A History of Nvidia Stream Multiprocessor (2020)

Nvidia’s move from fixed-function GPU stages to unified stream multiprocessors began with Tesla in 2006. Earlier chips assigned separate hardware to vertex processing, fragment generation, and fragment merging, forcing designers to predict how heavily each stage would be used. The G80 replaced those units with SMs that could run different kinds of shader work, balancing demand automatically. Its SIMT model groups 32 threads into warps; performance depends on keeping warps available to hide memory latency and avoiding threads taking divergent control-flow paths. Tesla also introduced CUDA, making GPU computing more accessible. The account traces how this core structure endured even as each generation changed its scale and design.

Fermi improved per-SM throughput, floating-point accuracy, and CUDA’s support for C++ features. Kepler increased parallelism and efficiency by removing the hardware scheduler and expanding SMX units, though their scheduling became more complex. Maxwell then prioritized power efficiency, aligning core counts with warp size and simplifying scheduling; its later generation packed twice as many SMs as Kepler with only a 25 percent increase in die area. Pascal largely retained Maxwell’s SM design, relying on a 16-nanometer process and faster GDDR5X memory to raise compute and bandwidth. Across these generations, advertised peak performance rose from 345.6 gigaflops for the G80 to 11,340 for the GTX 1080 Ti. The progression reflects a recurring trade-off among parallel throughput, scheduling complexity, energy use, manufacturing limits, and feeding data to thousands of threads.