The Promise of Heterogeneous Computing
summary
Software grows more abstract while hardware fragments into GPUs, TPUs, FPGAs, DSPs and memory‑adjacent chips, creating a gap; Google ships a TPU for TensorFlow and Nvidia builds GPUs with ~150 billion transistors, PCIe still limits bandwidth (~60 GB/s), and many chips remain hard to access. Bitfusion builds software to pool different kinds of compute across servers into a virtual mega server, letting developers run remote accelerators from a laptop so specialized hardware is actually usable.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:00The fundamental problem in computing
- 3:50The era of frequency
- 6:15What happened to developers
- 7:30Integrated GPUs
- 9:00Discrete GPUs
- 11:50FPGAs
- 14:50Applicationspecific processors
- 15:30Ecosystem maturity
- 16:05Quantum computers
related talks
26mKeynote: Artificial Innovation Accelerates Artificial Intelligence
39mDensity! The not-so-secret trend driving the future of Data Analytics
58mThe Next Decade of Software Development
09mAccelerating compute with software defined hardware (FPGA's)
52mComputing Performance 2022: What's on the Horizon
13mDecentralized ZettaFlop Uber GPU/XPU Network for Al and Spatial Computing
21mAnalyst Notebook
03mAccelerate with purpose