Best bang for the buck build?

[zizc]

Hi, I'm trying to build a dedicated system for running F@H, but I don't know what are the recommended specs, the minimum is on the website. How much should I budget for the entire build? Any recommended GPUs and CPUs? My budget is around $100-$600.

[bikeaddict]

An RTX 4060 or 4060 Ti is probably the best value and most power efficient in new GPUs. About 3 million PPD for $300 or 4M PPD for $400. There may be deals on a used 3060 Ti, 3070 or 3070 Ti on eBay, but they are less power efficient.

https://folding.lar.systems/gpu_ppd/overall_ranks

If you want to fold on CPU, I went with old HP Z440 workstations and Xeon e5-2690 v4 CPUs a couple years ago. Not sure if newer used Ryzen 9 rigs are a better value now.

https://www.ebay.com/itm/384284924158?V ... 4284924158
https://www.ebay.com/sch/i.html?_from=R ... c&_udhi=35

[zizc]

An RTX 4060 or 4060 Ti is probably the best value and most power efficient in new GPUs. About 3 million PPD for $300 or 4M PPD for $400. There may be deals on a used 3060 Ti, 3070 or 3070 Ti on eBay, but they are less power efficient.

https://folding.lar.systems/gpu_ppd/overall_ranks

If you want to fold on CPU, I went with old HP Z440 workstations and Xeon e5-2690 v4 CPUs a couple years ago. Not sure if newer used Ryzen 9 rigs are a better value now.

https://www.ebay.com/itm/384284924158?V ... 4284924158
https://www.ebay.com/sch/i.html?_from=R ... c&_udhi=35

Hi, so are you saying that if I get a RTX 4060, that will be pretty good value and I can get something a lot cheaper for the other parts like CPU, RAM, etc?

[bikeaddict]

If you only want to do GPU work units, a CPU like an older i5, i7 or Xeon quad core that can boost to 3-4GHz would be plenty to feed data to the GPU.

If you want to also do CPU work units, a 14- or 16-core V4 Xeon would be able to do around 400K+ PPD.

The power supply of older prebuilts is the only concern. The 4060 or 4060 Ti will likely need an 8-pin PCIE cable. The HP Z440 with the 700W PSU has two 6-pin cables that would require an adapter.

https://www.amazon.com/Female-TeamProfi ... 107&sr=8-3

The Green PC Gamers site has tutorials on older workstations and upgrading hardware.

https://www.greenpcgamers.com/
https://www.youtube.com/@GreenPCGamerscom/videos

[zizc]

I'm quite new to this so GPU work units are better than CPU work units in terms of performance right?

[bikeaddict]

An RTX 4060 GPU will do about 7-8x the points per day (3M vs 400K) with the same power consumption as an old Xeon. But there is a need to process CPU WUs (Project Type 0xa8) according to the stats page, so the researchers could use the help if you want to spend the money on hardware and electricity.

https://apps.foldingathome.org/serverstats

[thiblec71]

@bikeaddict, can you explain me how you read the serverstats page? Where do you see which system is in need or prioritized?

[bikeaddict]

The 0xa8 and 0xa9 rows are the current active CPU project type. The 0x22 row is the current active GPU project type.

The three Jobs columns show how many tasks are waiting for clients to download and process. And the Assign Rate shows how many jobs of each project type are being distributed to clients every hour.

For 0xa8 CPU projects, waiting job count is huge and the assign rate is fairly small.

[thiblec71]

Thanks Bikeaddict for those answers.

Would you know how F@H allocate folding power between the different projects/universities?

[pyrocyborg]

If I may, and for future reference, if you already chose your hardware, here's my two cents.

First of all, I have plenty of PCs here. Some are older generation Intel (think circa 2017 and earlier down to 2012 with my two oldies i5-3450), but none of them are CPU folding due to how energy inefficient it is (in points per watts) in my case. On top of that, folding with a Nvidia GPU uses an entire thread of your CPU at all time, so going all in with both CPU folding and GPU folding without limiting your CPU folding cores could net in an actual loss of PPD as your CPU being at 100% would hinder your GPU efforts.

