Radeon Instinct MI25
AMDdatacenter| Overview | |
|---|---|
| Architecture | Vega10 (GCN 5.0) (2017) |
| Launch date | 2017 (AMD's own performance footnote dated Jun 2, 2017) |
| Process node | 14nm FinFET |
| GPU target (gfx) | gfx900 |
| Memory | |
| VRAM | 16 GB HBM2 |
| Memory bandwidth | 484 GB/s |
| ECC | Yes |
| Compute — vector | |
| FP64 | 0.768 TFLOPS |
| FP32 | 12.3 TFLOPS |
| FP16 | 24.6 TFLOPS |
| Cores & clocks | |
| Compute Units | 64 |
| Shader cores | 4,096 |
| Board & system | |
| TDP | 300 W |
| Form factor | Full-Height, Dual-Slot |
| Cooling | Passive |
| PCIe | PCIe Gen3 x16 |
| Multi-instance support | MxGPU SR-IOV (hardware virtualization, not MIG-style GPU partitioning) |
Compatible ROCm versions
none — no ROCm release has ever listed gfx900 as a supported GPU target (gfx). ROCm support is granted per target, so this card is not covered by any version, including the current one. The unofficial HSA_OVERRIDE_GFX_VERSION workaround is the usual route people take — see ROCm on consumer Radeon GPUs for what it does and where it breaks.
Added 2026-09 as AMD's first Radeon Instinct card (predates this collection's previous oldest AMD entry, 2020's MI100, by three years) — the user specifically asked for older Radeon Instinct cards alongside NVIDIA's K80/M40. AMD's own current site no longer hosts this datasheet (amd.com/system/files/documents/radeon-instinct-mi25-datasheet.pdf 404s); pulled via the Wayback Machine's 2018 archive of that exact URL, read via pdftotext — an archived copy of AMD's own document, not a third-party remake.
fp16TFLOPSVector (24.6) and fp32TFLOPS (12.3) are Vega10's "Rapid Packed Math" throughput via its standard stream processors — correctly placed in this schema's vector field, not fp16TFLOPS/matrix, because Vega predates AMD's dedicated Matrix Cores entirely (introduced with CDNA/ MI100 three years later). No INT8 figure is published for this card at all (unlike the later MI50) — matrix/tensor-throughput fields are correctly all omitted, not zero. fp64TFLOPS (0.768, i.e. "768 GFLOPS") is explicitly stated as a 1/16th-of-FP32 rate in AMD's own footnote. eccSupport is "Yes" but per AMD's own footnote 3, "ECC support is limited to the HBM2 memory and ECC protection is not provided for internal GPU structures" — partial, unlike MI50's full-chip ECC the following generation.
No Infinity Fabric / GPU-to-GPU interconnect field populated — Infinity Fabric Link was introduced with MI50/MI60 the following generation; MI25 is PCIe Gen3 x16 only for both host and (if used) peer-to-peer traffic, confirmed by MI50's own datasheet footnote describing "previous Gen Radeon Instinct" cards as PCIe-Gen3-only. targetId (gfx900) doesn't cross-link to any `rocm` entry in this collection — every ROCm version tracked here (oldest: 5.0.0) already dropped Vega10/gfx900 support, so the empty ROCm-compatibility section on this GPU's page is accurate, not a data gap (the same situation as this round's K80 addition on the CUDA side). No transistor count or MSRP — AMD doesn't publish Instinct list prices (see this collection's MI100 entry for the same note).