GPU Memory Bandwidth Bottlenecks

The cluster discusses how GPU compute performance is primarily limited by memory bandwidth rather than raw processing power, highlighting issues like slow PCIe transfers from system RAM to VRAM, unified memory trade-offs, and potential direct disk-to-GPU copying solutions.

➡️ Stable 0.6x Hardware
3,464
Comments
19
Years Active
5
Top Authors
#9675
Topic ID

Activity Over Time

2008
4
2009
16
2010
18
2011
32
2012
54
2013
62
2014
66
2015
95
2016
190
2017
189
2018
126
2019
162
2020
330
2021
280
2022
284
2023
535
2024
508
2025
483
2026
30

Keywords

RAM BAR CPU SSD DDR4 OK M1 engine.md MUCH SIMD gpu memory bandwidth ram cpu bottleneck gpus data vram pcie

Sample Comments

sva_ Nov 13, 2023 View on HN

It is remarkable how much GPU compute is limited by memory speed.

rp1 Dec 19, 2021 View on HN

True, bandwidth is a concern. It might be possible to copy memory directly from disk to the GPU, which may be more efficient.

aappleby Aug 25, 2024 View on HN

You'll be limited by memory bandwidth more than compute.

bronxbomber92 Sep 13, 2020 View on HN

How much would the trade offs change if GPUs shared the same main memory as CPU?

BoredPositron Aug 30, 2025 View on HN

Working with the vram of both gpus on the same card is aeons faster than taking a system ram roundtrip.

jakelarkin Oct 6, 2018 View on HN

probably the processor-memory bus. GPUs have wider ones

buildbot Nov 3, 2023 View on HN

Your PC can't use the GPU memory bandwidth for the CPU whatsoever. So why would you add the bandwidth?

p1esk Nov 29, 2019 View on HN

But you don't want to do that. The bandwidth between GPU and system RAM is 10x less than the bandwidth between GPU and GPU RAM, so the GPU would spend > 90% of time waiting for data. In this scenario, using CPU would most likely be faster.

z3phyr Mar 14, 2013 View on HN

Though, GPUs are meant for highly specialized number-crunching, they doesn't have direct access to system memory. One also gets additional latency transferring data between RAM and GPU memory.Also, GPU is a temporary technology. Wait for some time, and see CPUs getting more GPU like.

bsenftner Aug 26, 2025 View on HN

It's not just the GPU memory, it's also I/O memory. That speeds up a lot: just update the pointer to where the memory is, no copying out of I/O memory.