS virtual machine instead of a bare-metal box.
Intel also detailed the Diamond Rapids Xeon processor, which scales to 256 performance cores with up to 1.28GB of last-level cache and 1.6TB/s of memory bandwidth. It has no launch date yet.
Intel D270 accelerator at a glance
- Memory
- 160GB to 480GB
- Interface
- 1280-bit LPDDR5X
- Bandwidth
- 1.5TB/s
- Power
- 350W
- Architecture
- Xe3P
- Form Factor
- PCIe card
Via Storagereview ↗ · Also from Hardware Busters ↗ · Other coverage: VideoCardz ↗, Wccftech ↗
Spot a mistake? Tell us


