Hi everyone
what is the type of on-chip memories(register and shared memory) in GPUs (for example NVIDIA GPUs) especially ones that are used in autonomous vehicles? are they SRAM-based or GDDR-based (e.x. GDDR5)?
Best
NVIDIA does not often talk about the internal details of their GPUs, but from the available literature by NVIDIA-affiliated authors it is clear that on-chip memories are SRAM. In general, building DRAM in the logic processes used to build processors is not very efficient and therefore rarely done.
Mark Gebhart, et. al, “Unifying Primary Cache, Scratch, and Register File Memories in a Throughput Processor.” In 45th Annual IEEE/ACM International Symposium on Microarchitecture, December 2012, pp. 96-106:
However in GPUs, register files are already large, highly banked, and built out of dense SRAM arrays, not unlike typical cache and scratchpad memories.
John Nickolls, “Graphics and Computing GPUs.” In Computer Organization and Design: The Hardware/Software Interface, DA Patterson and JL Hennessy, 4th ed., Morgan Kaufman 2009, Appendix C:
Shared memory for low-latency access to data shared by cooperating threads
in the same CTA/thread block (implemented in on-chip SRAM)
1 Like
Thank you so much