Skip to content

Add SM86_RTX3060_MOBILE tested config (975 MHz base clock, GA106) - #140

Open
AndrewVu23 wants to merge 1 commit into
accel-sim:devfrom
AndrewVu23:rtx3060-config
Open

Add SM86_RTX3060_MOBILE tested config (975 MHz base clock, GA106)#140
AndrewVu23 wants to merge 1 commit into
accel-sim:devfrom
AndrewVu23:rtx3060-config

Conversation

@AndrewVu23

Copy link
Copy Markdown

Summary

Adds a tuned tested-config for the RTX 3060 Laptop (GA106, SM_86): 30 SM, 12 mem partitions, 3 MB L2, core clock 975 MHz (the base clock ncu profiles at), DP 64,64

Motivation / root cause

The core clock is set to 975 MHz , which is what the card actually runs at during ncu profiling (ncu --clock-control=base default), and also derived from gpc__cycles_elapsed_avg / gpu__time_duration after running different testsuites. Not the boost like in RTX 3070.

Changes

- configs/tested-cfgs/SM86_RTX3060_MOBILE/gpgpusim.config, config_ampere_islip.icnt

Test plan

  • run_simulations.py -C RTX3060M-SASS completes on the ubench traces
  • L2 math checks: 64×128×16×2×12 = 3 MB L2 = n_sets x line x assoc x n_sub x n_mem

Limitations / known gaps

WSL can't read memory/partition counts. That's why there are some configs that needed to be patched by hand:
-gpgpu_n_mem 12
-gpgpu_n_sub_partition_per_mchannel 2 SM_86 constant
-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:P,A:192:4,32:0,32 RTX 3070 verbatim
-gpgpu_memory_partition_indexing 2 SM_86 constant
-icnt_flit_size SM_86 constant
-gpgpu_clock_domains 975:975:975:3000.5 measured + DRAM from RTX 3070

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant