gsrcrxsi

sentient12

gsrcrxsi
the reason I ask, is that when I run Einstein@home (which has a good bit of reliance on the memory subsystem), the card runs pretty low average clock speed ~1665-1725 core and ~285W when the memory is clocked at 19.5Gbps, and this gives quite good performance because Einstein really benefits from fast GPU memory/bandwidth
 
if I remove the memory OC, and run the card at stock memory speeds (compute = P2 clocks = 18.5Gbps), suddenly average core clocks increased to ~1710-1800 range, power draw increased to ~290W, but actually worse performance because of the nature of this computation.
 
ideally I'd want to be able to run the fast memory, AND high clocks and have the option to pull all the way to 350W if I wanted. but I get the feeling that pushing the memory hard is the weakest link and causes hard throttling when under load.

What memory temps are you seeing in this workload? I didn't find a strong correlation between the core clock and the memory clock/temp with synthetics. I was pretty surprised to see 80C on the memory with the core at least 10c lower. 
 
I think it's pretty clear why they added a third 8 pin power connector on the FTW3 cards.

No idea on memory temps. I run Linux and memory temp monitoring is not available here. but when running Furmark in Windows, which had a fairly high memory controller load (~60%), memory temps were about 66C.
 
I never claimed that memory TEMPERATURE was any culprit. just memory LOAD. if i'm right that the memory or memory controller is getting power from the PCIe slot only, then more load = more power draw = hitting the PCIe slot power limit earlier under heavy memory load situations, with power limiting in place, it restricts core clocks (doesn't pull back on memory at all) and reduces overall power use, even if there's headroom. that is all just an estimated guess based on observed behavior so far.

When I hit the highest point on the pcie slot power I also got the most total board power, including the pcie plugs. So your theory doesn't seem to hold true for my card.
 
This is the highest I have seen yet, but it was during the 0 memory +165 GPU run and on Fire Strike Extreme.
 
Board Power Draw [W] , PCIe Slot Power [W] , 8-Pin #1 Power [W] , 8-Pin #2 Power [W] , Power Consumption (%) [% TDP]
 
            330.0   ,                    67.8   ,                       135.4   ,                    126.7   ,                              94.3
 
This was the highest during the 0 memory +165 GPU run and on Port Royal.
 
Board Power Draw [W] , PCIe Slot Power [W] , 8-Pin #1 Power [W] , 8-Pin #2 Power [W] , Power Consumption (%) [% TDP]
 
           317.3   ,                     67.1   ,                        131.1   ,                    119.2   ,                              90.7

gsrcrxsi
I don’t think you understand what I’m suggesting.

But there’s likely more than one thing at play here. I doubt there’s the kind of single component or setting or limit “smoking gun” that you’re looking for. I just thing something related to memory is a factor.

What I was getting at is that when my pcie slot hit over 66w it didn't throttle the pcie power plugs. However that was just for a 1/10th of a second and it also was not when my GPU hit it's highest clock speed either. The highest GPU clock speed happened at 1890MHz and held for 4/10ths of a second with the total board power bouncing at 310.8, 305.8, 308.1, 310.8 and then dropped off to 1845MHz.
 
I wish there was a way to post my log file.

gsrcrxsi
If GPUz says perfcap reason is PWR. You’re being throttled by a power limit. It’s always based on the PCIe slot. 65-67 W is all the same, the GPU doesn’t have fine enough control to hold PCIe slot power right at 66W, there will be some variance and overshoot, not to mention you only have 1 sec telemetry so you’re not even seeing all the peaks and valleys that are happening. This is the nature of control systems.

It’s pretty easy to determine that PCIe slot power is the limit when if you take the PCIe power reading and divide by 66W, you get the same power use % that PX1 reports. Whereas GPUz is doing the right calculation of board power/TDP.

Your right, I was thinking the GPU-Z log was showing every 10th of a second. So if the pcie slot power is the issue, then it's a hardware issue in the way EVGA designed the board?

good info, I'm considering this card

Wow. I'm 2 minutes and 30 seconds away from the 3967 and 2 minutes and 56 seconds away from the 3953 I was thinking on buying whatever card that came first but now after reading this whole thread I might stay away from any XC3. It's really unforgivable that the XC3 underperforms a FE. I mean they cost the same and they should perform the same, the only difference is that the FE will reach the actual power limit unlike the XC3. Honestly I don't think this is an honest mistake from EVGA. Probably EVGA capped XC3 power limit to make the FTW3 to look much a better choice. The same way EVGA power limits the FTW3 compared to the Kingpin card.
 