Now, from my experience with fahcore 22, GPU folding on an older generation PC in 2023 (with a CPU with a manufacture date of probably 2017 and earlier) will not be as fast as using a modern PC. As I said, I have plenty of PC with CPU manufacture date ranging from 2012 to 2017, and despite having the same GPU slotted in a full PCIe 3.0 x 16 slot (that actually runs at x16, and not one of these fake 16 slots paired to a CPU that can't even handle it), they will have an average PPD lower by anywhere between 5 and 10% compared to a modern PC. Even if I swap the GPU around, it's still the same situation, so these differences are not due to one GPU being "better" than the other by a fair margin.

Here's a few examples from my current Work units listing and these are all under Windows :

RX 6800 XT folding project 12424 :
- In a 3.0x16 slot with a i5 3450 (2012) : 4.6M PPD using 200w or so.
- In a 3.0x16 slot with a 3600x (2019) : 5.0M PPD using 170w or so (which means higher PPD with less power)

RX 6700 XT folding project 12424 :
- In a 3.0x16 slot with a i5 3450 (2012) : 2.5M PPD using 170w.
- In a 3.0x1 slot with a i3 12100F (2022) : 2.5M PPD using 120w (so yes: same PPD with less power, and in a x1 slot!)

RTX 3070 folding project 18725 :
- In a 3.0x16 slot with a Celeron G3930 (2017) : 4.2M PPD
- In a 3.0x16 slot with a i3 12100F (2022) : 4.5M PPD
- In a 3.0x16 slot with a 5600G (2021) : 4.6M PPD

In some cases, especially with AMD cards, the older CPUs really can't handle undervolting or lowering the GPU without causing a whole load of issues. On top of lower PPD, the GPU have to be kept at a higher power limit to prevent crashing. I had a lot of bad work units when I started folding again in 2022 on AMD cards due to that.

Still, as long as you choose a modern CPU with 16 PCIe lanes available to your main PCIe x16 slot and a PCIe revision of a least 3.0, you'll be fine folding with your GPU, even if you have a pretty cheap CPU and motherboard combo. For reference, most if not all CPU manufactured since 2020 are PCIe 3.0 compatible, so you shouldn't run into any issues there.

As such, the best bang or your buck build depends on your objective, but if you pay for your own power, having an efficient rig might be beneficial in the long run, which means a fairly recent Nvidia GPU (RTX 4060 Ti, for example) paired with a modern CPU and motherboard combination. If you don't pay for power for some reason, then an older GPU might be fine as you won't care too much about efficiency.

[BobWilliams757]

Interesting testing on your rigs @Pyrocyborg. Though the differences aren't huge, it does show that CPU impacts GPU more than some claim. It might be interesting to test other factors and figure out what really drives those differences... is it clocks alone, clocks and cores, ability to boost? (Not suggesting I expect you to test these things) The fact that the 5600G gave a slight boost over the I3 12199F says clocks or cores might impact the speed, even with newer CPUs.

But in any case, I agree with your end assessment. The 4000 series GPU's combined with just about anything are going to be more efficient overall. Those in a situation to have someone else pay for power might find more points efficiency in older gear.

[pyrocyborg]

I may not have wrote it properly, but these data points were taken in the span of a few minutes when I wrote the message. As such, they're from different GPU on different PCs. The slight difference between the 3070 on a i3 12100F and on a 5600G might come down to luck or the typical variation we get between a work unit, or even a variation due to it being a different card (even if it's the same model from the same manufacturer). Even the 3070 that's folding on an older G3930 is currently hovering between 4.2M and 4.4M for project 18725, so the difference isn't always as important, but folding on older hardware still yield a little bit lower PPD on average compared to folding on a modern PC.

