Hi all, I'm trying to run a highly scaled CPU configuration and finding that individual Gem5 instances are using up to 250GB of memory with my new configuration!! I'm wondering if anybody knows the most impactful parameters for memory usage so I can decide what to try and tone down while still testing what I want. My configuration is: Pipeline width: 12 Instruction queue entries: 512 ROB entries: 2048 LSQ entries: 192 SSIT/LFST entries: 512 L1i and d cache: 512KiB L2 cache: 32MB I have a feeling it's either the L2 cache or ROB entries, but is it possible there's some synergistic behaviour at play too? My previous configuration still had 576 ROB entries and a L2 cache of 4MB but didn't consume anywhere near the same amount of memory, so it seems odd to be entirely down to this. Sent with [Proton Mail](https://proton.me/) secure email.

Hi, just sharing some info I've found that look like a possible Gem5 bug: Below around 1024 physical registers (for all three data types) memory usage is at normal levels (around half a gb). At 1024 and beyond, memory usage reaches incredibly high levels that even my machine with 500GB of memory can't support. The process ends up either OOM killed or crashes with a bad malloc error. I hope this info ends up useful to someone! Sent from Proton Mail mobile -------- Original Message -------- On 19 Jan 2024, 01:25, muke101 wrote: > Hi all, I'm trying to run a highly scaled CPU configuration and finding that individual Gem5 instances are using up to 250GB of memory with my new configuration!! I'm wondering if anybody knows the most impactful parameters for memory usage so I can decide what to try and tone down while still testing what I want. My configuration is: > > Pipeline width: 12 > Instruction queue entries: 512 > ROB entries: 2048 > LSQ entries: 192 > SSIT/LFST entries: 512 > L1i and d cache: 512KiB > L2 cache: 32MB > > I have a feeling it's either the L2 cache or ROB entries, but is it possible there's some synergistic behaviour at play too? My previous configuration still had 576 ROB entries and a L2 cache of 4MB but didn't consume anywhere near the same amount of memory, so it seems odd to be entirely down to this. > > Sent with [Proton Mail](https://proton.me/) secure email.