FitMyLLM
AMD· CDNA 2.0

AMD Radeon Instinct MI200

Running LLMs on the Radeon Instinct MI200 — the long read: which models fit at which quantisation, and the settings worth changing. · Or what else $10,000 buys

VRAM
64 GB
FLAGSHIP
BANDWIDTH
1640
GB/S
MODELS Q4
373/449
83%
7B Q4 SPEED
~208
BLAZING
▸ MODEL COVERAGE @ Q483% OF ALL
▸ ESTIMATED SPEED· BY MODEL SIZE @ Q4

Average speeds at Q4 quantization. Actual performance varies by model architecture and context length.

3B
~486
TOK/S
7B
~208
TOK/S
14B
~104
TOK/S
32B
~46
TOK/S
70B
~21
TOK/S
▸ MEASURED RIG REPORTS

We rent the machine and time every model on it: decode, VRAM peak, concurrency, watts and cost per million tokens. Yours may already be one of them — and two are free to read in full.

SEE THE REPORTS
▸ SPECIFICATIONS
VRAM
64 GB
BANDWIDTH
1640 GB/s
FP16 COMPUTE
181 TFLOPS
TDP
300W
MEMORY
HBM2e
ARCHITECTURE
CDNA 2.0
STREAM PROCESSORS
6,656
COMPUTE UNITS
104
PCIE
Gen 4 x16
MSRP
$10,000
347
FAST MODELS · >30 TOK/S
Real-time chat speed
373
USABLE · >10 TOK/S
Comfortable for all tasks
373
TOTAL COMPATIBLE
Fit in VRAM at Q4
▸ RENT IT IN THE CLOUD

Buying Radeon Instinct MI200 costs $15–$40k and isn’t practical for most teams. Spin one up by the hour instead:

Spin up in ~60s. Pay by the second. Cancel anytime.

Some links are affiliate links — we may earn a small commission at no extra cost to you. This helps keep FitMyLLM free and independent.

