Folding Forum

I was wondering if the 3070 Ti might be a better buy than the 3080 for folding?
The 3080 has +2k more cores, but not all of them will be used.
It also has a 320bit memory bus, vs 256 for the 3070Ti.
I'm not sure if that would help FAH, since it still has plenty of VRAM speed, and not sure how much a 256 bus bottlenecks PPD vs a 320bit bus.
Third, it has more RAM, which I know for a fact FAH doesn't use.
It barely gets 2GB of RAM on large WUs. I can't imagine it needing more than 6GB.

Anyone has PPD values for both GPUs?

https://folding.lar.systems/gpu_ppd/overall_ranks indicates 3080 has higher throughput.

I am getting about 3M PPD on the 3070 (regular) and 5-8M on the 3080 Ti. Occasionally, there will be a project for which the 3080 Ti only gets 1-2M

jchang6 wrote:I am getting about 3M PPD on the 3070 (regular) and 5-8M on the 3080 Ti. Occasionally, there will be a project for which the 3080 Ti only gets 1-2M

3070 vs 3080Ti is 3 steps up.
3070 Ti vs 3080 is only 1 step up.

MeeLee wrote:

jchang6 wrote:I am getting about 3M PPD on the 3070 (regular) and 5-8M on the 3080 Ti. Occasionally, there will be a project for which the 3080 Ti only gets 1-2M

3070 vs 3080Ti is 3 steps up.
3070 Ti vs 3080 is only 1 step up.

boys gotta have their toys - temper tantrum galore without

Neil-B wrote:https://folding.lar.systems/gpu_ppd/overall_ranks indicates 3080 has higher throughput.

I just put all the online results in an excel chart.
If you're going to purchase a new 3000 series GPU on amazon, your best bang for the buck is apparently a 3080 Ti.
Since it's only $75 more expensive than a 3080, and $125 more expensive than a 3070; but performance wise, it's about 47% higher PPD than a 3070 Ti.

PPD 3080 = *Edit: Should have been a performance metric here that kind of got lost...
Price 3080 = +3% of a 3070 Ti (about $50 difference on amazon).
PPD 3080 Ti = +46% of a 3070 Ti.
Price 3080 Ti = +10% of a 3070 Ti (about $170 difference on amazon).

A word of caution re LAR stats (as with any stats) is that the state of the kit running the gpus and their setup is unknown/variable ... My ROG-STRIX-RTX3070-O8G-GAMING actually performs far better than the average on lar - and that is on Win 10 - however it is optimised for FaH.

I am looking at (the mythical) ROG-STRIX-LC-RTX3080Ti-O12G-GAMING as my next acquisition so kind of matches your research ... but as with everything these days it will come down to availability and at what price they are being sold.

The folding.lar resource is great but like all statistical systems is open to a lot of variables. Are the cards measured while the rest of the system is completely idle? Have they been undervolted (common on 3xxx cards)? Are they adequately cooled? have they selected a preferred cause?

My undervolted water-cooled 3080 which is not the active gpu in the system (I use the iGPU for running my desktop) is still getting >7m with Alzheimer's selected as the preferred cause. folding.lar always says my output is above average.

Now there's news of 3000 series super cards (3060S, 3070S, 3080S).
Should be somewhere in between the regular version and the Ti version.
The Ti versions have significantly more shaders. The Super versions only have slightly more shaders, but some memory upgrades and other minor things (like a few more RT cores).

Here's a video link explaining why it may not be possible: https://www.youtube.com/watch?v=J-kw_UMPqTo (From: JayzTwoCents channel on YouTube).

Paul

psaam0001 wrote:Here's a video link explaining why it may not be possible: https://www.youtube.com/watch?v=J-kw_UMPqTo (From: JayzTwoCents channel on YouTube).

Paul

He's only claiming this about the 3080. Not the other GPUs (3070 super and 3060 super).
The Super cards are actually GPU cores that haven't made it due to manufacturing defects, and have some shaders disabled, and a few memory upgrades, and sold as a lower end model.
That's how Nvidia gets rid of their excess on 'bad' GPU cores.
Like a 3060 is a 3070 with a toned down GPU core and some memory upgrades over the 3060 (or memory downgrades over the 3070).

I would counter argue what JZ says.
Selling Super cards makes a ton of sense!
it releases a bottleneck Nvidia currently has; getting gpus to the customer.
And as far as TSMC not letting them make this, is also untrue.
Nvidia has a Queue with TSMC.
It's up to them to decide what they will manufacture in this queue.
All Nvidia needs to do is give them the schematics at the time when TSMC opens their queue.
And whether they will run 1000 wafers of a 3080, or 500 wafers of 3080 and 500 of 3070 Super, doesn't matter to TSMC.
Switching schematics is like a half an hour job (just getting the right negatives or blueprints in the etcher).