Turing Review: GTX 1660 Super and GTX 1650

[bruce]

I've been very happy with my 1650 on FAH. It put me in the "big leagues" sufficiently that I could help with p13405. And 75 w is nice.

A small point about your review: The FAH servers are no longer at Stanford.

[MeeLee]

Edit:
EVGA precision X allows for a wider range of clock control than MSI afterburner - you can take the memory clock 2x lower on precision X. Given the 1660 super has tons of bandwidth due to the gddr6, my non scientific results show that you can clock down the ram to -1000, and leave the GPU clock higher and get the best mix of ppd per watt.

Can you share a stable overclock setting for 1660 Super on MSI Afterburner or EVGA Precision X? I don't want to break WUs for it. Also how much PPD you get with the overclock?

I can only tell you the approach on RTX GPUs, but I can imagine the GTX 16xx series GPUs work similar.
First, open your overclock utility, put fans on max, and power slider also.
Second, start folding, and open the log window.
Third, go back to your OC program, and record your GPU frequency.
Fourth, start increasing your GPU frequency, in 10Mhz increments every other minute. On most GPUs, this can take from 3 to 30 minutes.
Most GPUs of this calibre can run anywhere from 1935 to 2100Mhz boost frequency, so it'll depend on how good the cooling is, and how lucky you are with the silicon lottery.
Because you can add +200Mhz on your GPU doesn't make it better than someone else who can only add +100Mhz. It's mostly about at which frequency the GPU can run at. Both +200 and +100 Mhz could run at the same frequency.

Once the log window in FAH gives an error, dial it down by 30-40Mhz.
You can only make 3 errors, before the WU is cancelled, so make them count.
Every time you get an additional error, lower the overclock by 10MHz.
Once you have a fairly stable setting, and have no errors for the hour (you can just check the logs),
You can overclock the VRAM. Depending on the OC program, some will allow you to overclock by +750Mhz, others by +1500Mhz (both should be the same, as one will display the actual VRAM speed, and the other will display the doubled speed; they are Double Data Rate sticks).
Vram modules of GTX 1660 or faster, are 15Gbit modules, so you should be able to safely overclock them to that point.

Personally, I would approach an overclock differently.
Instead of overclocking to the max of the card,
I would look at the boost speeds you get fully stock.
Then cut power to the GPU by the maximum amount, and apply the overclock to nearing the stock (default) GPU boost settings, by doing the same procedure as above.
Your GPU will run slightly slower than with a full power overclock, but it'll run cooler, and more efficient.
The power saved on one GPU, allows you to add a second, smaller GPU (also power capped) together, while pulling nearly the same watts on the wall.
Running 2 GPUs also allows you to do more folding.

[Paragon]

I've been very happy with my 1650 on FAH. It put me in the "big leagues" sufficiently that I could help with p13405. And 75 w is nice.

A small point about your review: The FAH servers are no longer at Stanford.

Yeah there really is something to be said for efficient, low power builds. Sometimes I see people stand up these dual 2080 Ti rigs and whatnot and run them for a few months, and they disappear suddenly. I'm figuring the wake-up call was the power bill.

I'll update the bit on the servers too, I keep referencing Stanford instead of the consortium, bad habit I suppose. Have been doing this for a long time

[PantherX]

...I can only tell you the approach on RTX GPUs, but I can imagine the GTX 16xx series GPUs work similar.
First, open your overclock utility, put fans on max, and power slider also.
Second, start folding, and open the log window.
Third, go back to your OC program, and record your GPU frequency.
Fourth, start increasing your GPU frequency, in 10Mhz increments every other minute. On most GPUs, this can take from 3 to 30 minutes.
Most GPUs of this calibre can run anywhere from 1935 to 2100Mhz boost frequency, so it'll depend on how good the cooling is, and how lucky you are with the silicon lottery.
Because you can add +200Mhz on your GPU doesn't make it better than someone else who can only add +100Mhz. It's mostly about at which frequency the GPU can run at. Both +200 and +100 Mhz could run at the same frequency.

Once the log window in FAH gives an error, dial it down by 30-40Mhz.
You can only make 3 errors, before the WU is cancelled, so make them count.
Every time you get an additional error, lower the overclock by 10MHz.
Once you have a fairly stable setting, and have no errors for the hour (you can just check the logs),
You can overclock the VRAM. Depending on the OC program, some will allow you to overclock by +750Mhz, others by +1500Mhz (both should be the same, as one will display the actual VRAM speed, and the other will display the doubled speed; they are Double Data Rate sticks).
Vram modules of GTX 1660 or faster, are 15Gbit modules, so you should be able to safely overclock them to that point...

There's one issue and one potential issue with that approach:
Issue - It assumes that the WU assigned to your system is the most stressful. Even if that is the case, a new Project can be released which can push the GPU in other limits which may cause the overclock to be unstable.
Potential issue - You assume that the WU can restart three times. Sometimes, that can not happen which means you dump WUs and since these are live WUs, it hurts the science.

