SSFFT: Energy-Efficient Selective Scaling for Fast Fourier Transform in Embedded GPUs
summary
Proposes SSFFT, a software method that uses a theoretical model to statically pick thread-block and thread counts per FFT and to choose between a regular kernel or a merged-kernel that combines two FFT steps, reducing L2 thrashing by partitioning inputs with partition size ≤25% of L2 cache. Evaluated against NVIDIA cuFFT on embedded GPUs, it reports 1.29× speedup overall (peaks 1.35× for 1K, 1.27× for 256K) and 1.26× throughput per watt by improving SM and shared-memory utilization.
AI-written summary · about
chapters (7)
- 0:00SSFFT introduction
- 1:25CUFFT static limits
- 7:51SSFFT design principles
- 10:25SSFFT memory optimizations
- 12:19Performance vs energy results
- 14:48SSFFT impact and future
- 15:30Q&A
related talks
44mSparse Fourier Transform Algorithm for Real-Time Applications
41mEnhancing Your Gaming Experience on Linux With Sched_ext
54mFast C++ by using SIMD Types with Generic Lambdas and Filters
30mHighLoad Python SIMD, GPU, and Horizontal Scaling Looking into Silicon
09mFlexAttention + FlashAttention-4: Fast and Flexible
01h 02mSqueezing the Hardware to Make Performance Juice
42mEfficient GPU Implementation of Affine Index Permutations on Arrays
08mLightning Talk: Achieving SOTA GEMM Performance: A CuTeDSL Backend for PyTorch Induc...