AMD Super Pi History To Be Rewritten, Courtesy The Stilt

[Sin]

AMD Super Pi History To Be Rewritten, Courtesy Of The Stilt

AMD's typically underwhelming Super PI performance, that was usually attributed to architectural limitations when it comes to the X87 instruction set, appears to have been nothing more than a blunder on the part of the developers responsible with BIOS development and optimization for AMD platforms. Finnish overclocker, The Stilt, figured out how to considerably improve performance by going through the BIOS developers guides. The exact same guides available to the BIOS R&D teams of motherboard vendors, a surprising fact considering a single man managed to outdo an entire industry. Here is the download link to the patch: click

The Stilt posted a video in which he showed a 4.1GHz A10-6800K completing SuperPI in 17 minutes and 34 seconds. The fastest 5GHz Richland SuperPI 32M is around 18 minutes and 15 seconds. A lot faster! For more information, check out the thread in the HWBot forums.



View at TechPowerUp Main Site

 

[xvi]

There has to be a drawback. Power consumption? Lower performance in other areas?

(p.s. Submit your scores to HWbot while you can)

 

[robal]

I don't know how to feel about this.
Stilt said "sad"...
I'm more like "horrified".

Excellent work. Respect !

 

[dj-electric]

Testing ATM, will post results

 

[Mathragh]

If anything, it seems to lower the Physx performance in Fluidmark(although it probably does nothing), which is supposedly X87 heavy.

Without patch: scored 1403, 1406 and 1417 for an average of 1408,7
with patch: scored 1407, 1380, 1378 for an average of 1388,3

Will try to do some superPI later, but apparently either Kanter was wrong, or the patch doesn't work for all x87 code.

Not trying to say the patch doesn't work, but I was thinking of physx being a good candidate for seeing some improvements.

Edit: it could of course also mean the patch isn't working correctly; The program reports my µcode being out of date, and that I should update my bios. So either Asus is lazy, or theres a bug in the program (I've got the latest bios). Also, it keeps saying fix required, even after I've pressed fix(from the second click on, it reported there was nothing left to fix).

Edit2: Oops, after some deeper digging in the original post, it seems the code is protected on Zambezi, meaning that indeed there is nothing that can be fixed for my CPU. Bummerrrr.

 

[natr0n]

There is a multicore super pi and way better benchmarks exist.

I don't know what the purpose of this is really.

 

[Mathragh]

There is a multicore super pi and way better benchmarks exist.

I don't know what the purpose of this is really.

The underlying idea is that the handling of X87 code is sub-optimal in all of AMD's latest CPU's. So not only superPI should see a benefit, but all other programs relying on those instructions aswell. SuperPI was mentioned as the program benefiting greatly because it is almost totally reliant on X87 only.

 

[Ralfies]

The underlying idea is that the handling of X87 code is sub-optimal in all of AMD's latest CPU's. So not only superPI should see a benefit, but all other programs relying on those instructions aswell. SuperPI was mentioned as the program benefiting greatly because it is almost totally reliant on X87 only.

Are there any important programs that rely on x87 instructions these days?

 

[QuackDuck]

The underlying idea is that the handling of X87 code is sub-optimal in all of AMD's latest CPU's. So not only superPI should see a benefit, but all other programs relying on those instructions aswell. SuperPI was mentioned as the program benefiting greatly because it is almost totally reliant on X87 only.

you have a Zambezi.... the patch doesn't work 100% on zambezis...

 

[Mathragh]

you have a Zambezi.... the patch doesn't work 100% on zambezis...

Yeah, I just read =( perhaps some day he/they will crack the password on the register.

Are there any important programs that rely on x87 instructions these days?

Yeah, physx is being one of them in software(CPU) mode.

Apart from that, not that many mainstream programs. Most programmers try to use as least x87 code as possible, since it is ancient, and usually quite inefficient. This might also be the reason why AMD didn't bother with fixing this, although noone is sure.

 

[AphexDreamer]

Ever since I tried the benchmark I knew something was up. So I would blame the benchmark.

 

[TheLaughingMan]

What else uses this code?

Seriously after doing some research, nothing uses x87 micro-code any more. Compiles, while they still have access to these instructions, never generate code that uses it. Anything that could have been with x87 is almost exclusively SSE of some form. Improving performance of obsolete code to make AMD look better in this one benchmark is silly. Especially when the benchmark is mainly used as a stress test.