Anyway 100+W difference on the FTW3 compared to the XC3 is just too much of a difference to ignore it. So I'll just wait for the FTW3 and ditch my XC3 opportunity when it comes.

speedysloth

tyranus7
Wow. I'm 2 minutes and 30 seconds away from the 3967 and 2 minutes and 56 seconds away from the 3953 I was thinking on buying whatever card that came first but now after reading this whole thread I might stay away from any XC3. It's really unforgivable that the XC3 underperforms a FE. I mean they cost the same and they should perform the same, the only difference is that the FE will reach the actual power limit unlike the XC3. Honestly I don't think this is an honest mistake from EVGA. Probably EVGA capped XC3 power limit to make the FTW3 to look much a better choice. The same way EVGA power limits the FTW3 compared to the Kingpin card.
 
Anyway 100+W difference on the FTW3 compared to the XC3 is just too much of a difference to ignore it. So I'll just wait for the FTW3 and ditch my XC3 opportunity when it comes.

The FTW3 is a 3-pin design, it will always use more power than the XC3. I think it might be a bit irrational to assume that EVGA did this to make people buy the FTW3 cards. They might have done this to account for some thermal or design issue with the card. Maybe they wanted to control the temperatures? Maybe they thought the card will last longer? We can't say for sure unless we get a reply from EVGA themselves and at this point I do not think they will reply to this 'problem'.

But let's forget for one minute about the power drawing. Is not acceptable that the XC3 performs worse than the FE if both cost the same. I mean the XC3 doesn't seems to be doing bad on thermals also what's the point on XC3 Hybrid to also get power capped? The logic behind water cooled cards is that you can actually overclock because the card has better temps, but with a XC3 you're just doomed and limited by the power limit and not by thermals. The XC3 has clearly design issues if the card can't perform at FE levels, and we know that the issue in this case is the fact that the card can't reach the theoretic power limit of 350W.
 
The FTW3 has it's own set of issues when talking about power. Mostly the card can't reach 450W but at least for most users it goes around 400-430W and that's plenty of power to beat the FE and justify the extra $200.

FWIW, finding out with the FTW3 and memory clock if i go past 1200mhz, it starts to artifact at high temp (~78c) at 1225mhz.
 
The performance (and power draw) also goes down after 1200mhz on my card, using unigine heaven benchmark. It does get to around 430w at the max stable overclock i found to be 120+ core and 1200+ memory.
 
Suspect its either the memory VRM only being 3 phase (where as other cards in this class like MSI Suprim X have a 4-phase memory vrm and reviewed to overclock around 1600+ mhz), possible error correction is kicking in and holding the performance back, or the memory is eating largely into the power budget and it starts to lower the available power to the GPU.
 
If you are overclocking memory along with raising the power limit, test with a program where it places near 100% load on the GPU when looking in one direction and there is little animation (to keep GPU usage persistent).
Leave core clock at stock and raise memory until you see FPS (and or power draw) actually start to decline, then overclock the core.
 
With all 3 RTX 30 series cards i've used so far (gigabyte 3060 ti gaming oc pro, evga 3070 ti ftw3 ultra, and evga 3080 ti ftw3 ultra), the power limit at most reaches around 20w below the max of the vbios so the software reporting through NVAPI is either inaccurate with the ampere architecture or that is the way this architecture is designed to account for keeping power within limits when near the ceiling

talon951

gsrcrxsi
I'm thinking that EVGA might have set the slot power too low (someone else mentioned that their 3080 slot power was set to 78 watts or so?), so it's always maxed out. I have a feeling that some parts of the memeory subsystem might be strung to the PCIe slot power, and I think the throttling behavior is a little too agressive. these are all just my guesses. only EVGA could know, and only if they dive into their own BIOS and specs to determine what's causing the undesireable behavior. If the memory is pulling power from the slot, there's probably not much they can do about that since it's hardware, but maybe there's something they can due to tune the limiting behavior. like allowing slightly more power through PCIe slot, or tuning power balancing to allow more power through the cables under certain situations, I dunno really. It's in EVGA's court to at least check if there is anything they can do.

Like I said before, if the slot power limit is actually set to 66w, then the only way to improve it is it increase that limit.  The bios doesn't control how the power is balanced between the 3 connections (8pins and slot).  The bios simply puts the brakes on when a defined limit is hit.  EVGA went to a 8 amp fuse (was 10 amp) on the slot around the time they released the new revision of the 3090 FTW3.  So it makes sense they would have a lower power limit on the slot to keep well below the fuse capability (96w).
 