I even swapped some GPU around today for efficiency sake (we got a new CPU/Motherboard combo) and some of the cards I moved away from older PCs are now faring a little bit better, somewhere in the ballpark of 5% higher PPD for the same power usage. However, There does not seem to be any noticeable difference when it comes to folding on modern hardware (for example, hardware manufactured since 2020) as long as you have a least a full thread per Nvidia GPU. Any difference in PPD might come down to be typical variance.

It's very hard to see what affects the folding performance and I'm clearly not knowledgeable enough to ponder why, but as you pointed it out, it's possible that boost might have something to do with it as some of the earlier models I have are devoid of that function.

[BobWilliams757]

In some cases, especially with AMD cards, the older CPUs really can't handle undervolting or lowering the GPU without causing a whole load of issues. On top of lower PPD, the GPU have to be kept at a higher power limit to prevent crashing. I had a lot of bad work units when I started folding again in 2022 on AMD cards due to that.

I had intended to comment on this as well, just out of curiosity. Did you have any trends with that issue, or were there failures across a wide selection of projects? I had one single project that seemed to error on my 1660 Super, and it seemed that lower power limits caused it to happen more often. Some would be fine regardless, but the trend seemed to revolve around a power limit of below 60% or so. Undervolting/overclocking didn't seem to matter, as it was doing it with stock settings. No other project has ever given me any issues.

And similar to you, I don't claim to know every factor that causes some of these variances and such. I just pay attention to the ones I can control and try to go from there.

[pyrocyborg]

I had intended to comment on this as well, just out of curiosity. Did you have any trends with that issue, or were there failures across a wide selection of projects? I had one single project that seemed to error on my 1660 Super, and it seemed that lower power limits caused it to happen more often. Some would be fine regardless, but the trend seemed to revolve around a power limit of below 60% or so. Undervolting/overclocking didn't seem to matter, as it was doing it with stock settings. No other project has ever given me any issues.

And similar to you, I don't claim to know every factor that causes some of these variances and such. I just pay attention to the ones I can control and try to go from there.

That's a good question. I do not remember exactly, but it did happen on a lot of projects and with different GPUs. The main constant was that these were with older CPU. For example, I wanted to lower my 6700 XT power usage from 170w down to something closer to 120w, yet, anything lower than the default values resulted in either a drivers crash or multiple errors and bad work units. Same thing with my 6800 XT while it was on a i5 3450 : I could not get them to go under 195-205w without crashing for some reason, while I can on the newer PCs.

On the other hand, my newer PCs seem to be able to have a little bit more leeway with the AMD cards, but anything too low still result in an error. For example, two hours ago, I tried increasing the Core value of one of my 6700 XT (on a new PC) to see if I could manage to get better PPD in a 1% increase in power to 1% increase in PPD ratio, but I made a mistake and instead of writing 2500 as the value of choice, I did enter 1500. The card crashed while folding, and I got the "Particle coordinate is nan" error. Hopefully, the work unit was still fine and I didn't waste it.

With AMD RX 6000 cards, lower core voltage and lower core speed do tend lower the power usage while still maintaining good PPD, and power limit has no influence at all, but there is a point where PPD drops by a lot, and another point where stability is severely affected.

With Nvidia cards, it seems like adjusting the power limit is the way to go. How low it can go is dependant on the card family, the model itself and whether it won the silicon lottery or not. For example, my EVGA FT3 RTX 3090 can go as low as 55% with a loss of maybe 10% in PPD. Anything lower than that, and the PPD drops in a non linear fashion. A Gigabyte Eagle 3060 Ti could be set a 71% before it would lose a lot of PPD in exchange for a few watts.

Overclocking or downclocking is a no-go with Nvidia RX 3000 cards when it comes to folding. Reducing the core or memory values tend to reduce stability (which we don't want) and don't give anything in return. Increasing the core or memory values tend to induce a risk of errors and give almost nothing in exchange, so yeah, stock values it is.

I never tested the limit, but it's possible the Nvidia cards would crash or throw errors at one point.