Instinct MI210
AMDdatacenter| Overview | |
|---|---|
| Architecture | CDNA 2 (2022) |
| Process node | TSMC 6nm FinFET |
| GPU target (gfx) | gfx90a |
| Memory | |
| VRAM | 64 GB HBM2e |
| Memory bandwidth | 1,600 GB/s |
| ECC | Yes |
| Compute — vector | |
| FP64 | 22.6 TFLOPS |
| FP32 | 22.6 TFLOPS |
| Compute — matrix / tensor | |
| FP64 | 45.3 TFLOPS |
| BF16 | 181 TFLOPS |
| FP16 | 181 TFLOPS |
| INT8 | 181 TOPS |
| Cores & clocks | |
| Compute Units | 104 |
| Shader cores | 6,656 |
| Peak clock | 1,700 MHz |
| Board & system | |
| TDP | 300 W |
| Form factor | PCIe dual-slot FHFL |
| PCIe | PCIe 4.0 x16 |
| Interconnect | Infinity Fabric (3rd gen) |
| Multi-instance support | N/A |
Compatible ROCm versions
Added 2026-09-06 during a `gpus` coverage pass -- a single-GCD, PCIe add-in-card sibling of the already-tracked MI250X (same CDNA 2 generation, same gfx90a target), commonly deployed as a single-GPU on-prem/workstation ROCm accelerator where MI250X's dual-die OAM module isn't practical.
AMD's own product PDFs (instinct-mi210-brochure.pdf, amd-instinct-gpu-family-brochure.pdf) repeatedly timed out on direct fetch -- unlike most entries in this collection, these figures come from a secondary spec aggregator (flopper.io) and a semiconductor-analysis writeup (chipsandcheese.com's "AMD's Radeon Instinct MI210: GCN Lives On"), cross-checked for internal consistency against this site's already-verified MI250X entry rather than against AMD directly: fp64TFLOPSMatrix (45.3) is almost exactly 2x fp64TFLOPS (22.6), the same vector:matrix ratio MI250X's page publishes explicitly (47.9 -> 95.7); vramGB (64) and memoryBandwidthGBs (1600) are almost exactly half of MI250X's dual-GCD board-level figures (128 / 3276.8), consistent with MI210 being a single-GCD product built from the same die.
computeUnitCount (104) and shaderCoreCount (6656, derived from 104 CUs x 64 stream processors/CU -- the same ratio MI250X's page confirms directly, 14080/220=64) come from chipsandcheese specifically, not flopper.io; 104 CUs is noticeably less than a literal half of MI250X's 220 (110), which is plausible binning/yield variance between the two products rather than an error, but is flagged here since it's the one figure that doesn't cleanly reconcile.
int4TOPS and interconnectBandwidthGBs are omitted -- no MI210-specific figure was found in any source consulted, and int4TOPS on MI250X (383, identical to its int8TOPS) isn't confirmed to hold for MI210 specifically rather than assumed. No msrpUSD -- AMD doesn't publish Instinct list prices, matching every other Instinct entry in this collection.