Hardware Detail GPUPrevious gen

AMD Instinct MI300X 192GB

AMD Instinct MI300X 192GB profile for local AI model planning.

AMDAMD INSTINCT2023-12192GB VRAM
AI Memory
192GB VRAM
Runtime headroom is estimated.
LLM Score
70
server
Power
600W
Average AI load
Value Score
6.1
Planner signal

Task Fit

Code100

Strong fit for code workloads in the current planner.

Chat100

Strong fit for chat workloads in the current planner.

RAG100

Strong fit for rag workloads in the current planner.

Agent97

Strong fit for agent workloads in the current planner.

Vision92

Strong fit for vision workloads in the current planner.

Image generation88

Strong fit for image generation workloads in the current planner.

Embedding80

Usable for embedding, with tradeoffs around speed, memory, or runtime maturity.

Voice75

Usable for voice, with tradeoffs around speed, memory, or runtime maturity.

Video generation73

Usable for video generation, with tradeoffs around speed, memory, or runtime maturity.

Runtime Support

ollamaPartial
llama.cppgood
LM StudioPartial
vllmPartial
sglangPartial
ComfyUIgood
diffusersPartial

Specs

CPU cores—
GPU cores—
System RAM—
Unified memory—
VRAM192GB
Usable model memory192GB
CUDANo
MetalNo
ROCmSupported

Cost and Ops

Hardware price$15,000
Build cost$3,000
UPS / accessories$200
5-year estimate$20,102
Expected lifespan6 years
Residual value18%
Build requiredYes

Models This Hardware Can Run

Estimated from default model artifacts, VRAM, unified memory, and recommended RAM.

62 compatible defaults
DeepSeek-R1-Distill 70B
Compatible

DeepSeek · Q4_K_M

80GB recommended
DeepSeek-R1-Distill 32B
Compatible

DeepSeek · Q4_K_M

48GB recommended
Qwen3 235B-A22B
Compatible

Alibaba · Q2_K

192GB recommended
Phi-4 14B
Compatible

Microsoft · Q4_K_M

24GB recommended
GLM 5.1 FP8
Compatible

Z.ai · FP8

96GB recommended
GLM 4.7 Flash
Compatible

Z.ai · Q4_K_M

32GB recommended
Kimi K2.7 Code
Compatible

Moonshot AI · Q4_K_M

48GB recommended
Kimi K2.6
Compatible

Moonshot AI · Q4_K_M

48GB recommended
Kimi K2 Thinking
Compatible

Moonshot AI · Q4_K_M

48GB recommended

Best For

huge VRAMROCm data centersingle accelerator large models

AMD Instinct MI300X 192GB profile for local AI model planning.

Source Confidence

Specsreviewed
PriceEstimated
PowerEstimated
PerformanceEstimated
Updated2026-06-18

Similar Hardware

NVIDIA RTX 6000 Blackwell 96GB
NVIDIA · 96GB VRAM · $12,376 / 5y

NVIDIA RTX 6000 Blackwell 96GB profile for local AI model planning.

NVIDIA A100 80GB
NVIDIA · 80GB VRAM · $12,156 / 5y

NVIDIA A100 80GB profile for local AI model planning.

NVIDIA H100 80GB
NVIDIA · 80GB VRAM · $29,822 / 5y

NVIDIA H100 80GB profile for local AI model planning.