My suggestion to overclocking is to cool down the GPU to the best of your ability and let the GPU work out its own stable settings. This is based on Nvidia doing their own research so that's a lot of R&D that I can't match. Plus, I don't like to baby-sit systems so like the set-and-fold system

[pcwolf]

I picked up a Zotac GTX 1650 OC for a second card, one without additional power plug so it runs entirely off the PCIe 75w.

Currently delivering 335,000 PPD. Temp: 62c Fan speed: 68% Clock: 1875MHz Power: 65w

Main card is NVidia RTX 2060 FE delivering 1,600,000+ PPD. I turned the power limit down to 150w using nvidia-smi to keep the temperature ~75c. In this for the long haul. I wanted the 1650 in the second slot in the tower case, which contributes heat to the 2060. The combination works well.

ASRock X470 Taichi with Manjaro Linux

[Endgame124]

Edit:
EVGA precision X allows for a wider range of clock control than MSI afterburner - you can take the memory clock 2x lower on precision X. Given the 1660 super has tons of bandwidth due to the gddr6, my non scientific results show that you can clock down the ram to -1000, and leave the GPU clock higher and get the best mix of ppd per watt.

Can you share a stable overclock setting for 1660 Super on MSI Afterburner or EVGA Precision X? I don't want to break WUs for it. Also how much PPD you get with the overclock?

-1000 on memory
+100 on GPU

This combo was fairly close to stock ppd (around 650k) while also dropping power to around 90 watts (as measured from a ups with other stuff on it, so power numbers are not precise)

[MeeLee]

...I can only tell you the approach on RTX GPUs, but I can imagine the GTX 16xx series GPUs work similar.
First, open your overclock utility, put fans on max, and power slider also.
Second, start folding, and open the log window.
Third, go back to your OC program, and record your GPU frequency.
Fourth, start increasing your GPU frequency, in 10Mhz increments every other minute. On most GPUs, this can take from 3 to 30 minutes.
Most GPUs of this calibre can run anywhere from 1935 to 2100Mhz boost frequency, so it'll depend on how good the cooling is, and how lucky you are with the silicon lottery.
Because you can add +200Mhz on your GPU doesn't make it better than someone else who can only add +100Mhz. It's mostly about at which frequency the GPU can run at. Both +200 and +100 Mhz could run at the same frequency.

Once the log window in FAH gives an error, dial it down by 30-40Mhz.
You can only make 3 errors, before the WU is cancelled, so make them count.
Every time you get an additional error, lower the overclock by 10MHz.
Once you have a fairly stable setting, and have no errors for the hour (you can just check the logs),
You can overclock the VRAM. Depending on the OC program, some will allow you to overclock by +750Mhz, others by +1500Mhz (both should be the same, as one will display the actual VRAM speed, and the other will display the doubled speed; they are Double Data Rate sticks).
Vram modules of GTX 1660 or faster, are 15Gbit modules, so you should be able to safely overclock them to that point...

There's one issue and one potential issue with that approach:
Issue - It assumes that the WU assigned to your system is the most stressful. Even if that is the case, a new Project can be released which can push the GPU in other limits which may cause the overclock to be unstable.
Potential issue - You assume that the WU can restart three times. Sometimes, that can not happen which means you dump WUs and since these are live WUs, it hurts the science.

My suggestion to overclocking is to cool down the GPU to the best of your ability and let the GPU work out its own stable settings. This is based on Nvidia doing their own research so that's a lot of R&D that I can't match. Plus, I don't like to baby-sit systems so like the set-and-fold system

Yeah, the real overclocking starts after this.
By watching the logs daily, for failed WUs, and clocking the GPU frequency down.
Sometimes a setting that works on a cold day, might error on a hot day, if the outside temperature is, or affects the PC's ambient temperature.
Once you have less than 1 bad WU per week, and corrected for them, you can be pretty sure the OC was done correctly.
Overclocking to perfection actually takes several weeks; where one can overclock down to the Mhz.

[PantherX]

...By watching the logs daily, for failed WUs, and clocking the GPU frequency down.
Sometimes a setting that works on a cold day, might error on a hot day, if the outside temperature is, or affects the PC's ambient temperature.
Once you have less than 1 bad WU per week, and corrected for them, you can be pretty sure the OC was done correctly.
Overclocking to perfection actually takes several weeks; where one can overclock down to the Mhz.

It might work after you have adjusted for all available GPU project now. However, next year if they release a more stressful Project or a CUDA optimized FahCore or a new version of FahCore, you're back to square one Plus, you do end up wasting a few WUs if the system is being monitored. If not, a powerful GPU can end up with a variety of failures and churn through a large number of WUs.

However, once FAHBench is updated to the latest version and optimizations, using that before getting your system ready for live WUs would be fantastic!

