Spec Card SPEC CARD
Intel Data Center GPU Max 1550
Intel's flagship data center GPU (the GPU product line alongside Gaudi 3), the primary accelerator of Aurora, the first exascale supercomputer (63,744 units deployed — the largest single-card deployment worldwide), and the standout example of a 47-tile multi-process-node Foveros+EMIB multi-chip package
Overview
Intel Data Center GPU Max 1550 is the flagship model of Intel's data center GPU line, known internally by the codename Ponte Vecchio. Using Foveros 3D packaging and EMIB bridge technology, it integrates sixteen TSMC 5nm compute tiles, multiple Intel 7 Rambo cache tiles, HBM2e memory tiles, and SerDes tiles into a single package — 47 functional tiles in total, spanning five process nodes, with a transistor count exceeding 100 billion. The chip features 128 Xe-cores, over a thousand XMX matrix engines, and 128 GB of HBM2e memory. It connects to the host via PCIe Gen 5 and provides Xe Link high-speed interconnect. As the primary accelerator of the Aurora supercomputer, it targets data center workloads such as large-scale scientific computing, AI training, and high-performance data analytics.
Core Specs
| Architecture | Xe-HPC microarchitecture, codename Ponte Vecchio (PVC), a 2×Xe-Stack (64 Xe-cores per stack) dual-technology multi-chip package combining Foveros 3D and EMIB 2D: 47 functional tiles across five process nodes (16 compute tiles TSMC N5 + 8 RAMBO cache tiles + 2 base dies Intel 7 + 8 HBM2E + 2 SerDes tiles TSMC N7, 11 EMIB), >100 billion transistors |
| Memory | 128GB HBM2e (each Xe-Stack with its own HBM2e stacks; Aurora single-card STREAM measured 2.1TB/s) |
| Memory Bus Width | 1024 bit (per ARK's official Graphics Memory Interface spec) |
| Memory Bandwidth | 3276.8 GB/s |
| Compute | Official peak: FP64 52 / FP32 52 / XMX TF32 419 / XMX BF16 839 / XMX FP16 839 TFLOPS / XMX INT8 1678 TOPS; Aurora fleet-wide per-card measured GEMM medians (official paper Table 3): FP64 29.2 / FP32 44.0 / TF32 212.8 / BF16 420.3 / FP16 421.6 TF/s, INT8 854.3 TOPS (architecture-based: 32,768 FP64 FLOPs/clock × up to 1.6 GHz, consistent with both tables) |
| Compute engines | 128 Xe-cores (including 1024 XMX matrix engines + 1024 Xe Vector engines; XMX 4096-bit wide, vector 512-bit dual-issue FP64 FMA), Register File 64MB / L1 64MB / L2 (Rambo) 408MB |
| Power | 600W TDP (ARK official; Aurora ECP deployments configured CPU/GPU at a balanced 4kW, with GPU sustained around the 500W tier) |
| Interconnect | Xe Link high-speed interconnect fabric (ARK: Xe Link Max Frequency 53 Gbps; Aurora ships dual-stack all-to-all with dual planes, 28 GB/s theoretical per link) + PCIe Gen5 x16 host interface (64 GB/s design bandwidth) |
| Form Factor | OAM accelerator module (OAM supports 4/8-GPU all-to-all topology); each Aurora blade packs 2× Xeon CPU Max + 6× Max 1550 |
| Launch Year | 2023 |
| Transistors | >100 billion (47 functional tiles, five process nodes) |
Specs verified against official documentation (
official benchmarks , updated 2026-10-10) · actual price subject to the seller's quote
Use Cases
- Large-scale scientific computing and numerical simulation: with dual-precision vector engines and high-capacity HBM2e memory, well suited to traditional HPC workloads such as computational fluid dynamics, climate modeling, and molecular dynamics
- Large-scale AI model training and fine-tuning: XMX matrix engines support BF16, FP16, TF32 and other precisions, making the card suitable for large-scale deep learning training tasks
- Exascale supercomputing infrastructure: as the primary accelerator of the Aurora supercomputer, suited to national-level computing platforms requiring massive GPU networking and global interconnect
- Mixed Precision inference and data analytics: low-precision paths such as INT8 can serve large-scale inference services and analytics workloads demanding high-throughput data processing
Same-Family Models
| Model | Memory | Power |
|---|
| Intel Gaudi 2 Compare | 96GB HBM2E (6× HBM2E stacks) | OAM module: 600W TDP |
| Intel Data Center GPU Max 1100 Compare | 48GB HBM2e (ARK official; SURF node configuration listed as 'Max 1100 48G XeLink') | 300W TDP |
| Intel Gaudi 3 Compare | 128GB HBM2e (8×16GB stacks, 3.6GHz, unified addressing across dual dies) | OAM HL-325L: 900W TDP (passive cooling; liquid-cooled also 900W) | PCIe HL-338 card: 600W TDP (passive cooling) |
Video Overview VIDEO
Procurement & Inspection Notes
- This model is an OAM-form-factor data center accelerator with power consumption at the high end of the data center range. Before deployment, verify that rack power delivery and blade or node cooling (Liquid-cooled or forced Air-cooled) are compatible; for multi-card nodes, conduct a dedicated assessment of facility power redundancy and thermal density
- Verify multi-chip package integrity: the card is built from dozens of functional tiles via Foveros and EMIB. On inspection, use official diagnostic tools to confirm that all Xe-cores, XMX engines, and the full HBM2e capacity are intact and functional, and that device identifiers match official records
- Confirm software stack and interconnect configuration: at purchase, verify supported oneAPI toolchain versions, and check that Xe Link bridge cards, PCIe Gen 5 motherboard slots, and in-chassis cabling space meet the data center layout requirements for multi-card interconnect. Also note the officially announced expected discontinuation date — evaluate long-term supply and maintenance cycles accordingly
FAQ
How does the Intel Data Center GPU Max 1550 differ from the Gaudi series?
Both belong to Intel's data center accelerator portfolio, but they serve different purposes. The Max 1550 is a general-purpose GPU based on the Xe architecture, with a full graphics and compute pipeline, large HBM2e memory, and Xe Link interconnect, covering both HPC scientific computing and AI training/inference. The Gaudi series, by contrast, is a dedicated ASIC accelerator aimed at deep learning training and inference. Workloads requiring a general-purpose GPU programming model (such as SYCL, oneAPI, and the existing GPU ecosystem) or mixed HPC/AI loads typically go with the Max 1550.
How large is the Max 1550 deployment in the Aurora supercomputer?
The Aurora supercomputer deploys 63,744 Intel Data Center GPU Max 1550 units alongside more than 20,000 Xeon CPU Max processors across over 10,000 blade nodes — one of the largest single-model GPU deployments within a single system to date, with theoretical peak compute exceeding the 200 EFLOPS class (on a mixed-precision basis).
Which precisions and data types does the Max 1550 support?
The card supports FP64 and FP32 double- and single-precision vector compute, and its XMX matrix engines support TF32, BF16, FP16, as well as low-precision types such as INT8 and INT4. Each precision delivers different peak throughput: AI training typically uses BF16/FP16, inference can use INT8/INT4, and traditional HPC simulation relies primarily on FP64. On the software side, development is done through the oneAPI and SYCL programming models.