So it's a hardware issue in the way EVGA designed the board and can't be fixed?

Using timespy/Witcher 3, it seems with 30~40% load on the Memory Controller, it limits the GPU Voltage to .83 to .85. I was able to get higher scores by "overclocking" at those voltages to get about 1800mhz or so avg GPU Clock. Best Advice which I don't think EVGA will fix this, even if they can is to find out what voltages are being used in your favorite games and overclock at those voltages.    

kram36

Aruzedragon
Using timespy/Witcher 3, it seems with 30~40% load on the Memory Controller, it limits the GPU Voltage to .83 to .85. I was able to get higher scores by "overclocking" at those voltages to get about 1800mhz or so avg GPU Clock. Best Advice which I don't think EVGA will fix this, even if they can is to find out what voltages are being used in your favorite games and overclock at those voltages.    

Bottom line, EVGA needs to fix this problem.

Agreed but just giving some tips on how to get the most of out of the cards. Anyways was able to score a 18,938 with a overclock 1800 at .831 and slowing scaling to 1905 to .9
 
https://www.3dmark.com/spy/21602265
 

I was able to get a look at memory controller use for my Einstein@home app, ~85% load
 
so we have:
Einstein, 85% mem controller load = 285W limit
Furmak, 60% mem controller load = 304W limit
Timespy, 30% mem controller load = 320-330W limit
DX12 RT, ~1-5% mem controller load = 350W limit
 
i see a pattern here.

Cats4Life

gsrcrxsi
I was able to get a look at memory controller use for my Einstein@home app, ~85% load
 
so we have:
Einstein, 85% mem controller load = 285W limit
Furmak, 60% mem controller load = 304W limit
Timespy, 30% mem controller load = 320-330W limit
DX12 RT, ~1-5% mem controller load = 350W limit
 
i see a pattern here.

That's for sure. My card was just delivered today. I'll likely build tomorrow or the day after when I get the chance. I'll contribute my results then.

Can't wait, more data is always helpful. One thing I am really curious on how some people can get the normal 350 watts in most usages.

Arnett

kram36
Bottom line, EVGA needs to fix this problem.

Unfortunately, I'm confident that they won't. They don't see this as a problem because it's operating exactly as EVGA designed it to: just barely performing over a stock 3080 FE. With the last couple generations, many reviewers didn't compare aftermarket SKUs against each other because "All <1080 Ti | 2080 | 1660 Super> cards are basically the same, within 1% or so. Just go with the one you like / trust." EVGA saw that as an opportunity. We got too comfortable thinking the same model GPUs were within the same performance profile. But EVGA just taught us there's more than just frequency. I was afraid that something like this would happen. That's why I kept bugging you in the other thread regarding the XC3 vs FTW3 benchmarks. >$200 difference just felt like there was something they weren't telling us.
I'm really bummed that they did this, but I'm not going to find a better card than the FTW3 for the same price, so I'll stick with EVGA for now. But regarding the step up program, is there any estimate of priority? Since the FTW3 queue has only cleared about 3 minutes over 5 weeks, it would take a small eternity if step up requests are put at the end of the line.
 
 

The 3080 Ti XC3 isn't performing as good as the 3080 Ti FE card. Same price for less performance? Sounds like a good way to loose customers and damage your Company's reputation.

EVGATech_LeeM has time to address post in this thread, but can't be bothered to address this thread.
 

EVGATech_LeeM
The header is only to have it sync with PX1.  You would not connect the card header to your motherboard.  Nothing would be gained that way.

If you want to control your fan RGB/speed away from the card, then you'll need to plug those connectors into your motherboard or different devices.

 
https://forums.evga.com/FindPost/3431031
 
Come on EVGATech_LeeM, time to address this thread.

blizz93
I have been reading a bit of this thread and I see that some people mention that the performance of the 3080 ti xc3 is worse than a FE but at least the results that I see from benchmarks of my gpu is that the performance is equal or a little better than a FE and I'm also seeing the same power limit that many say around 344w and 350w this is the last result I got on time spy just for comparison.


https://www.3dmark.com/spy/21633989

Very few hit anywhere near 340W, most of the time it's averaging around 300W during the Time Spy run. You have an anomaly in the 3080 Ti XC3 cards, not the norm.