inference/research
verifiedhigh · 0.85updated 2026-10-03

AMD Instinct MI210

amd CDNA2 (gfx90a) 2022

Fields

id
accelerators/amd-instinct-mi210
Vendor
amd
Architecture
CDNA2 (gfx90a)
Release year
2022
Process (nm)
6
Form factors
pcie
VRAM (GB)
64
Memory type
hbm2e
Memory bus (bit)
4,096
Memory bandwidth (GB/s)
1,600
Bandwidth basis
4096-bit interface x 1.6 GHz memory clock = 1.6 TB/s, stated directly in AMD's spec table. Half the MI250X's 8192-bit bus, and exactly half its bandwidth - this part is a genuine single-GCD CDNA2 card, unlike the MI250 whose memory does not halve.
Peak FLOPS
precisionfp32tflops22.6denseyesvendor_claimyessource_idamd-mi210-product-pagebasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp32 for AMD Instinct MI210 (source of record: 'AMD Instinct MI210 product page', kind: spec-sheet). Recorded dense:true as printed.precisionfp32tflops45.3denseyesvendor_claimyessource_idamd-mi210-product-pagebasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp32 for AMD Instinct MI210 (source of record: 'AMD Instinct MI210 product page', kind: spec-sheet). Recorded dense:true as printed.precisionfp16tflops181denseyesvendor_claimyessource_idamd-mi210-product-pagebasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp16 for AMD Instinct MI210 (source of record: 'AMD Instinct MI210 product page', kind: spec-sheet). Recorded dense:true as printed.precisionbf16tflops181denseyesvendor_claimyessource_idamd-mi210-product-pagebasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own bf16 for AMD Instinct MI210 (source of record: 'AMD Instinct MI210 product page', kind: spec-sheet). Recorded dense:true as printed.precisionint8tflops181denseyesvendor_claimyessource_idamd-mi210-product-pagebasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own int8 for AMD Instinct MI210 (source of record: 'AMD Instinct MI210 product page', kind: spec-sheet). Recorded dense:true as printed.precisionint4tflops181denseyesvendor_claimyessource_idamd-mi210-product-pagebasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own int4 for AMD Instinct MI210 (source of record: 'AMD Instinct MI210 product page', kind: spec-sheet). Recorded dense:true as printed.
Unified memory
no
Consumer part
no
notes
AUDIT 2026-10-03 (AMD convention check, no value changed): AMD's Instinct product pages print a 'Peak X Performance' row (DENSE) and a separate 'Peak X Performance with Structured Sparsity' row (SPARSE, exactly 2x the dense value) only from the CDNA3/MI300 generation onward. The CDNA2 pages for MI210 (single-GCD CDNA2) print NO structured-sparsity variant for any precision - only 'Peak Half Precision (FP16)', 'Peak INT4', 'Peak INT8' and 'Peak bfloat16' - so every dense flag on this record is correct under AMD's own labelling. Concretely, on these pages AMD prints FP16 == INT8 == INT4 == BF16 at the same number (181.0 on the MI210, 362.1 on the MI250, 383.0 on the MI250X), all dense and all equal. That the four precisions are numerically identical is the CDNA2 property, not a sparsity artifact: unlike CDNA4 (MI350), CDNA2 has no 2:4 structured-sparsity path to double anything. The single-die CDNA2 part, and the last MI-series card sold as a normal PCIe add-in rather than an OAM module. 104 CUs / 6,656 SPs - exactly half the MI250X's 220/14,080 - so 181 TFLOPS FP16 vs 383, 22.6/45.3 TFLOPS FP32 vector/matrix vs 47.9/95.7. Unlike the MI250, the MI210's memory DOES halve with the compute: 64 GB HBM2e on 4096 bits at 1.6 TB/s, versus the MI250X's 128 GB / 8192-bit / 3.2 TB/s. Same 1.6 GHz memory clock in both cases. tdp_w is null because AMD's MI210 page has no TDP or TBP field at all (the MI100/MI250X pages do). The card is passively cooled, full height, 267mm, double slot, and draws auxiliary power from the PCIe slot; the 300W-class figure commonly quoted for MI210-family cards is not confirmed on this page, so it is left null rather than invented. 3 Infinity Fabric links at 100 GB/s each (the MI250X OAM has 8) - this is a real scaling difference for multi-GPU inference on PCIe cards: no all-to-all mesh. Same generation limits: no TF32, no FP8, no structured sparsity (MI300X datasheet footnote). All flops entries dense-only. INT8 and INT4 are both 181 TOPs - no 4-bit advantage on CDNA2. ROCm LLVM target gfx90a. Matrix FP32 is 2x vector FP32, the same CDNA2 asymmetry as MI250X.

Notes

AUDIT 2026-10-03 (AMD convention check, no value changed): AMD's Instinct product pages print a 'Peak X Performance' row (DENSE) and a separate 'Peak X Performance with Structured Sparsity' row (SPARSE, exactly 2x the dense value) only from the CDNA3/MI300 generation onward. The CDNA2 pages for MI210 (single-GCD CDNA2) print NO structured-sparsity variant for any precision - only 'Peak Half Precision (FP16)', 'Peak INT4', 'Peak INT8' and 'Peak bfloat16' - so every dense flag on this record is correct under AMD's own labelling. Concretely, on these pages AMD prints FP16 == INT8 == INT4 == BF16 at the same number (181.0 on the MI210, 362.1 on the MI250, 383.0 on the MI250X), all dense and all equal. That the four precisions are numerically identical is the CDNA2 property, not a sparsity artifact: unlike CDNA4 (MI350), CDNA2 has no 2:4 structured-sparsity path to double anything. The single-die CDNA2 part, and the last MI-series card sold as a normal PCIe add-in rather than an OAM module. 104 CUs / 6,656 SPs - exactly half the MI250X's 220/14,080 - so 181 TFLOPS FP16 vs 383, 22.6/45.3 TFLOPS FP32 vector/matrix vs 47.9/95.7.

Unlike the MI250, the MI210's memory DOES halve with the compute: 64 GB HBM2e on 4096 bits at 1.6 TB/s, versus the MI250X's 128 GB / 8192-bit / 3.2 TB/s. Same 1.6 GHz memory clock in both cases.

tdp_w is null because AMD's MI210 page has no TDP or TBP field at all (the MI100/MI250X pages do). The card is passively cooled, full height, 267mm, double slot, and draws auxiliary power from the PCIe slot; the 300W-class figure commonly quoted for MI210-family cards is not confirmed on this page, so it is left null rather than invented.

3 Infinity Fabric links at 100 GB/s each (the MI250X OAM has 8) - this is a real scaling difference for multi-GPU inference on PCIe cards: no all-to-all mesh.

Same generation limits: no TF32, no FP8, no structured sparsity (MI300X datasheet footnote). All flops entries dense-only. INT8 and INT4 are both 181 TOPs - no 4-bit advantage on CDNA2.

ROCm LLVM target gfx90a. Matrix FP32 is 2x vector FP32, the same CDNA2 asymmetry as MI250X.

Cross-references

Interconnect (2)