DBTB INT Subbu Rama r
summary
Describes a working software to virtualize GPUs that pools heterogeneous accelerators in data centers and claims measurable gains: higher application throughput or lower infrastructure cost by sharing GPUs and other processors. Argues data growth makes compute demand explode and recommends choosing FPGAs and pattern‑matching or specialized deep‑learning chips for suitable jobs rather than defaulting to CPUs or GPUs. Emphasizes practical experience, patience, and that the virtualization approach is already deployed.
AI-written summary · about
related talks
19mML training acceleration with heterogeneous resources in Bytedance
09mAccelerating compute with software defined hardware (FPGA's)
39mDensity! The not-so-secret trend driving the future of Data Analytics
50mThe democratization of tomorrow’s technology
09mDemocratizing Access to GPUs for Faster MLAI Innovation with Generative AI
21mAnalyst Notebook
24mManipulating Data in an Artificial Intelligence World
24mTiresias: A GPU Cluster Manager for Distributed Deep Learning