Folding Forum

Good evening guys and gals. I've been trying to figure it out on my own, but can't seem to crack it.
I had an offer to trade my 3060 for an 2070 blower by zotac + some money, but I can't figure if I'm gonna increase or decrease my PPD, due to the multitude of gpu chip variations.

According to FOLDING.LAR.SYSTEMS:
1)TU106 [GeForce RTX 2070] M 6497 - GPU PPD AVERAGE: ALL OS 1,597,615 [https://folding.lar.systems/gpu_ppd/bra ... 070_m_6497]
2)TU106 [GeForce RTX 2070 Rev. A] M 7465 - GPU PPD AVERAGE: ALL OS 2,397,883 [https://folding.lar.systems/gpu_ppd/bra ... v_a_m_7465]

The ZOTAC 2070 Blower uses a TU106-400-A1 chip, which seems to be option (1)?
Even so, FP64 performance should be better than that of the 3060? Something doesn't add up, I'm pulling 1,57~1,6M on a 42% PWR Limited 3060.

RTX 3060 RTX 2070
FP64 (double) performance FP64 (double) performance
199.0 GFLOPS (1:64) 233.3 GFLOPS (1:32)

https://www.techpowerup.com/gpu-specs/g ... 3060.c3682
https://www.techpowerup.com/gpu-specs/z ... ower.b6356

Am I missing something? Again...

*https://www.gamersnexus.net/hwreviews/3 ... tra-review
*https://www.youtube.com/watch?v=eZJcXHOn42Y

3060 is much faster than a 2070.

A GPU must do double precision Floating Point math. (FP64) However, F@H tries to minimize the use of FP64 as it is MUCH slower then FP32. (it only needs to be twice as slow, but the vendors often slow them to 1/32 or 1/64 the speed of FP32. The fastest I have seen is 1/8 as fast)

If it was me, I would use the speed of single precision Floating Point math (FP32) to compare.

Looking only at FP32 a 2080 should be falling behind the 3060. That is not the case however.

JimboPalmer wrote:If it was me, I would use the speed of single precision Floating Point math (FP32) to compare.

Yes, for the exact reason he gave. The bulk of the work is FP32. (The programmers were pretty effective at minimizing the need for FP64 but it's >0.)

... but even rated FP32 GFLOPS can be deceptive. Rated speed is proportional to the number of Shader*ClockRate which works out pretty well for large proteins where all the shaders can be kept busy for most of the time. There's always some serial calculations and for small proteins, you can't keep them all busy.

bruce wrote:

JimboPalmer wrote:If it was me, I would use the speed of single precision Floating Point math (FP32) to compare.

Yes, for the exact reason he gave. The bulk of the work is FP32. (The programmers were pretty effective at minimizing the need for FP64 but it's >0.)

... but even rated FP32 GFLOPS can be deceptive. Rated speed is proportional to the number of Shader*ClockRate which works out pretty well for large proteins where all the shaders can be kept busy for most of the time. There's always some serial calculations and for small proteins, you can't keep them all busy.

As I said above, that doesn't hold much water either. The 3060's 12.74 TFLOPS underperform the 2080's 10.07 TFLOPS, in terms of PPD. A 25% increase in FP32 still leaves the 3060 almost a million behind in PPD.

iero wrote:

bruce wrote:

JimboPalmer wrote:If it was me, I would use the speed of single precision Floating Point math (FP32) to compare.

Yes, for the exact reason he gave. The bulk of the work is FP32. (The programmers were pretty effective at minimizing the need for FP64 but it's >0.)

... but even rated FP32 GFLOPS can be deceptive. Rated speed is proportional to the number of Shader*ClockRate which works out pretty well for large proteins where all the shaders can be kept busy for most of the time. There's always some serial calculations and for small proteins, you can't keep them all busy.

As I said above, that doesn't hold much water either. The 3060's 12.74 TFLOPS underperform the 2080's 10.07 TFLOPS, in terms of PPD. A 25% increase in FP32 still leaves the 3060 almost a million behind in PPD.

Still waiting for someone to enlighten me.

iero wrote:
Still waiting for someone to enlighten me.

The model number, in your case M 6497 vs. M 7465 is displayed in the FAHControl window. In the case the 2070 is the slower one, your PPD wil not decrease that much (1,4M vs. 1.7M), but if it is the faster Rev. A, then you get 2,3M PPD. I know, it is a bet, but you wrote that you can get some money extra...On Ebay, I recently got a GTX1060 3GB for 100€, which may easily close the gap of 0,3M PPD. I think, your "business" partner offers you that additional price.

