AMD RX 480

[Madpoo]

Originally Posted by airsquirrels

Ok I got the 1080 Founders Edition running.

It is still failing 43/107 self tests using 8.0 and compute 6.1, so it seems the mfaktc issue is still present. I skipped the self test to get performance numbers and I'm seeing 1071 GhzDay/Day up to 75 bits, which is 23% faster than my Fury X results.

Also compiled CudaLucas. M49 (4M FFT) runs at 3.4ms/iter, 70 hour ETA

70 hours to do M74207281 isn't bad. It sounds like despite the skepticism because it's 1/32 FP64, it's still not a bad LL testing bit of silicon. I mean, it has a lot of FP cores so 1/32 of a lot is still a good amount... isn't that how it works?

Sure, there are faster things out there, and in the end it may be more useful to have something like that doing TF work (more overall throughput), but I say use it for whatever you have more fun with.

[airsquirrels]

70 hours at a mere 180 watts. That's the most impressive part.

[henryzz]

The switch from 28nm to 16nm will have helped a lot.
Hopefully they will keep up to Intel in future rather than slip so far behind.

[jasong]

Originally Posted by henryzz

The switch from 28nm to 16nm will have helped a lot.
Hopefully they will keep up to Intel in future rather than slip so far behind.

Is Intel legitimately ahead, or did they cheat in some way that I don't know about.

I'm specifically thinking of them bribing companies into delaying or not allowing Athlon-based computers to be sold.

Athlon is a brand of cpu that came out while Intel was building P4s. AMD did significantly worse than it could have because Intel used it's clout to prevent companies from selling the Athlons.

It's kind of like the Windows tax inflating the Windows numbers even while people immediately install Linux on the computers they bring home.

[airsquirrels]

Interestingly enough, performance at the 2048K FFT is actually even more impressive:

FFTSize: 2048K Exponent: 38410247 (0.10%) Error: 0.18750 ms: 1.7378 eta: 18:27:14
Card 1 (Graphics Device - 75.00C, 100% Load [1833/607]@145.70W/180.00W, M38410247 using 2048K) GhzDay: 68.23

That's not far off from my Titans, and significantly better than my Fury Xs.

[Madpoo]

Originally Posted by jasong

Is Intel legitimately ahead, or did they cheat in some way that I don't know about.

I'm specifically thinking of them bribing companies into delaying or not allowing Athlon-based computers to be sold.

Athlon is a brand of cpu that came out while Intel was building P4s. AMD did significantly worse than it could have because Intel used it's clout to prevent companies from selling the Athlons.
....

As far as I know, Athlons were always horrible at FP, nothing new about that. Their other stats were impressive and they made some decent advances that pushed both Intel and AMD ahead by competing. But I don't think they were ever competitive in the precision computing arena, were they?

[ATH]

That is very impressive performance considering FP64= 1/32 * FP32.

But it still seems like a waste to use this for LL when it is the fastest mainstream graphic card for trial factoring.

Is there actually anything faster than this except the new Tesla P100 for trial factoring?

[mackerel]

Originally Posted by Madpoo

As far as I know, Athlons were always horrible at FP, nothing new about that. Their other stats were impressive and they made some decent advances that pushed both Intel and AMD ahead by competing. But I don't think they were ever competitive in the precision computing arena, were they?

http://s5.postimg.org/jg4d7uciv/llr_perf.png

That's a performance comparison I did in 2005. I never gave architecture much thought at the time and just wanted something fast. I guess if you were to look up the actual clock and estimate the values from the image (it is unlikely I will ever find the original sheet if I even still have it), you could estimate an IPC for the processors for prime finding application the time. The software I used at the time was LLRnet, so presumably that was a network wrapper for LLR. I was running what was the rieselsieve project at the time. Without in depth analysis, it does look like this might be an area where Intel's higher clock actually gave some advantage. Even HT looks like it helped throughput some although I didn't use that for LLR.

For context, Intel was still in full Netburst swing. Multi-core hadn't come along yet. AMD had come out with 64 bit extensions although no one used it at the time. I was likely running either Windows 2000 or XP then depending on the age of the system.

Fast forward, I think where Intel really jumped ahead in CPU FP was with Sandy Bridge, with another big jump from Haswell. I got a feeling that the post ATI merger AMD wanted FP to be done by the GPU allowing them to cut back on spending CPU resource there.

[Mark Rose]

Yeah, once the Athlon came out, it was king until the Core arrived in 2006. That's when Intel dropped Netburst and went back and extended the PIII design.

[jasong]

Originally Posted by Madpoo

As far as I know, Athlons were always horrible at FP, nothing new about that. Their other stats were impressive and they made some decent advances that pushed both Intel and AMD ahead by competing. But I don't think they were ever competitive in the precision computing arena, were they?

Good point, I temporarily forgot the forum I was in when I asked that question.

[henryzz]

Originally Posted by mackerel

http://s5.postimg.org/jg4d7uciv/llr_perf.png

That's a performance comparison I did in 2005. I never gave architecture much thought at the time and just wanted something fast. I guess if you were to look up the actual clock and estimate the values from the image (it is unlikely I will ever find the original sheet if I even still have it), you could estimate an IPC for the processors for prime finding application the time. The software I used at the time was LLRnet, so presumably that was a network wrapper for LLR. I was running what was the rieselsieve project at the time. Without in depth analysis, it does look like this might be an area where Intel's higher clock actually gave some advantage. Even HT looks like it helped throughput some although I didn't use that for LLR.

For context, Intel was still in full Netburst swing. Multi-core hadn't come along yet. AMD had come out with 64 bit extensions although no one used it at the time. I was likely running either Windows 2000 or XP then depending on the age of the system.

Fast forward, I think where Intel really jumped ahead in CPU FP was with Sandy Bridge, with another big jump from Haswell. I got a feeling that the post ATI merger AMD wanted FP to be done by the GPU allowing them to cut back on spending CPU resource there.

AMD really lost out LLR/Prime95 performance until their SSE2 implementation caught up in speed with Phenom II. I remember moving from a 2.4 Ghz Athlon 64 to 2.4 Ghz Q6600 and getting an 8x speed improvement.