Folding Forum

Just got around to formally benchmarking my RX 580 against last year's set of cards...it's a noticeable improvement over the RX 480, but it still doesn't compete on an efficiency basis with the equivalent time period Nvidia 10xx series cards.

https://greenfoldingathome.com/2019/11/ ... me-review/

Summary: 375K PPD, 250 Watts of power drawn from the wall. So, 1500 PPD/Watt (system level efficiency). The card was reporting a power draw in the driver of around 110 Watts, which seems low for a 185 Watt TDP card. Still, folding doesn't use 100% of the card's power.

Hopefully I can test some of the newer RTX series cards soon...just waiting on funds

Paragon wrote:...it still doesn't compete on an efficiency basis with the equivalent time period Nvidia 10xx series cards.

https://greenfoldingathome.com/2019/11/ ... me-review/

Summary: 375K PPD, 250 Watts of power drawn from the wall. So, 1500 PPD/Watt (system level efficiency). The card was reporting a power draw in the driver of around 110 Watts, which seems low for a 185 Watt TDP card. Still, folding doesn't use 100% of the card's power.

Based on info that's many years old (and perhaps still accurate), AMD and NV shaders process instructions differently. Each NV shader is pretty much independent of the others so if shaders need to be doing different things, that can be handled efficiently. AMD shaders operate in groups, so as long as the software needs a lot of them doing exactly the same thing, it can be efficient. An equal number of shaders at the same speed doesn't imply that OpenMM will process at the same speed. That's probably also related to power efficiency but I"m not sure how it's related.

OpenMM may sort the operations differently when it creates the Kernels but since they're both running OpenCL, I'm not sure how it would do that. On the other hand, if two FAHCores were compiled for CUDA and for OpenCL, that would give NV an added boost ... but require more complexity and Development costs. (I have no idea how much of a boost that would be.)

bruce, you're desribing the old VLIW4 or 5 (Terascale) architecture used on older AMD cards (from 2xxx to 6xxx). AMD used GCN architecture from 7xxx to Polaris (including the RX 580 and Vega cards) which closer to NV approach of independent shaders ... it explains why AMD cards with GCN showed a great improvement in FAH performance when compared to their NV counterparts.

Note : since Navi (and RX 5700 XT), AMD moved to the RDNA architecture which is using the same independent shaders.

TY

@Paragon: If you will test AMD Vega GPUs be sure to try 2 slots on same GPU. Because those have so many shaders ~4000 that only one small work unit cannot use them all.

@Foldy, The 2080 and 2080 Super has 3072 shaders, while the 2080 Ti has 4,352 shaders. They seem to process WUs just fine?

I heared with small work units RTX 2080 ti also does not get fully used and needs 2 work units running concurrently. With big work units that is not the case.

Running multiple WUs is generally not recommended. It takes a lot of extra manual effort because once you've started 2 small WU, they might be replaced with a large + small WU which will slow down both of them -- and then you have to manage the unloading process, too, untill you happen to get small WUs again, when the sequence starts over.

I have rtx 2080 and I see over %80 usage all the time. So multiple wus are not needed.

And I get about 1.525.000 ppd with only one gpu. It can get up to 1,6 time to time.

Dual work units seems only useful for AMD Vega 56/64 or Radeon VII and rtx 2080ti because they have ~4000 shaders. If you get small+big work unit they get slowed down but in total science done and PPD is more then if GPU runs only one small work unit. For other GPUs with less shader count dual work units doesn't make sense.

Are dual AMD GPUs different than dual NV GPUs.

For NV, you can have a pair of GPUs that share the load of rasterizing a single video stream (for gamers who really want improved video frame rates). They do not share the resources that FAH uses (3D calculations) so FAH treats them as independent GPUs. By default, a slot is created for each GPU so that they process WUs independently. That's not the same as running two WUs on the same logical gpu ... which is technically not supported by FAH, (but FAH gurus have figured out how to do it themselves -- and is only useful if the assigned WU is too small to more than half of your 4000+ shaders).

Manually matching up two streams of WUs that saturate about half of your shaders can be a PITA -- so you'd need to see your powerful GPU running at about 50% or less to make it worthwhile. Otherwise, you're slowing down the throughput of each WU's performance which decreases the PPD total non-linearly.

In Linux, I've never seen any GPU drop under 90%.
They occasionally fold 10-25Watts under their set wattage (125, 136, 145, or 182 Watts). Their PPD still is quite good (worst case a 2080 Ti running at 1,27M PPD).
Like Bruce mentioned, it's not worth mixing 2 WUs, because they get processed in 4 hours or less, and the next WUs might actually be 2x large WUs; in which case you'll lose a lot of the early return points.