# Please add "AMD Radeon AI PRO R9700" to "My Hardware"

> Source: <https://discuss.huggingface.co/t/please-add-amd-radeon-ai-pro-r9700-to-my-hardware/171780#post_7>
> Published: 2026-09-18 06:52:19+00:00

The R9700 is a good candidate for “My Hardware”: RDNA4 architecture, 32 GB GDDR6, high FP8/INT8 throughput and a memory profile that fits 30B–70B models in Q4_K_M. Adding it would help AMD users get accurate Fit Finder results and improve hardware coverage for ROCm/DirectML setups.

Below is a small diagnostic script I use to classify a GPU for local‑AI workloads.

It doesn’t benchmark anything — it simply evaluates the specs and produces a profile that tells which model sizes fit and how strong the GPU is for FP8/INT8 inference.

``` python
def gpu_ai_profile(memory_gb, fp8_tflops, int8_tops, architecture):
    """
    This function builds an 'AI profile' for a GPU based only on its specs.
    It answers three practical questions:
    1. Which model sizes fit in VRAM?
    2. How strong is the GPU for FP8/INT8 inference?
    3. What architecture family does it belong to?
    """

    profile = {}

    # VRAM capacity → determines which GGUF sizes fit
    if memory_gb >= 32:
        profile["models_fit"] = ["30B", "34B", "40B", "70B (Q4_K_M, borderline)"]
    else:
        profile["models_fit"] = ["7B", "13B", "20B"]

    # FP8 throughput → good indicator for multimodal and MoE models
    profile["fp8_class"] = (
        "high" if fp8_tflops >= 300 else
        "medium" if fp8_tflops >= 100 else
        "low"
    )

    # INT8 throughput → relevant for GGUF quantized inference
    profile["int8_class"] = (
        "high" if int8_tops >= 300 else
        "medium" if int8_tops >= 100 else
        "low"
    )

    # Architecture tag
    profile["arch"] = architecture

    return profile

# Example using the R9700 specs:
print(gpu_ai_profile(
    memory_gb=32,
    fp8_tflops=383,
    int8_tops=383,
    architecture="RDNA4"
))
```