▸ COMPATIBLE MODELS· 373
S
Falcon-H1R Tiny 90M0.09B
FALCON·256K CTX· CHAT· REASONING
16198
TOK/S · 1% VRAM
S
SmolLM2 135M0.135B
SMOLLM·2K CTX· CHAT
10798
TOK/S · 1% VRAM
S
nomic-embed-text-v1.5 100M0.14B
EMBEDDING·8K CTX· CHAT
10413
TOK/S · 1% VRAM
S
GPT-2 124M0.14B
GPT2·1K CTX· CHAT
10413
TOK/S · 1% VRAM
S
SmolVLM 256M0.256B
SMOLLM·8K CTX· CHAT· VISION
5694
TOK/S · 1% VRAM
S
Gemma 3 270M0.27B
GEMMA·32K CTX· CHAT
5399
TOK/S · 1% VRAM
S
Snowflake Arctic Embed M v2.00.305B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
4780
TOK/S · 1% VRAM
S
bge-large-en-v1.5 335M0.335B
EMBEDDING·1K CTX· CHAT
4352
TOK/S · 1% VRAM
S
mxbai-embed-large-v10.335B
EMBEDDING·1K CTX· EMBEDDING
4352
TOK/S · 1% VRAM
S
Snowflake Arctic Embed L0.335B
EMBEDDING·1K CTX· EMBEDDING
4352
TOK/S · 1% VRAM
S
LFM2 350M0.35B
LFM·125K CTX· CHAT· TOOL_USE
4165
TOK/S · 1% VRAM
S
SmolLM2 360M0.36B
SMOLLM·8K CTX· CHAT
4049
TOK/S · 1% VRAM
S
GPT-2 Medium 345M0.38B
GPT2·1K CTX· CHAT
3836
TOK/S · 1% VRAM
S
Qwen 2.5 0.5B0.5B
QWEN·32K CTX· CHAT
2916
TOK/S · 1% VRAM
S
SmolVLM 500M0.5B
SMOLLM·8K CTX· CHAT· VISION
2916
TOK/S · 1% VRAM
S
Falcon-H1 0.5B0.52B
FALCON·128K CTX· CHAT
2803
TOK/S · 1% VRAM
S
BGE-M30.568B
EMBEDDING·8K CTX· EMBEDDING
2567
TOK/S · 1% VRAM
S
Snowflake Arctic Embed L v2.00.568B
EMBEDDING·8K CTX· EMBEDDING· MULTILINGUAL
2567
TOK/S · 1% VRAM
S
Qwen3-Embedding 0.6B0.6B
EMBEDDING·32K CTX· EMBEDDING· MULTILINGUAL
2430
TOK/S · 1% VRAM
S
Falcon-H1R Tiny 0.6B0.6B
FALCON·32K CTX· CHAT· REASONING
2430
TOK/S · 1% VRAM
S
Falcon Perception 0.6B0.6B
FALCON·4K CTX· VISION
2430
TOK/S · 1% VRAM
S
Qwen 1.5 0.5B0.62B
QWEN·32K CTX· CHAT
2351
TOK/S · 1% VRAM
S
LFM2 700M0.74B
LFM·125K CTX· CHAT· TOOL_USE
1970
TOK/S · 1% VRAM
S
Qwen3 0.6B0.75B
QWEN·32K CTX· CHAT· REASONING
1944
TOK/S · 1% VRAM
S
GPT-2 Large 774M0.81B
GPT2·1K CTX· CHAT
1800
TOK/S · 2% VRAM
S
Qwen 3.5 0.8B0.87B
QWEN·256K CTX· CHAT· CODING· MULTILINGUAL
1676
TOK/S · 2% VRAM
S
InternVL3 1B0.94B
OTHER·32K CTX· CHAT· VISION
1551
TOK/S · 2% VRAM
S
MiniCPM5 1B1.08B
MINICPM·128K CTX· CHAT· REASONING· MULTILINGUAL
1350
TOK/S · 2% VRAM
S
TinyLlama 1.1B1.1B
LLAMA·2K CTX· CHAT
1325
TOK/S · 2% VRAM
S
LFM2.5-1.2B-Thinking1.2B
LFM·122K CTX· CHAT· REASONING· TOOL_USE
1215
TOK/S · 2% VRAM
S
Llama-3.2-1B1.2B
LLAMA·4K CTX· CHAT
1215
TOK/S · 2% VRAM
S
LFM2 1.2B1.2B
LFM·125K CTX· CHAT· TOOL_USE· MULTILINGUAL
1215
TOK/S · 2% VRAM
S
Zamba2 1.2B1.2B
OTHER·4K CTX· CHAT
1215
TOK/S · 2% VRAM
S
EXAONE-4.0-1.2B1.3B
EXAONE·64K CTX· CHAT
1121
TOK/S · 2% VRAM
S
OPT 1.3B1.3B
OPT·2K CTX· CHAT
1121
TOK/S · 2% VRAM
S
MiniCPM-V 4.61.3B
OTHER·256K CTX· CHAT· VISION
1121
TOK/S · 2% VRAM
S
DeepSeek Coder 1.3B1.35B
DEEPSEEK·16K CTX· CODING
1080
TOK/S · 2% VRAM
S
Phi-1 1.3B1.42B
PHI·2K CTX· CODING
1027
TOK/S · 2% VRAM
S
Phi-1.5 1.3B1.42B
PHI·2K CTX· CHAT· CODING
1027
TOK/S · 2% VRAM
S
LFM2 8B A1B8.3BMoE
LFM·125K CTX· CHAT· TOOL_USE· MULTILINGUAL
972
TOK/S · 9% VRAM
▸ NEXT STEP

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your Radeon Instinct MI200.

WHAT THIS CARD IS WORTH

Radeon Instinct MI200 holds 373 of the models in our catalogue and is, in practice, a Q4_K_S card — the largest it takes is Llama-3.2-90B-Vision-Instruct at Q4_K_S.

