Different GPUs at same power limit (FPS/watt)

[BFG10K]

Reserved in case I decide to run my own tests.

 

[BFG10K]

I'm curious if anyone has seen or run their own tests with multiple GPUs at the same power limit. Like a 3050 against a 3060 and 3080 all locked to the 3050's default 130W power limit.

Or even better, something like a 2070 against 3xxx cards locked to 175W, to gauge multiple generations of FPS/watt.

The only thing I've found is this.

He tests the 3060TI and 3080TI against the 3060 @ 3060 default power limit of 170W. 3060Ti comes out slightly ahead with the 3080TI matching the 3060TI.

 

[MrTeal]

That's not really IPC. IPC is pretty specifically instructions (or in the case of a GPU maybe FPS) per clock, not per watt. To compare them, youd have to fix the clock rate. IPC is a pretty useless comparison in that regard for the same arch though, since obviously 3080 Ti > 3060 Ti > 3060.

Efficiency will generally be higher with lower clock rates and more shaders, but you'll have cases like the 3080 Ti at a low power limit where it has to pay the power cost of 384 bit GDDR6x even if the power limit doesn't allow it to effectively use it.

 

[biostud]

It is a bit more complicated when you don't have full control of vcore. I did these test on my 6800XT. For me it would be interesting to see how a regular 6800 would perform in the same power envelopes. (Borderlands 3 Ultra @ 5120x1440)

@200W avg. FPS 66.52 (actual GPU voltage ~950mV, max frequency 2300Mhz, Voltage 1050 mV)
@240W + 20% avg FPS 69.51 +4,5% (actual GPU voltage ~1015mV, max frequency 2400Mhz, Voltage 1050 mV)
@280W + 40% avg. FPS 71.25 + 7,1% (actual GPU voltage ~1150mV, max frequency 2500Mhz, Voltage 1050 mV)
@310W + 55% avg. FPS 72.54 + 9,0% (actual GPU voltage ~1150mV, max frequency 2600Mhz, Voltage 1150 mV)

 

[NTMBK]

That's not Instructions Per Clock.

(Of course given every GPU generation has a different instruction set, trying to compare IPC is meaningless.)

 

[coercitiv]

He tests the 3060TI and 3080TI against the 3060, all locked to the 3060's default power limit of 170W. 3060Ti comes out slightly ahead with the 3080TI matching the 3060TI.

I'm surprised the 3080 Ti is performing so poorly at 170W, I would expect much better performance.

One of my concerns when getting a 250-300W card such as the 6800XT was scalability, in the sense that I usually run my cards bellow spec in order to get a better noise profile. To give you a ballpark idea, I used to run my Vega 56 at 150W or lower through a combo of underclocking and frame limits, resulting in GPU fan speeds aorund 1000 RPM. As I upgraded to the 6800XT, I was wondering if that option was still available to me, in the sense of still being able to run games at <150W even if the cooling on the new card is obviously more capable and quieter (better fans, overall slower fans etc).

The first game I loaded on the new card, with the settings from the old card, was Division 2. To my complete surprise the software power readings for the 6800XT were around 50-70W. IIRC it was less than half the power consumption of the Vega 56 with equal performance (same FPS cap). We know from reviews the 6800XT is around 100% faster than Vega56, but it's still a nice surprise to see the card scaling so well in low power scenarios.

 

[ZGR]

I think the 3080 Ti is performing poorly because capping the power limit alone is the wrong way to tune RTX 3000.

Using a hand tuned voltage curve and capping the voltage to 0.713 V will make any RTX 3000 card sip power below 1.6 GHz.

Lowering power limit alone is just gonna get the GPU to throttle while still drawing quite a bit of power. Further lowering clocks manually until within TDP limit is going to have much better performance.

The stock voltage curve is abysmal on every RTX 3000 card I’ve tuned. A 1.5 GHz 3080 locked at 0.713 V is about a 150 W card without RT.

 

[BFG10K]

I'm surprised the 3080 Ti is performing so poorly at 170W, I would expect much better performance.

Yep, that's basically why I started this thread. I had assumed more execution units at lower clocks would be more efficient, but that's obviously not always the case.

In my situation with a 2070, even with a 60FPS cap it gets very noisy in more demanding games like Crysis Warhead, Tomb Raider 2013 and Resident Evil 2/3. Also the heat belched into the room @ 175W is significant.

I settled on a 130W cap; the fans are several hundred RPM lower and the amount of heat dumped out is far less. And I still meet my 60FPS target at the same settings.

But then I started thinking, would a 3050 perform the same...

 

[BFG10K]

Using a hand tuned voltage curve and capping the voltage to 0.713 V will make any RTX 3000 card sip power below 1.6 GHz.

