A history of NVIDIA Streaming Multiprocessor (part 2)
•x min read
gpu-computinghardwarenvidiaarchitecture
A few weeks ago, I stumbled upon this awesome blog post by Fabien Sanglard. It goes into great detail on the evolution of NVIDIA's streaming multiprocessors and the design choices behind the different architectures throughout the years.
The article was published on May 2, 2020, when the last released architecture was Turing (2018), and Ampere was still in the realm of speculation. Since then, NVIDIA has released four different architectures, LLMs exploded in popularity, and the demand for massively parallel processors has risen to an all-time high.
In this blog post, we are going to visit the new SM architectures and try to answer a few thoughts Fabien left by the end of his article:
What will be interesting is to see how Nvidia keeps on evolving now that their dies have three types of cores serving different purposes. Will we see dies entirely made of Tensor cores or RT cores? I am curious to find out.
Visited in this blog post:
| Year | Arch | Series | Die | Process | Consumer GPU | Data Center GPU |
|---|---|---|---|---|---|---|
| 2020 | Ampere | |||||
| 2022 | Ada Lovelace | |||||
| 2022 | Hopper | |||||
| 2024 | Blackell |