TOKENS/SEC PER $100
1.8
8B at Q4_K_M, so cards compare like for like
VRAM PER $100
0.64 GB
what memory costs on this card
THE BIGGEST IT TAKES
Llama-3.2-90B-Vision-Instruct90B · Q4_K_S56.3 GB20 tok/sEST
Hunyuan A13B80B · Q4_K_M51.9 GB111 tok/sEST
Qwen3-Coder-Next80B · Q4_K_M51.5 GB345 tok/sEST
Qwen3-Next 80B A3B80B · Q4_K_M51.5 GB345 tok/sEST
NVLM-D 72B79.38B · Q4_K_M52.3 GB22 tok/sEST
InternVL3 78B78B · Q4_K_M51.4 GB22 tok/sEST
Qwen2.5-72B72.7B · Q5_K_M55.5 GB21 tok/sEST
Qwen2-VL 72B72.7B · Q5_K_M55.5 GB21 tok/sEST
AGAINST CARDS OF SIMILAR MEMORY
RTX 6000D84 GB$7,5001.9 tok/s per $100
H100 SXM5 80GB80 GB$25,0000.1 tok/s per $100
H100 CNX80 GB$25,0000.2 tok/s per $100
A100 SXM 80GB80 GB$10,0000.4 tok/s per $100
Radeon Instinct MI20064 GB$10,0001.8 tok/s per $100

Neighbours in memory rather than in price: memory decides whether a card can do the job at all, so two cards of the same size at different prices is the comparison you are making. Ordered by memory, then bandwidth — not by the value column, which is worked out from bandwidth and price alone and therefore rewards a cheap card whatever its software stack does to that bandwidth in practice. Read it as one input, not as a ranking.

▸ DEVICE UNDER TEST

AMD Radeon Instinct MI200 — 64 GB VRAM.

RADEON INSTINCT MI200 SPEC
BRAND
AMD
VRAM
64 GB HBM2e
BANDWIDTH
1640 GB/s
FP16 COMPUTE
181 TFLOPS
FP32 COMPUTE
22.6 TFLOPS
STREAM PROCESSORS
6,656
TDP
300 W
ARCHITECTURE
CDNA 2.0
MSRP
$10000
▸ AI CAPABILITY
373/ 449 models @ Q4

With 64 GB VRAM and 1640 GB/s bandwidth, this GPU handles models up to 80B parameters.

Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~187 tok/s.

RENT IT — LIVE PRICES
Checking what the clouds are charging…
§ 01TOP MODELS FOR RADEON INSTINCT MI200
373 FIT · SHOWING 20
MODELSIZEVRAM Q4TOK/SAVG
Hunyuan A13B80B49.4 GB11281.1
Qwen3-Coder-Next80B49.4 GB48643.0
Qwen3-Next 80B A3B80B49.4 GB48649.0
NVLM-D 72B79.38B49.0 GB1848.7
InternVL3 78B78B48.2 GB1980.6
Qwen2.5-72B72.7B44.9 GB2039.7
Qwen2-VL 72B72.7B44.9 GB2055.5
Qwen 1.5 72B72B44.5 GB2049.7
Qwen2 Math 72B72B44.5 GB2049.7
Molmo 72B72B44.5 GB2054.1
DeepSeek R1 Distill Llama 70B70.6B43.6 GB2142.4
Llama 3.3 70B70.6B43.6 GB2144.8
Llama 3.1 70B70.6B43.6 GB2133.2
Llama 3 70B70.6B43.6 GB2144.1
Llama-3.1-Nemotron-70B70.6B43.6 GB2143.7
Cogito 70B70B43.3 GB21
Llama 2 70B70B43.3 GB2133.4
CodeLlama 70B70B43.3 GB2145.7
Dolphin Llama 3 70B70B43.3 GB2145.7
Tulu 3 70B70B43.3 GB2159.4