Folding Forum

One interesting feature of the new Radeon VII is that it features very high double precision speed.

Inherently, double precision (FP64) is one half the speed of single precision, (FP32) But in an effort to sell upmarket Workstation cards and Server cards, vendors hobble double precision in consumer cards. It does not usually impact gaming.

I have seen cards with a 32 to 1 ratio of single precision operations to double. The new Radeon VII has a very generous 4 to 1 ratio, better than most consumer gear.

Radeon VII double precision 3.5 TFLOPS, single precision 13.8 TFLOPS
Titan RTX double precision 0.389 TFLOPS, single precision 12.4 TFLOPS

So these cards are about 10% different in common math, but in scientific work the VII is impressive. It will be 9 times faster.

F@H mostly uses single precision when it can, only breaking out the slower double precision when needed to get accurate results. The VII may show us what that ratio is, because it is only a minor speed up on single precision.

(It appears that the 4 to 1 ratio is enforced in the BIOS, expect hacked BIOSes that then disappoint gamers, because their games don't use much double precision)

The Radeon VII is just a rebranded Radeon Instinct MI50 which are server class GPU intended for HPC ... it seems that DP throughput has been reduced in a half but it's still quite impressive DP power.

AnandTech has posted FAHBench SP and DP results.

I was thinking the same thing WRT The mix of Precision. I suspect that the VII will post some interesting results which may shed some light on how much DP is used. It will be interesting to see if it varies by Project or even within Projects.

Looks like it would do 1000k PPD except double precision boosts it more

So far, the FAHCores use mixed precision which is as JimbpPalmer described above, so gamers and FAH adherents won't be getting a majore increase in speed. FAHBench DP does not represent the production FAHCores but a pure DP core makes the professional GPUs look good.

The first WU I've run on the VII at complete stock (except fan speed tweaks).
Project 14152
Number of Frames Observed: 39
Min. Time / Frame : 00:00:59 - 1,688,170 PPD
Avg. Time / Frame : 00:01:00 - 1,646,140 PPD

Compared to some of my other cards (unfortunately my big cards haven't run this unit yet), this unit seems to be a typical PPD unit. So I'm quite impressed so far. Card power consumption is around 200W delta from idle. Not bad, but I'm hoping undervolting will improve.

Another: 13815

Avg. Time / Frame : 00:01:31 - 1,407,062 PPD

That seems in line with my Titan Xp

The general trend I'm seeing is good on the large atom count projects. On the smaller projects it drops substantially due to underutilization (power consumption drops by half as well).

14158 (A tiny atom count project)

Avg. Time / Frame : 00:01:16 - 419,895 PPD

Which is well below my Titan Xp at ~700k PPD (which is also well below its average).

So the Radeon VII matches a gtx 2070 or gtx 2080? Maybe the tiny atom count projects can be excluded from the Radeon VII and assigned to slower GPUs?

That's interesting ... it's been a while since we last saw numbers from high end AMD GPUs ...

I think I remember similar behaviour in the days I was using a R9 290X : it liked more the bigger projects.

In the early days of FAH, those involved in benchmarking found that ATI GPUs (all of them) liked bigger projects and it was tough to benchmark projects in a way that would keep those donors happy. All shaders ran identical code, even when the code contained branches, and the results of the unnecessary ops were simply ignored. I don't know if AMD still uses that same logic.

Yeah, I agree there is no way to fix it with benchmarking. The card power consumption drops by half on the smaller units. It isn't being fullyt utilized so it effectively is a "smaller" GPU at that point and should be getting lower PPD. I just haven't seen one to this level before in a GPU project (yes in CPU).

11719
1.13 million PPD (01:23 TPF)

See my comment here about trying to process a small number of atoms with a lot of shaders.

AMD does have ridiculous amounts of graphics memory.
Not sure if future FAH cores will make use of it or not...


@EXT: I regularly see only 87Watts used on my GTX 1060, while other cards run at 100-110W. So it's not limited to AMD.

FAH doesn't use a lot of GPU memory ... maybe up to a couple of hundred MB ...