AMD FX-8150 Bulldozer finally tested

[mongoled]

Actually, we already have such an issue known for Bulldozer, and NO bench-marked system has the patch installed!

The shared L1 cache is causing cross invalidations across threads so that the prefetch data is incorrect in too many cases and data must be fetched again. The fix is a "simple" memory alignment and (possible)tagging system in the kernel of Windows/Linux.

I reviewed the code for the Linux patch and was astonished by just how little I know of the Linux kernel... lol! In any event, it could easily cost 10% in terms of single threaded performance, possibly more than double that in multi-threaded loads on the same module due to the increased contention and randomness of accesses.

Not sure if ordained reviewers have been given access to the MS patch, but I'd imagine (and hope) so! Last I saw, the Linux kernel patch was still being worked on by AMD (publicly) and Linus was showing some distaste for the method used to address the issue. One comment questioned the performance cost but had received no replies... but you don't go re-working kernel memory mapping for anything less than 5-10%... just not worth it!

Thanks for this great post!

There is some light!

So, maybe we will see a 20% improvement across the board in a couple of months time.

Excited again

[Olivon]

Good to see that there's some light in the lavatory pan !

Somebody seen rintamarootta, 2good4you or even JF

[haylui]

Good to see that there's some light in the lavatory pan !

Somebody seen rintamarootta, 2good4you or even JF

no, not in Singapore.

[Michaelius]

Yes it would tell you how the chip handles gaming at settings you'll never play it.

In these 6 tests the BD cpu averaged 130 fps with AA. The i7 averaged 131.5 fps with AA.

Answer me this - if that BD chip was $50 would you even consider buying anything else for gaming? What about $100? How much is that 1.5 extra fps worth? This is the reality of enthusiast gaming, not ridiculous 600x480 resolutions.

You would think people on enthusiast forum should know that fast CPU is essential for gaming in 2011:

http://www.techspot.com/review/405-t...nce/page8.html
http://www.pcgameshardware.com/aid,7...efit/Practice/
http://www.bit-tech.net/hardware/cpu...rcraft-2-use/2
http://www.techspot.com/review/312-m...nce/page8.html

PS. 50-100$ seems like a reasonable price for AMD offering.

[zalbard]

Actually, we already have such an issue known for Bulldozer, and NO bench-marked system has the patch installed!

The shared L1 cache is causing cross invalidations across threads so that the prefetch data is incorrect in too many cases and data must be fetched again. The fix is a "simple" memory alignment and (possible)tagging system in the kernel of Windows/Linux.

I reviewed the code for the Linux patch and was astonished by just how little I know of the Linux kernel... lol! In any event, it could easily cost 10% in terms of single threaded performance, possibly more than double that in multi-threaded loads on the same module due to the increased contention and randomness of accesses.

Not sure if ordained reviewers have been given access to the MS patch, but I'd imagine (and hope) so! Last I saw, the Linux kernel patch was still being worked on by AMD (publicly) and Linus was showing some distaste for the method used to address the issue. One comment questioned the performance cost but had received no replies... but you don't go re-working kernel memory mapping for anything less than 5-10%... just not worth it!

Very interesting.
I suppose applications would not need any extra optimisations to benefit from this, then?
And if this is all true and this Windows patch is not ready then the launch will be a disaster. They should have waited some more.

[bamtan2]

Actually, we already have such an issue known for Bulldozer, and NO bench-marked system has the patch installed!

The shared L1 cache is causing cross invalidations across threads so that the prefetch data is incorrect in too many cases and data must be fetched again. The fix is a "simple" memory alignment and (possible)tagging system in the kernel of Windows/Linux.

I reviewed the code for the Linux patch and was astonished by just how little I know of the Linux kernel... lol! In any event, it could easily cost 10% in terms of single threaded performance, possibly more than double that in multi-threaded loads on the same module due to the increased contention and randomness of accesses.

Not sure if ordained reviewers have been given access to the MS patch, but I'd imagine (and hope) so! Last I saw, the Linux kernel patch was still being worked on by AMD (publicly) and Linus was showing some distaste for the method used to address the issue. One comment questioned the performance cost but had received no replies... but you don't go re-working kernel memory mapping for anything less than 5-10%... just not worth it!

link to these discussions? I don't see anything about bulldozer on lwn.net or lkml.org

[Olivon]

[flyck]

isn't this already known? and i thought they talked about 3% performance.. Seems to be about the same 'issue'.

[bamtan2]

^^power usage (dont care)

[radier]

Power usage is very important. My rig at IDLE is draining 87,5W yet at load I can use extreme CPU performance. This is how modern CPU should act.

[jimbo75]

You would think people on enthusiast forum should know that fast CPU is essential for gaming in 2011:

http://www.techspot.com/review/405-t...nce/page8.html
http://www.pcgameshardware.com/aid,7...efit/Practice/
http://www.bit-tech.net/hardware/cpu...rcraft-2-use/2
http://www.techspot.com/review/312-m...nce/page8.html

PS. 50-100$ seems like a reasonable price for AMD offering.

Starcraft 2 - First link shows the Phenom II scoring 26.7 fps at 1024x768, no AA/AF, clicking on the 2nd resolution shows 26.1 fps at 1920x1080, 4xAA/16AF. What would you play at? Oh incidentally, that puts it 0.6 fps behind it's main competition, the i5-750.

Mafia II - An i7 920 and i5 750 with a massive 60% overclock beats a stock 965 by 12%. 4% normally. This is supposed to be what, a major victory?

