llmrun

AMD · CDNA 2

Best AI Models for AMD Instinct MI210 (64 GB)

Runs 70B models well; 120B MoE only just fits.

64 GB HBM2e1638 GB/s181 TFLOPS FP16300 WCDNA 2 2022

What LLMs can the Instinct MI210 run?

#ModelSpeedllmrun ScoreQuant · VRAM
1
Qwen3.8 27BAlibabaHow to run →
Q4_K_M · 17.9 GB · 56.4 t/s
56.4 tok/sQ4_K_M · 17.9 GB
2
Qwen3.6 27BAlibaba
Q4_K_M · 17.9 GB · 56.4 t/s
56.4 tok/sQ4_K_M · 17.9 GB
3
Gemma 4 31B ITGoogle
Q4_K_M · 21.7 GB · 48.2 t/s
48.2 tok/sQ4_K_M · 21.7 GB
4
Gemma 4 26B A4B ITGoogle
Q4_K_M · 16.5 GB · 114 t/s
114 tok/sQ4_K_M · 16.5 GB
5
NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA
Q4_K_M · 19.3 GB · 84.2 t/s
84.2 tok/sQ4_K_M · 19.3 GB
See all 373 models on the Instinct MI210 →Best quant that fits · speeds estimated on this card

About the Instinct MI210

The Instinct MI210 runs 70B models well. CDNA 2 generation, released 2022.

GenerationCDNA 2released 2022

Instinct MI210 strengths and weaknesses

Strengths

  • Runs 120B MoE models: GPT OSS 120B at 96.5 tok/s in 63.5 GB
  • Room for long context: Llama 3.3 70B Instruct at Q4 uses 51.3 GB, leaving 12.7 GB free

Weaknesses

  • Slow prompt reading: A 4K-token prompt on Qwen3.8 27B waits 5.7 s for the first token
  • ROCm, not CUDA: No FP8 weights; MXFP4 only through llama.cpp, not vLLM

Instinct MI210 tokens per second by model size

Output speed depends on memory bandwidth (1638 GB/s).

  1. 9BQwen3.5 9B · Q4_K_M · 6.6 GB
    155tok/s
  2. 14BQwen3 14B · Q4_K_M · 11.9 GB
    104tok/s
  3. 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
    56.4tok/s
  4. 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
    91.3tok/s
  5. 70BLlama 3.3 70B Instruct · Q4_K_M · 51.3 GB
    21.1tok/s
  6. 120B MoEGPT OSS 120B · IQ4_XS · 63.5 GB tight · IQ4_XS
    96.5tok/s

Estimated speeds: Q4_K_M, 4K context, one user. How we estimate

All 373 models on the Instinct MI210

Featured only

Best quant that fits, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works

LLM models compatible with AMD Instinct MI210 — ranked by Score, recency and Fit
ModelVRAMScoreFit
Qwen3.8 27B
Q4_K_M·56.4 t/s
17.9 GBA82, Runs well
Muse Glimmer 30B
Q4_K_M·53.6 t/s
19.6 GBA81, Runs well
Gemma 4 31B IT
Q4_K_M·48.2 t/s
21.7 GBA79, Runs well
NVIDIA Nemotron 3.5 Lightning 30B A3B BF16
Q4_K_M·84.2 t/s
19.3 GBS88, Runs great
GLM 4.7 Flash
Q4_K_M·80.8 t/s
19.9 GBS88, Runs great
Qwen3.6 35B A3B
Q4_K_M·91.3 t/s
22.0 GBS90, Runs great
JEV 27B VL
Q4_K_M·56.4 t/s
17.9 GBA82, Runs well
Gemma 4 26B A4B IT
Q4_K_M·113.7 t/s
16.5 GBS94, Runs great
Flint
Q4_K_M·88.1 t/s
25.0 GBS89, Runs great
LFM2.5 24B A2B
Q4_K_M·104.7 t/s
15.0 GBS92, Runs great
MiniCPM V 4.7 35B A3B
Q4_K_M·95.3 t/s
22.1 GBS90, Runs great
GPT OSS 20B
Q4_K_M·134.5 t/s
13.1 GBS95, Runs great
MiMo V2.6 Distill Qwen 9B
Q4_K_M·158.3 t/s
6.5 GBS97, Runs great
Sage A1.5
Q4_K_M·91.3 t/s
22.0 GBS90, Runs great
Qwen3.5 9B
Q4_K_M·154.6 t/s
6.6 GBS97, Runs great

Instinct MI210 specs, power and PSU

Specs

Architecture
CDNA 2
VRAM
64 GB HBM2e
Memory bandwidth
1638.4 GB/s
FP16 (dense)
181.00 TFLOPS
Stream processors
6,656
TDP
300 W
Released
2022-03-22

Instinct MI210 FAQ

What is the largest model the Instinct MI210 can run?

GPT OSS 120B (116.8B) is the largest at 4-bit or better: at IQ4_XS it uses 63.5 GB of 64 GB, and runs at about 96.5 tok/s.

Can the Instinct MI210 run Qwen3.8 27B?

Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of 64 GB, and it is the best model that runs well on this card.

How many tokens per second does the Instinct MI210 give?

About 56.4 tok/s on Qwen3.8 27B at Q4_K_M, with the first token after 5.7 s on a 4K-token prompt.

What power supply does the Instinct MI210 need?

Plan a 600 W PSU (2× the 300 W TDP).

Is 64 GB enough for coding and agents?

Yes for coding: Qwen3.8 27B, the best of the 11 coding models that fit, runs at about 56 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.

Related

Best coding LLMs48 GB VRAMGPUs with 64 GB or more