3090 secret memory PerfCap

[Fuzzb3k]

Howdy folks. I've got a Gigabyte Aorus Master 3090 and I've noticed that it will throttle due to 'Thermal PerfCap' reported in HWiNFO, but the GPU itself is only 55C.

After digging a little further, it seems as though the memory chips on the rear side of the card might not be being cooled adequately. Whenever I overclock the memory it will instantly throttle the card due to thermal perfcap. It has been reported that the memory chips on the rear side can reach temps of 90C (!). It looks like there is no way to monitor the temperature of the memory chips right now, but it looks like Nvidia do have access to the temperature sensors on the chips and this is feeding into the throttling algorithm.

I'm going to try putting a heatsink/fan on the rear of the card to see if that helps, but this could explain why quite a few people have reported subpar performance with their 3090's. The 3080 doesn't have any memory chips on the rear side, so I guess it doesn't suffer from this particular problem.

Has anyone else noticed anything similar?

 

[Fuzzb3k]

Yeah I heard that EVGA have some sort of patent on monitoring the memory chips. Not sure how much truth there is to that, but seems strange that other manufacturers haven't implemented monitoring of the memory chips.

 

[Fuzzb3k]

I added a spare AMD Spire cooler heatsink that I had lying around to the top of the card and it has improved things massively. It no longer throttles due to thermals even if I overclock the memory to 700MHz. At the moment the heatsink is just sitting on top, but I might try adding some thermal paste underneath to see if I can improve the heat transfer.

Here's a thermal image of my current setup:



Even though the heatsink pipes are blue (colder), the heatsink itself feels very hot - I can't keep my finger on it for more than a couple of seconds.

 

[Fuzzb3k]

This an issue with the 3090FE too?

Can’t say I’ve noticed any problem OC’ing the VRAM so far.

Have you checked HWiNFO whilst running a GPU intensive app? You know you've run into this issue if you see 'Performance Limit - Thermal = Yes' regularly and your GPU clock speed keeps dropping down.

This is what mine looks like when it's throttled due to VRAM temperature exceeding the limit.

 

[Fuzzb3k]

I think the 3090FE is okay. Not seen the thermal issue once and for daily use memory is sat at +600 Mhz fine. ECC for mine kicks in around +900 or something but thermally is fine.

Would you mind sharing how I can test the VRAM for ECC errors?

 

[Fuzzb3k]

Sounds like an AIB oversight then TBH, which is a bit worrying really.

I might take off the backplate later and see if there's any thermal pads on the memory chips. But even if there is, there's no active cooling of the back side of the card as it's not joined in any way to the (massive) front cooler so it's really relying on dissipation from the backplate. Seems a bit rubbish considering all the effort Gigabyte put into the front cooler.

Surely it wouldn't have taken much effort to include a heatpipe or two to hook over to the back side?

 

[Fuzzb3k]

I pasted some Thermal Grizzly on the bottom of the Spire heatsink and that's done the job. No more thermal throttling. It's now only throttling due to the power limit and running a sustained memory overclock of +1300 and 2000 MHz GPU core clock.

 

[Fuzzb3k]

Hey, thanks for the link to this thread. Could you post the raw pic without the thermal mode please so I can see what you did?

I'm looking to get some thermal grizzly 3mm pads for the mem and backplate and then some 2.5 mm to replace the factory ones by core.

I had seen redpanda from redpanda mining have to do this. Didn't fancy ruining my warranty though.

Sure, here you go:



I pasted some Thermal Grizzly over the backside of the GPU and memory chips, then secured the heatsink with zip ties and pointed the fan at it. There wasn't enough clearance at the top to keep the fan mounted to the heatsink.

 

[Fuzzb3k]

That is truly savage and highly commendable

Haha, yes it was a pragmatic / hacky workaround to a problem. Being a software developer I'm used to doing those! I'm sure there are more elegant solutions requiring more time and effort, but so long as it works then I'm happy.

 

[Fuzzb3k]

Whilst I admire the ingenuity it seems a lot of money to pay for a GPU only to have to ghetto rig some extra cooling.

Very true. I'm having trouble accepting that the manufacturers didn't spot this as being a problem during the design stage. Surely they must test the cards at full utilization and monitor the effectiveness of their cooling solutions?