GPU Specs GPU specs define hardware characteristics used for roofline analysis, performance modeling, and comparison. Wafer includes built-in specs for common GPUs and supports custom configurations.
Quick Start # List available specs wafer config specs list # Show spec details wafer config specs show H100 # Set default spec wafer config specs default H100
Commands wafer config specs list List all available GPU specifications:
Output:
GPU Specifications: NVIDIA: Name Memory BW (TB/s) FP16 (TFLOPS) TDP H100 80GB 3.35 989.4 700W H200 141GB 4.80 989.4 700W A100 80GB 2.04 312.0 400W B200 192GB 8.00 2250.0 1000W RTX4090 24GB 1.01 165.2 450W AMD: Name Memory BW (TB/s) FP16 (TFLOPS) TDP MI300X 192GB 5.30 1307.4 750W MI250X 128GB 3.28 383.0 560W MI210 64GB 1.64 181.0 300W * = default
wafer config specs show Show detailed specifications for a GPU:
wafer config specs show < gpu-nam e >
Example: wafer config specs show H100
Output:
NVIDIA H100 SXM5 (80GB) Memory: Capacity: 80 GB HBM3 Bandwidth: 3.35 TB/s Bus Width: 5120-bit Compute: FP64: 33.5 TFLOPS FP32: 66.9 TFLOPS FP16: 989.4 TFLOPS (Tensor Core) BF16: 989.4 TFLOPS (Tensor Core) INT8: 1978.9 TOPS (Tensor Core) FP8: 1978.9 TFLOPS (Tensor Core) Architecture: SMs: 132 CUDA Cores: 16896 Tensor Cores: 528 (4th gen) L2 Cache: 50 MB Registers/SM: 65536 Roofline: Ridge Point (FP16): 295.3 FLOP/byte Ridge Point (FP32): 20.0 FLOP/byte Power: TDP: 700W
wafer config specs add Add a custom GPU specification:
wafer config specs add < nam e > [OPTIONS]
Options: Option Description --memoryMemory capacity (e.g., “80GB”) --bandwidthMemory bandwidth (e.g., “3.35TB/s”) --fp16FP16 peak TFLOPS --fp32FP32 peak TFLOPS --tdpThermal design power (watts)
Example: wafer config specs add my-custom-gpu \ --memory 48GB \ --bandwidth 2.0TB/s \ --fp16 400 \ --fp32 200 \ --tdp 350
wafer config specs remove Remove a custom specification:
wafer config specs remove < nam e >
Built-in GPU specs cannot be removed, only custom specs.
wafer config specs default Set the default GPU for analysis:
wafer config specs default < nam e >
The default is used when no --gpu flag is specified:
# Set default wafer config specs default H100 # Now roofline uses H100 automatically wafer roofline --bytes 1e9 --flops 1e12 --time-ms 0.5
Using Specs Roofline Analysis # Use specific GPU wafer roofline --gpu H100 --bytes 1e9 --flops 1e12 --time-ms 0.5 # Use default GPU wafer roofline --bytes 1e9 --flops 1e12 --time-ms 0.5
Baseline Discovery wafer baseline run "torch.matmul(A, B)" \ --shape A=1024,1024 \ --shape B=1024,1024 \ --hardware H100
Compare same workload across GPUs:
# Analyze for H100 wafer roofline --gpu H100 --bytes 2e9 --flops 4e12 --time-ms 1.0 # Analyze for A100 wafer roofline --gpu A100 --bytes 2e9 --flops 4e12 --time-ms 1.5
Built-in Specifications NVIDIA GPUs GPU Generation Memory Peak FP16 B200 Blackwell 192GB HBM3e 2250 TFLOPS H200 Hopper 141GB HBM3e 989 TFLOPS H100 Hopper 80GB HBM3 989 TFLOPS A100 Ampere 80GB HBM2e 312 TFLOPS RTX 4090 Ada 24GB GDDR6X 165 TFLOPS
AMD GPUs GPU Generation Memory Peak FP16 MI300X CDNA 3 192GB HBM3 1307 TFLOPS MI250X CDNA 2 128GB HBM2e 383 TFLOPS MI210 CDNA 2 64GB HBM2e 181 TFLOPS
Custom Spec File Create specs from a YAML file:
# my-gpu.yaml name : Custom GPU vendor : NVIDIA memory : capacity_gb : 80 bandwidth_tb_s : 3.0 compute : fp16_tflops : 500 fp32_tflops : 250 fp64_tflops : 125 power : tdp_watts : 400
wafer config specs add --from-file my-gpu.yaml
Next Steps
Roofline Analysis Use specs for roofline analysis.
Baseline Discovery Use specs with baseline.
Targets Configure GPU targets.
Workspaces Access cloud GPUs.
