DDR latency of RTX3090

I need to simulate the RTX 3090 on an emulator, mainly focusing on memory aspects. Is there a commonly recognized DDR latency value for the RTX 3090 that everyone agrees on? Thank you all for your assistance.

To my knowledge, no. It depends on what you measure and how you measure. Measured latency tends to fluctuate with load (higher load correlates with higher latency; a congested road is a useful analogy).

For GDDR6X, which is what is used for the RTX 3090, you could assume an average latency of around 250 nanoseconds. For comparison, DDR4 DRAM used for system memory would exhibit an average latency around 80 nanoseconds.

I have modeled memory subsystems in the past and my experience is that even if one models them in a fairly detailed way (caches, memory controllers, internal structure of the DRAM) and plugs in quite a number of parameters, the accuracy of the simulation of memory-intensive programs tends to be rather poor. After much frustration I was happy to be within 10% of observed application-level performance.

If you are creating a simple analytical model, simply plugging in a single latency value for the DRAM may be sufficient, but then analytical models generally exhibit worst-cases errors much larger than 10% when compared to reality.

Thank you very much for your help. I was wondering if you would be willing to share your source code that achieves measurement accuracy within 10% error. I would be immensely grateful.

Sorry, there is no code for me to share. This was (1) quite some time ago (2) in industry, so that code remains the property of the company I worked for at the time.

There are different kinds of modelling. The specific model I was working on was a proprietary trace-driven performance simulator that could be calibrated using existing (but possibly unreleased) hardware to evaluate design alternatives for next-generation hardware.

I got out of modelling work as I found it unsatisfying but am vaguely aware that trace-driven models have been (partially? largely?) replaced by execution-driven models in order to more accurately capture the behavior of modern highly-parallel processors. One promising product I helped evaluate back then was Simics by Virtutech. No idea what has become of it, as I got out of processor development altogether and went back to software.

I am very grateful for your selfless response, which means a lot to me. Thank you for taking the time.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.