The problem with undervolting is the same as with overclocking. You're taking a risk with stability because you never know which situation around the corner will cause a game crash. And when it does randomly crash, you won't know if it's your voltage or something else.

A lowered powercap is 100% stable and doesn't need testing.

 

[ZGR]

The problem with undervolting is the same as with overclocking. You're taking a risk with stability because you never know which situation around the corner will cause a game crash. And when it does randomly crash, you won't know if it's your voltage or something else.

A lowered powercap is 100% stable and doesn't need testing.

I am pretty sure most if not all RTX 3000 cards can do 1.45-1.50 GHz at 0.713 V stable no problem. But MSI Afterburner can do a few hundred MHz below that in case it isn’t.

I can’t imagine a dud 1200 MHz 3080 Ti at 0.713 V won’t be stable, and that will outperform the power limited 3080 Ti in the video above while barely consuming more power.

But you are not wrong. It can take weeks to actually verify an undervolt is stable. Lots of patience required to dial it in.

 

[DAPUNISHER]

I haven't done anything scientifically controlled. I tested 2 RDNA2 cards in Fallout 4 1080p max settings 60fps cap in place, Hi Res textures installed. RX 6400 stock v. RX 6600XT with Chill turned on. The 6600XT used less wattage than the 6400. It wasn't a big difference obviously since my XFX 6400 never pushes 50W in games I've tried on it. Might be noteworthy that the 6600XT also had better frame pacing in the most demanding situations, and held 60 where the 6400 could not. I will throw my 6800 in the mix later today or tomorrow. I can't see how it beats the 6600XT, which was using 20-30W 99.9% of the time. With low to mid 20s mostly.

 

[IEC]

It is a bit more complicated when you don't have full control of vcore. I did these test on my 6800XT. For me it would be interesting to see how a regular 6800 would perform in the same power envelopes. (Borderlands 3 Ultra @ 5120x1440)

@200W avg. FPS 66.52 (actual GPU voltage ~950mV, max frequency 2300Mhz, Voltage 1050 mV)
@240W + 20% avg FPS 69.51 +4,5% (actual GPU voltage ~1015mV, max frequency 2400Mhz, Voltage 1050 mV)
@280W + 40% avg. FPS 71.25 + 7,1% (actual GPU voltage ~1150mV, max frequency 2500Mhz, Voltage 1050 mV)
@310W + 55% avg. FPS 72.54 + 9,0% (actual GPU voltage ~1150mV, max frequency 2600Mhz, Voltage 1150 mV)

Similar to the sweet spot for my 6800XT on stock voltage curve. Literally just set max frequency slider in Adrenaline to 90% and it does 160-170W @ <949mV, frequency 2100-2200MHz with minimal performance loss and less than 60% of the power draw. I do not undervolt as I have zero tolerance for instability.

There is card to card variation but I would suspect a 6800 would take the crown for efficiency setups over the 6800XT. Techpowerup review of 6800 perf/W
At least at 1440p. At 1080p my 6600XT Red Devil card with a 90% frequency cap was routinely using 60W while gaming.

Both AMD and nVidia's latest generation of cards can be more efficient (esp a power limited 4090) - but you'll have to keep in mind these are early numbers on early drivers which clearly have a lot of room for improvement (looking at you, AMD).


I can cap my max frequency for my 7900XT at 90% with maybe 5% performance loss and cut power usage down to <260W, which would match the stock 4090 performance for efficiency. This gives me the performance I want at 1440p 240Hz while capping power and thermals at very reasonable levels (card does not exceed 60C hotspot and stays quiet). If I set a FPS cap in-game, this would further increase efficiency by demanding a lot less of the GPU.



If I had a specific application that favors nV or needed high-refresh rate 4K gaming, I would favor the RTX 4090 as the card to get for efficiency in that scenario.

At 1440p even the 6800/XT provides plenty of performance at a much better value given recent sales.

 

[biostud]

Similar to the sweet spot for my 6800XT on stock voltage curve. Literally just set max frequency slider in Adrenaline to 90% and it does 160-170W @ <949mV, frequency 2100-2200MHz with minimal performance loss and less than 60% of the power draw. I do not undervolt as I have zero tolerance for instability.

There is card to card variation but I would suspect a 6800 would take the crown for efficiency setups over the 6800XT. Techpowerup review of 6800 perf/W
At least at 1440p. At 1080p my 6600XT Red Devil card with a 90% frequency cap was routinely using 60W while gaming.

