Comparison of current GPUs PPD/watt and PPD/price

[foldy]

Volta gets FahBench Score 230. If bigger work units get available then this will be 50% more PPD compared to Pascal.

[redant71]

Hello there, i am new to this forum and have been searching for suitable first topic to try this out and check if the image in my signature works

First i would like to say that most of the nit-picking and numbercrunching on this forum go way over my head at the moment, how ever i have come to some conclusions: AMD GPUs are less PPD-productive on the F@H system... So i scrapped out my PC1 GPU's = Radeon HD5770 Vapor-X + another noname Radeon HD5770 connected with a CrossfireX bridge, replaced that with a new ASUS nVidia GeForce GTX 1070 and was godsmacked by the hight PPD i got. on the other PC (that had one of the 5770's earlier on, and then got a Radeon 6990 4GB -bought from Ebay/ Tradera (Sweden)) -for that one i got a used GTX 980SC EVGA GPU, after reading up on PPD/GPU performance on a list from overclock.net (http://www.overclock.net/t/475163/gpu-p ... d-database)

This other PC is now to performing way better than on the AMD GPU. I guess not all is due to more expensive and newer cards, looking in to list the only decent AMD Cards is the VEGA and FURY-GPU's and they come with a massive pricetag, even compaired to the 1070 (I did not get th Ti one since it too was a bit too pricy for my taste) Now i hit 1M PPD on two PC if letting them just run GPU folding, i also read on some post on here about turning of the CPU-folding to get higher GPU work done, i do not know how or why this work, but it the way i roll now =)

So.. Don't judge the newbie here, its my first post -i may be blunt and of topic but there it is! / Anton "redant71" from Sweden

[bruce]

Welcome to foldingforum.org, redant71

Historically speaking, we all started out as newbies ... and I'd be really happy if we had a continuous flow of newbies joining FAH and asking questions that (for those of us with more experience) are easy to answer. The only "stupid question" is one that you want to know the answer to but you're too embarrassed to ask.

Also historically speaking, FAH started out running only on the "math coprocessor" of very basic CPUs. Later it was adapted for SSE instructions. Then the first GPU code was developed for AMD GPUs. Later, code was provided for NVidia GPUs and most recently, the CPU code was updated to support AVX. All of those are still supported except that SSE is now a minimum requirement.

During that 15+ years, hardware has improved faster that Moore's Law would suggest. It's market driven, though -- driven a lot by what can be sold to the game-playing public. FAH has RESPONDED to those changes rather than doing anything to drive hardware's natural growth.

Has NV outpaced ATI/AMD in the GPU field? You'll find people here biased one way or the other and they don't hesitate to express their biases as facts. I'd say that more than half of them would answer yes, but certainly not 100%. A lot depends on how you compare the two. Many will only compare the fastest GPU of each platform, forgetting that they're not matching prices.

One thing is for sure, if you spend more, you get more, and older generation GPUs can still produce useful results unless they're really several generations out-of-date.

If you have $0 to spend, then fold with whatever you have. If you're going to buy something new, then start by setting a budget and compare what you can get from various sources. Don't let somebody with a bigger budget make you feel bad. FAH's policy is to fold with whatever you already have. Most people don't do that, but FAH is not in a position to tell you how to spend your money if that's what you want to do.

[redant71]

So i understand, thank you for welcomming me!
Yes i started out with what i had, as well as trying to maximize points with as many slots as possible, utilizing all CPU-cores but one and trying to get as much out of GPU slots as well at the same time. Later on i read that i might get more points just by using GPU due to som stream-issue on the mobo(?) i think it had to do with cores and memory usage but as i said i read up on what do do -followed the instructions and saw the difference, then again i really did not know why this happened or really if that was the only way to do that... This AMD-fanboi business paired with my belief that same price gear from AMD and intel/ Nvidia would perform allmost the same and the "truth" that AMD is cheaper most of the time because the mainstream products is half a year behind intel in devolopement and to a higher TPU/ larger nano-size of the actual circuits -prooved wrong, hopefully this is just about the core of the Folding at home and how it worked at that moment i took the chioce between sticking to my guns and switching to the green team just to get better advance in folding (and also helping better to calculate proteins for Cancer sience)