[VegaZhree3]

Thank you guys for the replies. My card is MSI Gaming version (non-X), there is also a Gaming X version, the difference is the boost clocks. I looked up a few reviews, and put +75 on GPU clock in Afterburner to match the Gaming X version. I'm folding more than 1 month and no problems.

Maybe i'll tune more after FAHBench is updated like @PantherX said.

[MeeLee]

...next year if they release a more stressful Project or a CUDA optimized FahCore or a new version of FahCore, you're back to square one Plus, you do end up wasting a few WUs if the system is being monitored. If not, a powerful GPU can end up with a variety of failures and churn through a large number of WUs.

However, once FAHBench is updated to the latest version and optimizations, using that before getting your system ready for live WUs would be fantastic!

Not really.
The OC settings that work now, could be re-tuned later. Even if the load increases, the GPU will be tuned close to optimal on the new WUs.
In any case, it's important to regularly check the log (on a daily basis) to see if there are any failed WUs.
Once a system has been running for a few months, it's easy to see when the PPD drops by an amount, if it could have been due to an bad OC setting (bad WU).


As far as losing WUs when doing overclock on them, is inevitable.
Even with this system, folding continuously, it's not that hard to have 99.8% non-failure rates.
And that would be if you buy a new GPU every 2 months, and tune it (on regular WUs). Most people will see closer to 99.9% Completed WUs; since once a GPU sends 4 or so bad WUs, FAHClient will disable the GPU from folding.

On most overclocking, I've gained 10% in speed or more.
I'd basically be doing >110% of WUs, with a <0.3% failure rate.
It would still mean more successful WUs are done than without OC.
With more WUs done, also a higher failure rate, but that's inevitable.

I would say overclocking really benefits people who are folding 24/7, or 5 days out of a week. Not for those folding a few hours to a few days a week on occasion.

FAHBench may work for some WUs, but may not work for others. I prefer to see failures on real WUs, because the numbers don't lie.
If FAHBench succeeds on an overclock, it still could mean the overclock isn't optimal for running WUs, or might still crash WUs.
Though FAHBench is a good way to initially OC the GPU, and later use regular WUs to fine tune.

[PantherX]

...FAHBench may work for some WUs, but may not work for others. I prefer to see failures on real WUs, because the numbers don't lie.
If FAHBench succeeds on an overclock, it still could mean the overclock isn't optimal for running WUs, or might still crash WUs.
Though FAHBench is a good way to initially OC the GPU, and later use regular WUs to fine tune.

Actually, it does contain a real WU in the options. Ideally, the researchers could create a WU that is known to be very stressful thus, there's no need to waste live WUs when a test one is more than enough to verify the GPU. However, let's see what happens when FAHBench is updated

[BobWilliams757]

I just wanted to give this thread a bump for several reasons. First to thank @Paragon for doing his reviews for various F@H related topics over time, and his engagement here on the forums when he does his testing. And also to update some numbers for what might be expected with the current cores and WU assignments we are getting post COVID rush.

About the time Paragon first posted this, I was looking into something for our primary PC, and wanted to keep it simple. For that reason at that time I decided I'd probably just go with a 1650 Super for several reasons. The cost was right, I already had a power supply with a 6 pin header, and it would still be a huge improvement in F@H contribution over folding with my onboard Vega 11. Then the world went sideways for a while, prices on most things PC related went through the roof, and I decided that the best thing to do was wait it out, which I mostly did.

But with new GPU's on the market, some stability coming back to pricing on earlier stuff, and other factors I decided to go ahead and get a new more efficient power supply also and when I found a great deal on one I jumped. This opened up options for GPU's more, since I now had 8 pin headers to deal with. Mostly due to the sound level factor, I decided that a 1660 Super or ti would be the best bet for this PC that gets used for other things probably 7-8 hours a day on average. I ended up paying a bit more than needed to get one of the quietest ones I could find, the MSI Super Gaming X.


Super quiet, and even at high fan loads not much worse than my (single 120mm case fan) case fan and stock CPU cooler combo. Under any power limited folding you hardly know it's even in there. Though I'm still doing some testing and tweaking on fan curves, best efficiency, etc so far the GPU has been great.

Of the 64 completed work units I've run since installing it....

Only 4 have been below 800k PPD

25 have been 1+ Million PPD, with more than half of those being power limited WU runs, and several at an average of less than 70w at the GPU.

So fat this thing fits the bill just fine. Though I might have been able to power limit a 2060 or something enough to fit the bill, I think unless I could really pull the power down the sound level just wouldn't be worth it, and even then might have caused me to spend even more on a high end cooling solution. So for now, I'm kind of glad I waited and changed my mind about the 1650 Super.


And I decided since he helped in the decision that I'd join Paragons team, so I'm now contributing my points to Nuclear Wessels. For those that don't get it.... well you might not ever get it. But I thought it was a great team name the day I first saw it!