 

[Irony]

Tried 1M with the fix

Here's with the fix enabled:

And here it is disabled:

It seems to make it alot worse; unless his labels are backwards and disable=enable.


Edit: So, I was just wanting to make sure cuz I couldn't remember my regular score at 4.5 so I OC'ed real quick to 5.2 and tried it because I know for sure that I normally do just over 17 seconds on 1M at 5.2. Down to 14 seconds for 1M with it set to disable. I think his settings are labeled backwards.

Now I tried running 32M at 5.2ghz.

 

[Jorge]

As I have been saying for years, many of the benches used to measure CPU/GPU/APU performance are tainted in that they are optimized for Intel products at the expense of AMD. This can be by intent or incompetence. When you run actual applications, you can see significant increases in AMD performance over many of the benches. When you run AMD processors on Linux, you can see even greater benefits than with Windoze. This ain't rocket science folks, it's reality. There is no financial incentive for benchmark makers to create proper, accurate benches when they profit from delivering benches that make Intel products look superior to AMD.

If all you care about is running bogus benches, then carry on. If you care about actually using your PC, then test with real applications and the best drivers.

 

[TheLaughingMan]

How would they benefit from making Intel products look superior?

 

[QuackDuck]

...
It seems to make it alot worse; unless his labels are backwards and disable=enable.
...
]

so you decided that... enable is good and disable is bad...

 

[boogerlad]

As I have been saying for years, many of the benches used to measure CPU/GPU/APU performance are tainted in that they are optimized for Intel products at the expense of AMD. This can be by intent or incompetence. When you run actual applications, you can see significant increases in AMD performance over many of the benches. When you run AMD processors on Linux, you can see even greater benefits than with Windoze. This ain't rocket science folks, it's reality. There is no financial incentive for benchmark makers to create proper, accurate benches when they profit from delivering benches that make Intel products look superior to AMD.

If all you care about is running bogus benches, then carry on. If you care about actually using your PC, then test with real applications and the best drivers.

Not just benches. Emulators and rendering run far better on Intel than they do on AMD.

 

[AphexDreamer]

So is there going to be a bios update or a way to mod a bios to have the fix?

 

[birdie]

As I have been saying for years, many of the benches used to measure CPU/GPU/APU performance are tainted in that they are optimized for Intel products at the expense of AMD. This can be by intent or incompetence. When you run actual applications, you can see significant increases in AMD performance over many of the benches. When you run AMD processors on Linux, you can see even greater benefits than with Windoze. This ain't rocket science folks, it's reality. There is no financial incentive for benchmark makers to create proper, accurate benches when they profit from delivering benches that make Intel products look superior to AMD.

If all you care about is running bogus benches, then carry on. If you care about actually using your PC, then test with real applications and the best drivers.

This post is so insane, meaningless and factually wrong I cringe thinking what's going on in your head.

 

[cdawall]

This post is so insane, meaningless and factually wrong I cringe thinking what's going on in your head.

It's been proven with quite a few different benchmarks. There should never be a switch asking if the processor is AMD or Intel, ever period end of story. cinebench proved that when AMD was allowed to use the intel data path it was substantially faster, yet there is a line of code that asks if(genuineIntel).

That is a load of BS. It should not be like that and yet it is. There are quite a few programs AMD has the ability to perform better in, yet due to program design it cannot. If you don't believe that look a little harder. It's not just AMD either there was testing done with a Via nano set to look like an Intel chip and it offered 15-30% better performance. I wouldn't try and argue something that is documented as an Intel owner I would just be mad that they are trying to prevent competition by making the competition look weaker. That is BS and anyone who has looked into it knows that.

 

[laszlo]

let's wait the next similar findings

i hope will follow soon

 

[de.das.dude]

this is why more things should be made open source.

i hope they fire the people who made the blunder. fools getting paid to do something and still dont do it better than some random dude who just did it himself.

to stilt.

 

[de.das.dude]

There is a multicore super pi and way better benchmarks exist.

I don't know what the purpose of this is really.

because occasionally a n00b intel fanboi comes along and yells "haha my intel i is better than your fx because it does pi calculation faster" :shadedshu

 

[arterius2]

because occasionally a n00b intel fanboi comes along and yells "haha my intel i is better than your fx because it does pi calculation faster" :shadedshu

well actually, my intel IS alot better than [your] fx because it does most things faster.

 

[m1dg3t]

My abbacus pWnz yoU ALL! Suk it!