As i said, i do not know if i got this wrong, but as i tried to maximize results of my AMD Radeon GPU's it came easier just by upgrading to the Geforce GPU's out of the box. At the same time i got better performance in some games, so this was not all about folding in sparetime... the PC's got more quiet and less warm too, and all together i guess it was worth the upgrade (about $500 for both, the new 1070 and the uded 980"SC" from ebay). Now i will read up how to tweak the gear i got but at the moment i am thrilled in advancing in my teams list, at the same time dissapointed on the over all team work for team 37451 (folding@sweclockers.com). Thanks again for the warm tone and happy new year!

[jrweiss]

After over 10 years of ATI/AMD GPUs, I finally bought my first nVidia. It was a hard decision, because my priorities are [lack of] noise and power consumption, NOT necessarily PPD. The choice was even harder now because there are no longer any fanless GPU cards available at the top end of the 'no extra PCIe power needed' zone. My Folding rigs are thus best described as 'low-power' rigs...

However, I now have 2 "low-power" Folding rigs: one with a new nVidia [MSI 1050ti]; vs one with a 1-year-old AMD [XFX RX460 fanless] GPU. I bought them at the same price point (~$125-140), and mainly for their similar power consumption (~70 Watts max). FWIW, the "Twin Frozr" fans on the 1050ti run at about 36%, keeping it at ~66C when Folding, and are all but inaudible at that speed.

The bottom line is that the 1050ti Folds at about twice the PPD of the RX460 - about 200K vs 100K PPD, on the current WUs. The trade is that the nVidia also demands full use of a CPU core (2 hyperthreads), vs the intermittent use of a single hyperthread for the AMD. So, my CPU slot is set for 4 cores on the nVidia rig, vs 6 cores for the AMD (leaving at least 1 core free for immediate use by other apps at all times). Considering that the CPU slot only generates 10-15K PPD, that trade is relatively insignificant. However, I have also noticed that the 1050 seems to cause more lag in graphics performance overall while Folding - not critical, but definitely perceptible.

Just remember that Folding is supposed to be a sideline. Buy your GPU for what you want to do with it in your primary computing tasks, and let the PPD come as they may. Somebody will ALWAYS come up with a bigger/better/faster Folding rig than you can afford...

[bruce]

@jrweiss
Presumably you're running the Windows 64-bit drivers for both GPUs and both are supported by hyperthreaded Intel CPUs. (Correct me if I'm wrong)
I think I'd happily donate a second hyperthread to producing twice the PPD. How do those GPUs compare in PCIe utilization as well as GPU utilization? [Maybe the nV drivers do a better job of overlapping I/O]

I just acquired a well-used Win10 machine without a GPU. I'm ordering ordering a half-height 1050Ti for it.

[artoar_11]

On 01 January, I calculated the consumption of my computer for one year. Works 90+% of time only whit GPU @ 2.0 GHz. Total idle time for the year no more than 10 days.

Total power consumption: 1740 kW (4.8 kW per day)
CPU: i5-2500k @ 4.20 GHz
MB: ASRock Z68 Ext3 Gen3 (PCI-E 2.0)
GPU: GTX 1070 ARMOR 8G OC (MSI)
Avg. PPD: 650k (Win7/64)

[foldy]

4.8 kW per day / 24 hours = 200 watts.

This is 200 watts from the wall for PC with gtx 1070 - makes sense.

[bruce]

The trade is that the nVidia also demands full use of a CPU core (2 hyperthreads) ... So, my CPU slot is set for 4 cores on the nVidia rig, vs 6 cores for the AMD (leaving at least 1 core free for immediate use by other apps at all times).

I challenge you to prove that 2 hyperthreads are required.

