Skip to content
Llama Builds

AMD · Vega 20 (GCN) · 2018

AMD Radeon Instinct MI50 32GB for local AI

Huge bandwidth per dollar on paper, but ROCm has dropped it and shortage pricing erased the bargain.

Updated · Speeds marked ~ are estimates; seemethodology

VRAM
32GB
Bandwidth
1,024 GB/s
Power
300W
Runs well
8 models

MI50 32GB

Typically about $600 used

Check price at AmazonFind it used on eBay

Can the MI50 32GB run LLMs?

The MI50 32GB has 32GB of HBM2 and 1,024 GB/s of memory bandwidth, which decides how big a model fits and how fast it generates text. It comfortably runs models up to about 35B parameters: the largest in our catalog is Qwen3.6 35B-A3B at Q5_K_M (28GB with 8K context, ~76 tokens/sec). Our top all-round pick for it is Qwen3.8 27B (Q6_K, ~19 tokens/sec). Llama 3.3 70B Instruct only runs with partial CPU offload, at about 2.5 tokens/sec with CPU offload.

Recommendations

Best models for the MI50 32GB

  1. Best overall · also coding, reasoning, vision

    Qwen3.8 27B

    Q6_K · 25GB · ~19 tokens/sec

    Runs well

  2. Runner-up · also fastest, largest that fits

    Qwen3.6 35B-A3B

    Q5_K_M · 28GB · ~76 tokens/sec

    Runs well
Read: the best LLMs to run on the MI50 32GB

Compatibility

What runs on the MI50 32GB

Every current model in our catalog, at the best quantization that fits in 32GB with 8K tokens of context. "Partial offload" means part of the model runs from system RAM (we assume 64GB), which is much slower.

Which open-weight models run on the AMD Radeon Instinct MI50 32GB, with the best quantization and estimated speed
ModelParametersBest quantMemorySpeedVerdict
Gemma 4 E4B8BQ8_010 GB~44 tok/sRuns well
Qwen3.5 9B9BQ8_012 GB~39 tok/sRuns well
Phi-414.7BQ8_018 GB~28 tok/sRuns well
gpt-oss-20b20.9B (3.6B active)MXFP415 GB~74 tok/sRuns well
Gemma 4 26B A4B26B (3.8B active)Q6_K23 GB~61 tok/sRuns well
Qwen3.8 27B27BQ6_K25 GB~19 tok/sRuns well
Gemma 4 31B31BQ5_K_M26 GB~19 tok/sRuns well
Qwen3.6 35B-A3B35B (3B active)Q5_K_M28 GB~76 tok/sRuns well
GLM-4.5-Air106B (12B active)Q4_K_M68 GB~9.2 tok/sPartial offload
Llama 4 Scout109B (17B active)Q4_K_M70 GB~6.6 tok/sPartial offload
gpt-oss-120b117B (5.1B active)MXFP468 GB~22 tok/sPartial offload
Qwen3.5 122B-A10B122B (10B active)Q4_K_M77 GB~10 tok/sPartial offload
DeepSeek V4 Flash284B (13B active)———Won't fitWhat it needs
GLM-5.3-Flash320B (18B active)———Won't fitWhat it needs
GLM-5.3744B (40B active)———Won't fitWhat it needs

Performance

MI50 32GB vs. alternatives on Qwen3.8 27B

Generation speed in tokens per second for a single user. Token generation is limited mostly by memory bandwidth, which is why bandwidth predicts these results so well.

  1. V100 32GB

    Q6_K · 32GB

    ~25tok/s

  2. MI50 32GB

    Q6_K · 32GB

    ~19tok/s

  3. R9700

    Q6_K · 32GB

    ~18tok/s

  4. Tesla P40

    Q4_K_M · 24GB

    ~11tok/s

Estimated (method)

Measured

MI50 32GB benchmark results

ModelQuantRuntimeGenerationPromptSource
Llama 2 7B ChatQ4_0llama.cpp (Vulkan)110 tok/s1120 tok/sllama.cpp discussion #10879: Performance of llama.cpp with Vulkan · October 2026
Llama 2 7B ChatQ4_0llama.cpp (ROCm (HIP))99 tok/s1055 tok/sllama.cpp discussion #15021: Performance of llama.cpp on AMD ROCm (HIP) · October 2026

Our take

Should you buy the MI50 32GB for AI?

The Instinct MI50 32GB is a Vega 20 datacenter card from AMD's 2018 lineup with 32GB of HBM2 on a 4096-bit bus for 1,024 GB/s. In 2025 it became a favorite for cheap LLM rigs when imported cards sold for around $150, but shortage pricing pushed listings to about $600 to $830 by October 2026. Current ROCm 7 releases no longer support its gfx906 GPU, so most owners run llama.cpp through Vulkan or older ROCm builds. It is passively cooled, so desktop use needs a fan shroud and good airflow.

What we like

  • 1,024 GB/s HBM2 bandwidth
  • 32GB per card
  • Works with llama.cpp through Vulkan

Watch out for

  • Not supported by current ROCm releases
  • Passive cooling needs a fan shroud
  • Prices roughly quadrupled during the memory shortage

Best for: Tinkerers building high-bandwidth rigs · Multi-card llama.cpp servers

Specifications

MI50 32GB specs that matter for AI

Memory
32GB HBM2
Memory bandwidth
1,024 GB/s
Memory bus
4096-bit
Architecture
Vega 20 (GCN)
Power
300W
Interface
PCIe 4.0 x16
Released
2018
Street price (used)
~$600 (2026-10)
Software
ROCm, Vulkan

AMD launched the MI50 with 16GB in 2018. Make sure you buy the 32GB version.

Sources: VideoCardz.net Radeon Instinct MI50 32GB specifications, Wikipedia, AMD Instinct

Cross-shop

Alternatives to the MI50 32GB

NVIDIA2018
32GB

HBM2 · 900 GB/s

V100 32GB

~$700 used

NVIDIA2016
24GB

GDDR5 · 346 GB/s

Tesla P40

~$230 used

AMD2025
32GB

GDDR6 · 640 GB/s

R9700

~$1,800

Frequently asked questions

How much VRAM does the MI50 32GB have?

The AMD Radeon Instinct MI50 32GB has 32GB of HBM2 with 1,024 GB/s of memory bandwidth on a 4096-bit bus.

What is the largest LLM the MI50 32GB can run?

Fully in GPU memory, the largest model in our catalog that runs comfortably is Qwen3.6 35B-A3B (35B (3B active)) at Q5_K_M, using about 28GB with 8K context at ~76 tokens/sec.

Can the MI50 32GB run Llama 3.3 70B?

Only with partial CPU offload. At Q4_K_M the model needs about 47GB with 8K context, so part of it spills into system RAM and speed drops to about 2.5 tokens/sec with CPU offload. For usable 70B speeds you want 48GB of VRAM or more (for example two 24GB cards).

What is the best LLM to run on the MI50 32GB?

Our top pick is Qwen3.8 27B at Q6_K, which fits in 32GB with room for 8K of context and runs at ~19 tokens/sec. See our full list of the best models for the MI50 32GB for coding, reasoning, and vision picks.

Is the MI50 32GB good for local AI in 2026?

Huge bandwidth per dollar on paper, but ROCm has dropped it and shortage pricing erased the bargain.