[Docs] New docs proposal: Dedicated CPU + GPU Hardware Compatibility Matrices + rework current Install Guide

[incertum]

Proposing to add several hardware compatibility matrices to new documentation combined with more guidance around system dependencies -- what do I need when? What are common hiccups? Kind of expand https://ml-explore.github.io/mlx/build/html/install.html or rework it a bit (already tons of really great guidance in there!).

I could envision:

• a new / separate and more detailed "Hardware + System Requirements" section.
• moving all the build instructions to a dedicated "Build from source" guide which points to the requirements guide.
• moving all pure install instructions to a dedicated "Install mlx" guide which points to the requirements guide.
• ... and perhaps have a few new tables (such as proposed below).

---

(1)

Add GitHub badges for supported CPU arch and platform support. Perhaps this could be suitable for the main README.md in the GitHub repo?

macOS

Linux

---

(2)

New doc on https://ml-explore.github.io/mlx/ -> link to it from main README.

macOS GPU Support (Apple Silicon)

GPU acceleration on Apple Silicon leverages the Metal framework.

Supported GPU Architectures

Chip Family	Metal GPU Family	Status	Key Differentiating Features & Capabilities
M1	apple7	✅	TBA
M2	apple8	✅	TBA
M3	apple9	✅	TBA
M4	apple10	✅	TBA
M5	apple11	✅	TBA

---

(3)

New doc on https://ml-explore.github.io/mlx/ -> link to it from main README.

Linux GPU Support (CUDA)

Supported GPU Architectures

The table below outlines recommended NVIDIA GPU architectures and their specific feature support. A green checkmark (✅) indicates a fully supported and recommended architecture.

Example GPU	Architecture (Compute Capability)	Status	Key Accelerated Data Types
H100	Hopper (sm_90)	✅	FP8, TF32, BFLOAT16, FP16, INT8
RTX 4090	Ada Lovelace (sm_89)	✅	FP8, TF32, BFLOAT16, FP16, INT8
A100	Ampere (sm_80)	✅	TF32, BFLOAT16, FP16, INT8
T4, RTX 2080	Turing (sm_75)	✅	FP16, INT8
V100	Volta (sm_70)	✅	FP16, INT8

---

More Questions:

• Maybe the minimum supported CUDA Toolkit version or the recommended one may even be tied to the GPU or other system aspects? Could the generic guidance of CUDA toolkit >= 12.0 be more nuanced?
• Some "stuff" we install as part of the mlx CI such as MPI or NCCL seems not listed or mentioned in the current system requirements section?
• Build from source (system requirements sub-section) seems to lack a section for building mlx natively on Linux.
• ...