Data Center GPU Max 1550
Inteldatacenter| Overview | |
|---|---|
| Architecture | Xe-HPC (Ponte Vecchio) (2023) |
| Launch date | Q1'23 |
| Memory | |
| VRAM | 128 GB HBM2e |
| Memory bandwidth | 3,276.8 GB/s |
| Cores & clocks | |
| Xe-cores | 128 |
| Shader cores | 1,024 |
| Matrix / Tensor cores | 1,024 |
| Peak clock | 1,600 MHz |
| Board & system | |
| TDP | 600 W |
| PCIe | PCIe 5.0 x16 |
| Interconnect | Xe Link |
Source: ark.intel.com product specifications page (intel.com/.../products/sku/232873/.../specifications.html), pulled 2026-09-01 (a prior note in this file claiming ark.intel.com was blocked with HTTP 403 was from an earlier research pass using a plain fetch tool without JS/browser rendering — a direct browser visit works fine).
fp64TFLOPS/fp32TFLOPS/fp16TFLOPS/bf16TFLOPS are intentionally omitted — ark.intel.com does not publish a per-precision FLOPS breakdown for this part at all (unlike AMD/NVIDIA datacenter parts), and Intel's own Data Center GPU Max Series product brief PDF (the one document that likely does have these numbers) returned HTTP 403 via fetch and rendered as a blank/undecodable canvas via direct browser navigation — couldn't be read either way this session. The previous version of this file carried a "derived" fp16TFLOPS (2x an unverified 52 FP32 TFLOPS figure) — removed, since neither number could be confirmed from an official Intel source and this site doesn't publish self-derived estimates. Widely-cited secondary figures (~52 FP32/FP64 TFLOPS-class, ~13-team sources disagreeing by 30%+) exist but are not used here per this site's primary-source-only policy.
matrixCoreCount/shaderCoreCount (1024 each) are ark's "Intel XMX Engines" and "Xe Vector Engines" counts — structural counts, not throughput figures; ark doesn't publish a "GPU Peak TOPS" figure for this SKU either (unlike Arc consumer parts), so int8TOPS is also omitted.
interconnect is "Xe Link" (multi-GPU fabric for OAM baseboards); ark publishes only "Intel Xe Link Maximum Frequency: 53 Gbps" (a per-link signaling rate), not an aggregate bandwidth figure comparable to AMD's Infinity Fabric or NVIDIA's NVLink GB/s numbers, so interconnectBandwidthGBs is left blank rather than converting/guessing.
No targetId: Intel doesn't publish a clean oneAPI/Level Zero device-architecture identifier comparable to CUDA compute capability or ROCm gfx targets, so there's no compatibility cross-link for Intel GPUs yet.
ark.intel.com lists "Expected Discontinuance: Jan 2026" for this SKU — as of this update (Sep 2026) this part may already be past its end-of-life date; Intel's successor roadmap for this line (Rialto Bridge/Falcon Shores) remains unclear.