Jupis_folder wrote:The model number, in your case M 6497 vs. M 7465 is displayed in the FAHControl window. In the case the 2070 is the slower one, your PPD wil not decrease that much (1,4M vs. 1.7M), but if it is the faster Rev. A, then you get 2,3M PPD. I know, it is a bet, but you wrote that you can get some money extra...On Ebay, I recently got a GTX1060 3GB for 100€, which may easily close the gap of 0,3M PPD. I think, your "business" partner offers you that additional price.

Thanks for the info! I was insinuating more about the most accurate way to relatively gaze the performance based to FP32 or something else.

iero wrote:

Jupis_folder wrote:The model number, in your case M 6497 vs. M 7465 is displayed in the FAHControl window. In the case the 2070 is the slower one, your PPD wil not decrease that much (1,4M vs. 1.7M), but if it is the faster Rev. A, then you get 2,3M PPD. I know, it is a bet, but you wrote that you can get some money extra...On Ebay, I recently got a GTX1060 3GB for 100€, which may easily close the gap of 0,3M PPD. I think, your "business" partner offers you that additional price.

Thanks for the info! I was insinuating more about the most accurate way to relatively gaze the performance based to FP32 or something else.

You can't look at the numbers posted online.
The 3060 numbers are hugely inflated that they don't make any sense, and yet, aftermarket numbers show a much higher performance thanks to better coolers, than stock Nvidia GPUs.
Then again, the reference design is the best design, if you just tie 1 case fan on the center. Then reference blows all other coolers out of the water.
The numbers online are also supposed to be stock speeds and stock numbers.
If you ramp up the fans, in an open case design, temps drop and performance increases.
The RTX 2000 series boost speeds benefitted greatly from cooler temperatures.
Apparently, the 3000 series GPUs benefit more from added power. (maybe not all of them).
Their temperatures remain mostly under 65C, which is hard to do on a 2000 GPU of the same wattage.
So while 2000 series GPUs had quite the overclock window especially when playing around with REDUCING wattage (rather than increasing it), at least some 3000 series actually require you to INCREASE the wattage, which in turn lowers the efficiency.
It's been a year, and I haven't found the time or energy to research 3000 GPU performance yet.
But I presume that a lot of the numbers that don't make sense, are because a 2070/2080 that's been fully optimized for folding, is being compared to a 3060 running stock.

MeeLee wrote: Then again, the reference design is the best design, if you just tie 1 case fan on the center. Then reference blows all other coolers out of the water.

Where are you getting this from? Are you talking about the reference PCB, cooler or both? Cause there is no way that either is true.

Either way, I was thinking along the lines that there has to be a number/metric or a formula of metrics that can reliably categorize Nvidia cards relatively to each other.

iero wrote:

MeeLee wrote: Then again, the reference design is the best design, if you just tie 1 case fan on the center. Then reference blows all other coolers out of the water.

Where are you getting this from? Are you talking about the reference PCB, cooler or both? Cause there is no way that either is true.

Either way, I was thinking along the lines that there has to be a number/metric or a formula of metrics that can reliably categorize Nvidia cards relatively to each other.

For most people's use case scenario the Nvidia Reference design would be superior in cooling, that is because it actually routs the heat away from the GPU, rather than with older (RTX 2000 - like) designs, where it'll just suck up it's exhaust air.
Yes, the Nvidia reference cooler is superior, but not as is.
Check videos of Jayztwocents, where he mounts an 8//10cm fan in the center on top of the GPU, to cool it even further.
The center part are for the memory modules, and they get very hot, since they have no extra fan to cool them..
Either way, numbers don't lie. performance is higher on an Nvidia GPU, than on a stock aftermarket cooler, once you add that one fan.

Allow me to disagree. First of all, there isn't a reference-founders design for the 3060, only the 3060ti. As for the passthrough design, there is nothing revolutionary about it. It has been done before.
If anything, it is a huge problem for most casual users. For example on my B450 motherboard the M.2 slot is located right above the GPU slot, and the Gigabyte Gaming OC 3060 design incorporates
the passthrough element which in turn allows my nvme to run at a chilling 68°C... I have to run the pc with the side panel open, or buy extra fans to improve exhaust...
As for the memory, GDDR chips are flip chip BGAs and as such are closer to the PCB than the plastic housing. So they benefit from any thermal interface at the back side of the PCB, preferably
thermal pads connected to a metal backplate[again, in an ideal word, finned] that has access to adequate cooling.

I didn't say anything about "founder editions", just reference design. And reference design means the stock nvidia cooler, which is different from the aftermarket ones.
The pass through design allows for the heat to be routed towards the cpu. And in any modern case, towards the case exhaust fan, which is under the psu.
If you also turn the psu upside down, then both the case fan and psu will suck out the hot air, and you won't need to open your case. Just allow for fresh and cool air to enter the bottom of the case (which often brings in dust, if done without filter)