---

name: gpu-architecture-advisor description: Provide architecture-specific GPU optimization advice for NVIDIA GPUs (Ampere, Hopper, Ada). Use when discussing GPU architecture features, compute capability, or hardware-specific optimizations. version: 1.0.0

GPU Architecture Advisor Skill

This skill provides architecture-specific optimization guidance for NVIDIA GPUs.

When I activate

I automatically activate when you:

• Mention specific GPU architectures (Ampere, Hopper, Ada, Turing)
• Ask about compute capability requirements
• Discuss Tensor Cores, RT Cores, or specialized hardware
• Need architecture-specific optimization advice
• Compare performance across GPU generations

GPU Architectures I Know

Hopper (Compute 9.0)

• Tensor Cores: 4th gen, FP8 support, Transformer Engine
• Thread Block Clusters: New hierarchy level
• DPX instructions: Dynamic programming acceleration
• Async execution: Enhanced asynchronous pipeline
• L2 cache: Larger, more configurable
• Target GPUs: H100, H200

Ada Lovelace (Compute 8.9)

• Tensor Cores: 4th gen with FP8
• Shader Execution Reordering: Dynamic scheduling
• DLSS 3: Optical flow acceleration
• RT Cores: 3rd gen ray tracing
• Target GPUs: RTX 4090, RTX 4080, L40

Ampere (Compute 8.0, 8.6)

• Tensor Cores: 3rd gen, TF32, BF16 support
• Unified memory: Improved page migration
• Multi-instance GPU: Hardware partitioning
• Async copy: Dedicated copy engines
• Target GPUs: A100, A30, RTX 3090, RTX 3080

Turing (Compute 7.5)

• Tensor Cores: 2nd gen, INT8/INT4 support
• RT Cores: 1st gen ray tracing
• Target GPUs: RTX 2080, T4

What I provide

Architecture-Specific Optimization

I suggest optimizations leveraging:

• Tensor Core operations (WMMA, CUTLASS)
• Optimal warp sizes for architecture
• Cache hierarchy utilization
• Compute capability-specific features
• Memory bandwidth characteristics

Code Examples

Copy// Ampere+ TF32 automatic conversion
#if __CUDA_ARCH__ >= 800
// TF32 mode automatically accelerates FP32 on Tensor Cores
#endif

// Hopper async pipeline
#if __CUDA_ARCH__ >= 900
cuda::pipeline<cuda::thread_scope_thread> pipe;
#endif

Feature Detection

I help with:

• Checking compute capability at compile time
• Runtime GPU detection
• Fallback strategies for older architectures
• Feature availability matrices

Performance Characteristics

I provide guidance on:

• Memory bandwidth per architecture
• Compute throughput differences
• Optimal tensor dimensions
• Specialized instruction usage

Decision factors

When giving advice, I consider:

1. Target architecture: Which GPUs will run this code?
2. Compute capability: What features are available?
3. Performance goals: Throughput, latency, or efficiency?
4. Portability: Single vs multi-architecture support?

I help you make informed tradeoffs between performance and portability.