The Future of Fast Code: Giving Hardware What It Wants
summary
Hardware offers huge gains but only with heavy specialization: simple Python matrix multiply hits ~10 MFLOPS, JavaScript ~500 MFLOPS, optimized C reaches ~1–40 GFLOPS with loop reordering and cache blocking and a tuned multicore build hits ~200 GFLOPS, a device neural engine reaches 16 TFLOPS on M2 neural engine after quantization, and H100-class GPUs approach petaflops with custom kernels. Compiler-and-DSL work (Halide/Taichi), EXO user-extensible intrinsics, and tools like Pitchfork (offline synthesis for DSP patterns) let teams target tensor cores and matrix-block instrinsics efficiently; automated synthesis plus differentiable program search produces image algorithms an order of magnitude faster or higher quality.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:15Hardware performance leap
- 8:39Clean code vs hardware
- 13:23Programming models lag
- 20:08Throughput design principles
- 28:35Energy costs and data movement
- 36:37Complex instructions and arithmetic
- 42:18Safety in high-performance systems
- 49:58Correctness and equivalence
- 51:13Q&A: Software-hardware co-design & FPGAs
- 53:19Q&A: HDL and spatial architectures
- 56:11Q&A: VLIW architectures?
- 57:59Q&A: Safety vs optimizations
- 1:00:07Q&A: Cache coherence overheads
- 1:01:54Q&A: Von Neumann bottleneck
- 1:03:17Closing remarks
related talks
44mFastware
29mReal Programming
01h 00mFastware
56mWrite Fast Code Like a Native
01h 50mUnlocking Modern CPU Power - Next-Gen C++ Optimization Techniques
01h 00mProgramming with less effort in C++...
01h 19mCo-Design for High-Performance Computing Software Systems
20mHigh-Performance Computing that’s Earth Friendly