Both AMD and nVidia's latest generation of cards can be more efficient (esp a power limited 4090) - but you'll have to keep in mind these are early numbers on early drivers which clearly have a lot of room for improvement (looking at you, AMD).
View attachment 75531

I can cap my max frequency for my 7900XT at 90% with maybe 5% performance loss and cut power usage down to <260W, which would match the stock 4090 performance for efficiency. This gives me the performance I want at 1440p 240Hz while capping power and thermals at very reasonable levels (card does not exceed 60C hotspot and stays quiet). If I set a FPS cap in-game, this would further increase efficiency by demanding a lot less of the GPU.

View attachment 75534

If I had a specific application that favors nV or needed high-refresh rate 4K gaming, I would favor the RTX 4090 as the card to get for efficiency in that scenario.

At 1440p even the 6800/XT provides plenty of performance at a much better value given recent sales.

Have you increased memory settings on your X6800XT? (fast timings and frequency?)

I'll try stock settings and frequency @ 90% and see what that brings me.

 

[IEC]

Have you increased memory settings on your X6800XT? (fast timings and frequency?)

I'll try stock settings and frequency @ 90% and see what that brings me.

No, default timings for 6800XT.

I'm using fast timings on the 7900XT due to idle power usage - enforcing fast timings vs default decreases idle power consumption on high-refresh rate multi-monitor setups by an additional 20-30% (bug?).

240Hz primary 1440p + 60Hz secondary 1440p = 31W power draw at idle. For comparison, my 6800XT draws 6W at idle but it's connected only to a 4K TV.

 

[biostud]

No, default timings for 6800XT.

I'm using fast timings on the 7900XT due to idle power usage - enforcing fast timings vs default decreases idle power consumption on high-refresh rate multi-monitor setups by an additional 20-30% (bug?).

240Hz primary 1440p + 60Hz secondary 1440p = 31W power draw at idle. For comparison, my 6800XT draws 6W at idle but it's connected only to a 4K TV.

Any reason you don't run fast timings and faster clock?

 

[IEC]

Any reason you don't run fast timings and faster clock?

No reason other than I don't need the extra memory performance.

 

[MrTeal]

Yep, that's basically why I started this thread. I had assumed more execution units at lower clocks would be more efficient, but that's obviously not always the case.

170W is a really big cut for the 3080 Ti; it's less than half its stock TDP. Even at the lower total TDP you're going to have fixed power costs for running that 384bit bus and 12 GDDR6X memory chips, and it's probably decently more than the 3060Ti and 3060. IIRC it was ~40W just for the memory ICs, so a big chunk.
Quick test with my FTW3 3080 10GB power limited to 45% (170W) running Furmark gives me the following in HWInfo:
GPU Power: 170W
GPU Core (NVVDD) Input Power: 57.6W (34%)
GPU SRAM Input Power: 35.7W (21%)
GPU FBVDD Input Power: 64.3W (37.8%)

At stock, it's
GPU Power: 377W
GPU Core (NVVDD) Input Power: 206W (54.6%)
GPU SRAM Input Power: 69.6W (18.5%)
GPU FBVDD Input Power: 75.8W (20.1%)

Frame buffer power is taking up 38% of the power budget at 170W on my card, and it would probably be more for the 12GB 3080/3080Ti. More shaders clocked lower for the same TFLOPS is generally more efficient, but a 3060 with a 192bit bus and 6 lower power regular GDDR6 ICs probably has 50-100% more power available to the core than a hugely downclocked 3080Ti.

 

[biostud]

Personally I think it is more interesting to find the optimal performance /watt for the each GPU, as there are so much variation between the cards.

Then do a comparison of the different cards @ optimal performance /watt.

 

[ZGR]

Quick test with my FTW3 3080 10GB power limited to 45% (170W) running Furmark

What is your core voltage during this test? I'm doing doing Furmark 3840x2160 with 8xMSAA.

At stock, my 3080 boosts up to 1.055 V which is just absurd. At 45% power limit in Furmark, it is around 0.712 to 0.755 V at 1180-1350 MHz and drawing around 140-144W via GPU Power sensor. Getting around 35 FPS with inconsistent performance.

With full power limit untouched but with locked voltage and clocks to 0.717 V and 1530 MHz respectively, the 3080 is drawing 148-153W at 1530 MHz and getting a constant 43 FPS. This is a large perf/w gain. I do feel like capping power limit alone is gimping higher TDP RTX 3000 cards unnecessarily.

I can't imagine any RTX 30XX not being able to do 1450 MHz at locked 0.717 V all day long. We can go as low as 800 MHz with this method in case someone has a serious dud.

The best part about locking clocks and voltage is we can do this on any modern GPU with MSI Afterburner. My friend's 7900 XTX profile looks similar but his clocks are much higher; automatically applying on Windows login.

 

[CP5670]

The 4090 can be limited to 350W instead of 450W with only 5-10% performance drop in most games, but anything RT falls off more. RT uses much more power than raster games in general.