This system has
> CPU: Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
> CPU ID: GenuineIntel Family 6 Model 26 Stepping 4
> CPUs: 8
> OS: Windows 7 Home Premium
> OS Arch: AMD64
> GPUs: 2
> GPU 0: Bus:3 Slot:0 Func:0 NVIDIA:4 GM107 [GeForce GTX 750 Ti]
> GPU 1: Bus:4 Slot:0 Func:0 NVIDIA:5 GM206 [GeForce GTX 960]

I've been running 3 slots: 2 GPUs plus one set at CPU:6 -- using all of the hyperthreads (including 2 to feed data to 2 GPUs).
I was getting a TPF of 2:30 p9415 on the GTX 960. I set the CPU slot to finish, and it did, long before that GPU finished. After freeing up 6 more hyperthreads, FAHControl is now reporting a TPF of 2:27 after more than an hour with (more than) a second hyperthread for each GPU.

2:27/2:30 = 147/150 = 98% so freeing up that second hyperthread produced no more than 2% improvement on p9415.

[jrweiss]

@jrweiss
Presumably you're running the Windows 64-bit drivers for both GPUs and both are supported by hyperthreaded Intel CPUs. (Correct me if I'm wrong)
I think I'd happily donate a second hyperthread to producing twice the PPD. How do those GPUs compare in PCIe utilization as well as GPU utilization? [Maybe the nV drivers do a better job of overlapping I/O]

Yes, both rigs are 4-core, 8-hyperthread i7s. The 4770K/1050ti rig is running the 388.43 WHQL drivers, and the 3770S/RX460 rig has the Crimson 17.11.1 WHQL drivers.

I don't track PCIe utilization, but CPUid HWMonitor sez the 1050ti takes 25-27% of the "bus interface". I don't know if that's a portion of the max GPU bandwidth (112.1 GB/sec) or the PCIe bandwidth (PCIe 3.0 x16). HWMon doesn't give the figure on the RX460 rig, but GPUz sez it's running at PCIE 3.0 x8 with the same max GPU bandwidth and a "Memory Controller Load" of 12-25% (vs 9-14% for the 1050ti). GPU loads are both in the 95-96% range (though I often see 98-99% with both of them).

Also, it's not just a matter of "a second hyperthread". The AMD GPU uses 1 hyperthread maybe 10% of the time, while the nVidia GPU uses 2 hyperthreads 100% of the time. Still, as I indicated before, I am happy with the 1050ti.

[jrweiss]

The trade is that the nVidia also demands full use of a CPU core (2 hyperthreads) ... So, my CPU slot is set for 4 cores on the nVidia rig, vs 6 cores for the AMD (leaving at least 1 core free for immediate use by other apps at all times).

I challenge you to prove that 2 hyperthreads are required.

I stand corrected on that issue.

Previously, every time I looked at the Task Manager (usually shortly after startup), Core 21 was using 25% of the CPU (1 core or 2 threads) on the 1050ti rig. I just checked, and it is now at a constant 13% (1 thread). By comparison, Core 21 on the RX460 rig varies from 1-3%, with occasional spikes every few minutes.

The 25% usage may be attributed to an extended startup time for the WU, but I have never noticed that in the AMD rigs...

[bruce]

During WU initialization, some "real work" is off-loaded to the CPU -- and perhaps for brief periods at other times. The NVidia Driver does us a spin-wait so that if it needs a few percent to actually move the data after initialization, it'll still show as 100% busy, wasting the other ~10% of the ~12% branching back to re-test if there's data to be moved. In contrast, the ATI driver posts a WAIT and then needs to process an interrupt when more data is ready.

[HaloJones]

These ppd numbers are extremely out of date. Core22 is far more efficient and achieves at least 10-20% more than these numbers

[jrweiss]

Dunno 'bout that...

Current WU on the 1050ti is showing ~180K PPD. I've seen over 200K on a few previous WUs, but 200K PPD is still in the ballpark. Remember that there is a LOT of variation between WUs...

[HaloJones]

Shows 650K for a 1070. I'm getting over 900K. Shows 650K for a TitanX Maxwell and I'm getting over a million ppd.

Those numbers in that spreadsheet are from 2018 when Core21 was the best we had.