Multi-Chip Support

KernelGenBench supports six hardware platforms with automatic device detection and unified execution pipeline.

Supported Platforms

Platform	Description	Notes
NVIDIA	A100 GPUs	Primary baseline
Ascend NPU	Huawei AI accelerators	—
MUSA	Moore Threads GPUs	—
Hygon DCU	Hygon data center accelerators	—
Iluvatar	Iluvatar AI chips	—
MetaX	MUXI accelerators	—

Auto-Detection

Device type is automatically detected at runtime:

# Check detected device
python -c "from runtime import get_device_type; print(get_device_type())"

Unified Commands

All platforms use the same commands — the framework handles device differences automatically:

# Same command works on all platforms
python scripts/generate_kernel_and_verify.py \
    --server-type openai \
    --model-name gpt-4o

Platform-Specific Behavior

Dataset Selection

Platform	Default Dataset
NVIDIA	KernelGenBench (210 operators)
Others	KernelGenBench-aten (110 operators)

On non-NVIDIA platforms, vLLM and cuBLAS operators are unavailable.

Anti-Hack Layers

Layer	NVIDIA	Non-NVIDIA
L1: AST Static Scan	✓	✓
L2: Ghost Replay	✓	✓
L3: Hardware Profiling	✓	✗

L3 profiling is NVIDIA-only due to tool availability.

Tolerance Settings

Numerical tolerances are automatically adjusted per platform to account for different floating-point implementations.

Cross-Platform Challenges

Compiler Maturity

Non-NVIDIA platforms have:

• Less mature Triton compilers

• Incomplete backend support

• Different memory models

Performance Impact

• Non-NVIDIA platforms require ~2× more tokens and time

• Cross-platform degradation can be severe

• Platform-specific optimizations needed

Hardware-Specific Templates

The framework injects platform-specific code templates:

• Import statements

• Runtime configurations

• Memory constraints

• Device-specific constants