You didn't really look very hard at the majority of those links, did you? Only Witcher 2 shows the Phenom's in a bad light, being 44% faster than the next-gen SB and 20% slower than the real competition. Congratulations, you finally got a game where the cpu could actually be said to matter at gaming settings, 2 and a half years after the "slow" cpu was released.

[naokaji]

^^power usage (dont care)

There is a place called Europe, where electricity is extremely expensive. Power consumption is the only reason in the near future for me to throw out my i7 920 and get something else, SB-E is out of the question as it sucks just as much power, Sandy Bridge is not enough extra performance and Bulldozer? well, if you believe the numbers then it sucks just as much power as a 6 Core Phenom / any S1366 Cpu but the performance is only a improvement if you compare it to something ancient like a Phenom II x4 or a C2Q... which is why I have a hard time believing the numbers, I mean surely AMD would not be so dumb to release it's new Cpu with a performance that can only compete with its own Phenom II x4? just doesn't make sense especially as the price suggests its a Phenom II x6 successor, but unfortunately for AMD lab501 is one of the more reliable sites out there.

[gosh]

link to these discussions? I don't see anything about bulldozer on lwn.net or lkml.org

this maybe? http://comments.gmane.org/gmane.linux.kernel/1170214

more: http://www.spinics.net/lists/linux-t.../msg13140.html

[Hondacity]

Out of curiosity, what's the performance impact if the workaround is
not enabled?

Up to 3% for a CPU-intensive style benchmark, and it can vary highly in
a microbenchmark depending on workload and compiler

this might be the ummm....

[R101]

It seems to me that AMD just expected Bulldozer to clock higher.. If there was any significant penalty in OS kernel, AMD would be crazy not to have managed it long before the CPUs hit the shelves.

[AAbenson]

power usage on the FX8150 is just horrible
...i dont see how big companies that have many servers running 24/7 will buy the FX...AMD are shooting themselves in the foot big time now--->bad power usage,bad computing power

[Chrono Detector]

What, so it consumes higher power and isn't really that much powerful? And what I can see that it can't even keep up with 1100T in some benchmarks which is embarrassing. Bulldozer looks less appealing now.

[villa1n]

Actually, we already have such an issue known for Bulldozer, and NO bench-marked system has the patch installed!

The shared L1 cache is causing cross invalidations across threads so that the prefetch data is incorrect in too many cases and data must be fetched again. The fix is a "simple" memory alignment and (possible)tagging system in the kernel of Windows/Linux.

I reviewed the code for the Linux patch and was astonished by just how little I know of the Linux kernel... lol! In any event, it could easily cost 10% in terms of single threaded performance, possibly more than double that in multi-threaded loads on the same module due to the increased contention and randomness of accesses.

Not sure if ordained reviewers have been given access to the MS patch, but I'd imagine (and hope) so! Last I saw, the Linux kernel patch was still being worked on by AMD (publicly) and Linus was showing some distaste for the method used to address the issue. One comment questioned the performance cost but had received no replies... but you don't go re-working kernel memory mapping for anything less than 5-10%... just not worth it!

I guess we can expect the same issue on interlagos? That kind of glitch would seem pretty important in terms of a datacenter environment... kvm / vmware is pretty much all multi.

[donitsi]

this maybe? http://comments.gmane.org/gmane.linux.kernel/1170214

more: http://www.spinics.net/lists/linux-t.../msg13140.html

So there really are some performance problems they try to solve.

But these fixes are coming too late. (too small too?) First impression is the thing that matters and bd seems crap at launch
Plus can these things be fixed in Windows?

[Viking1962]

...lol wtf how did your comment end up quoted before you made it? who here is versed in temporal mechanics?

hahaha, he used a Bulldozer. Thats why his post came before my

[mav2000]

Something is not right here..I am confused about a few things:

1. How can a company make a new product which is lower/equal to the older product line. I dont see OCZ coming out with a Vertex 4 which is slower than the Vertex 3, unless there is some significant change in life of the said SSD

2. How are the planning to price it at 285 USD, if most reviews have shown it to be equal to or below PH II?

3. If the discussed performance is true, then should AMD have marketed this as a 4 core, 8 thread chip?

On 1, I have no answer, but I guess if anything changes then it will be on launch date ...maybe..

On 2, I have heard the number of chips available is going to be extremely low on launch...could this be the reason for the high prices? Low supply, normal to high demand??

[Hornet331]

Not sure if ordained reviewers have been given access to the MS patch, but I'd imagine (and hope) so! Last I saw, the Linux kernel patch was still being worked on by AMD (publicly) and Linus was showing some distaste for the method used to address the issue. One comment questioned the performance cost but had received no replies... but you don't go re-working kernel memory mapping for anything less than 5-10%... just not worth it!

the queston was answered, amd said ~3%

On Wed, Jul 27, 2011 at 11:57:45AM -0400, Avi Kivity wrote:
> Out of curiosity, what's the performance impact if the workaround is
> not enabled?

Up to 3% for a CPU-intensive style benchmark, and it can vary highly in
a microbenchmark depending on workload and compiler.

http://permalink.gmane.org/gmane.linux.kernel/1172226

[M.Beier]

3% isnt enough to save this.

[STaRGaZeR]

if all these benchmarks are legit (and, at this point there is very little reason to doubt them since multiple people got pretty much the same results), the the only thing we are going to see the next few days is lots of spin control.

Jf-amd will have his work cut out for him, going into full damage control mode.

good to see that there's some light in the lavatory pan !

Somebody seen rintamarootta, 2good4you or even jf

QFT

[FlanK3r]

Some dude is doing live stream benching Bulldozer , with 8 threads he got 24 GFLOPs in LinX 0.6.4

because last LINX dont stress this CPU now, u need PRIME