Anand benches with BD hotfixes

[Rvenger]

Don't know if anyone saw it.

Here it is
http://www.anandtech.com/show/5448/the-bulldozer-scheduling-patch-tested

 

[Kristijonas]

Too bad they didn't bench i7 2600k (hotfix) 😉

 

[janas19]

A patch that supposedly "optimizes" the scheduler for Bulldozer architecture, yet the performance gain is only 1-2%.

Ok AMD, why not just let Bulldozer be a failure and move on?

 

[Kristijonas]

Ok AMD, why not just let Bulldozer be a failure and move on?

No need for AMDs approval on this one, hehe.

 

[Rvenger]

And still no huge gains with the Windows 8 Developer Preview...

 

[dma0991]

The more AMD tries to prove themselves with BD, the more disappointed I am. They should just scrap BD and divert their attention and resources on something worthwhile.

 

[Cerb]

The more AMD tries to prove themselves with BD, the more disappointed I am. They should just scrap BD and divert their attention and resources on something worthwhile.

All they have is BD+n and Bobcat+n. Maybe they could invest in cache development?

 

[Coup27]

I actually think it's done more harm than good. Chances are AMD had to pay MS to develop this "hotfix" as it's their hardware which changed things.

If the hotfix had given a big bump, like 15% or more than it certainly would have been worth while. When their internal testing came out around 1-2%, someone should have thought that this isn't worth it and people will just laugh at it when it's released.

 

[exar333]

The more AMD tries to prove themselves with BD, the more disappointed I am. They should just scrap BD and divert their attention and resources on something worthwhile.

At least BD is still THE CPU for gaming. Oh wait...

 

[Rvenger]

They should have added some L3 cache to Llano and upped the clock speeds instead of making triple core versions, and the fx4100.

Shoulda... woulda... coulda... well they didn't...

 

[dma0991]

All they have is BD+n and Bobcat+n. Maybe they could invest in cache development?

BD had too much cache and too little logic that actually does the work. The transistor count means nothing when they have themselves a huge cache but slow as hell. They should either put their focus more on their GPU, mobile APU or join forces with ARM and improve their presence in anything related to mobility. I think it would be very unlikely we would see BD or its future iterations taking the performance crown from Intel anytime soon.

At least BD is still THE CPU for gaming. Oh wait...

Those gamers have to be certified as visually impaired. 😀

 

[jones377]

I remember when JFAMD argued that the windows scheduler didn't need any extra tuning for Bulldozer and it should treat all "cores" equally. Many of us argued against that, saying the scheduler should use the same principle as it does with a CPU+SMT(HT). Look who was right.......

He was right in that Bulldozer doesn't pretend to be an SMT CPU to the OS apparently. If it was designed that way, this hotfix patch wouldn't have been necessary!

 

[bunnyfubbles]

BD had too much cache and too little logic that actually does the work. The transistor count means nothing when they have themselves a huge cache but slow as hell. They should either put their focus more on their GPU, mobile APU or join forces with ARM and improve their presence in anything related to mobility. I think it would be very unlikely we would see BD or its future iterations taking the performance crown from Intel anytime soon.


Those gamers have to be certified as visually impaired. 😀

well, with intel content to keep mainstream buyers hamstrung with 4 cores, AMD could theoretically make a comeback with the enthusiasts if they can offer a platform that doesn't cost $1000 just to get started.

As software becomes more and more multithread aware they won't have to compete with intel's sheer single thread or even IPC dominance if AMD can offer a brute force multi core solution that won't break the bank.

I'm extremely happy with my Sandybridge 2600K, but when I look down the road all I see form intel with Ivy and Haswell on a mainstream platform is the same 4 core strategy with all the emphasis on developing the integrated GPU, basically only progressive clockrate increases... That leaves us with socket 2011 with its current vastly overpriced chipset and CPUs that start @ $600 if you want something that can distinguish itself from the mainstream lineups.

While BD looks like a complete failure to me right now, I could very well see myself being enticed by a future version of the product right around the time Haswell hits if AMD can tighten up some things and make an attractive solution for a reasonable price, as I'm currently not at all happy with what Intel has done with s2011 in general, and I won't be content with mere clockrate and IPC increases from s1155 or s1150 with Haswell.

 

[Vesku]

Hmm, unless I missed something I think he was actually right. I mean you don't get more than a few % improvement from extra scheduler awareness. Judging from Linux results they'd be better off trying to incorporate compiler/code improvements into the Windows codebase rather than spend an inordinate amount of time on the scheduler.

Still miffed at the first iteration of BD in general. Missed targets in both fabrication and design and that's after quite a long delay from original timeline estimates.

I remember when JFAMD argued that the windows scheduler didn't need any extra tuning for Bulldozer and it should treat all "cores" equally. Many of us argued against that, saying the scheduler should use the same principle as it does with a CPU+SMT(HT). Look who was right.......

He was right in that Bulldozer doesn't pretend to be an SMT CPU to the OS apparently. If it was designed that way, this hotfix patch wouldn't have been necessary!

 

[jones377]

Hmm, unless I missed something I think he was actually right. I mean you don't get more than a few % improvement from extra scheduler awareness. Judging from Linux results they'd be better off trying to incorporate compiler/code improvements into the Windows codebase rather than spend an inordinate amount of time on the scheduler.

Still miffed at the first iteration of BD in general. Missed targets in both fabrication and design and that's after quite a long delay from original timeline estimates.

It increases performance, sometimes by a little, sometimes by a bit more. AMD needs every scrap of extra performance they can get. They should have thought of this when they designed the CPU. My guess is that marketing put their foot down because doing it the "SMT" way would mean windows would report it as a 4-core........

 

[taltamir]

the hotfixes are NOT BD specific changes but smarter scheduling for multi core processors.
The exact changes lead me to believe that they should also improve performance on an intel multi core processors, both with HT and without.

Anyone knows of anyone who tested those hotfixes with intel CPUs?

 

[Vesku]

Just pointing out that performance to effort expended-wise they would be better off getting Microsoft to optimize for BD in the actual code base and compiler used. Unfortunately for AMD they don't seem to have a great relationship with MS, look how long it took to get this scheduler change through. So they are stuck with hunting for crumbs to fill out their BD mystery meatloaf.

 

[Joseph F]

A patch that supposedly "optimizes" the scheduler for Bulldozer architecture, yet the performance gain is only 1-2%.

Ok AMD, why not just let Bulldozer be a failure and move on?

Because they have too much invested in it at this point. 🙁
I don't think that they would have released it in this state (or at all) if they had enough money to ditch it and develop something else.

 

[Natfly]

I remember when JFAMD argued that the windows scheduler didn't need any extra tuning for Bulldozer and it should treat all "cores" equally. Many of us argued against that, saying the scheduler should use the same principle as it does with a CPU+SMT(HT). Look who was right.......

He was right in that Bulldozer doesn't pretend to be an SMT CPU to the OS apparently. If it was designed that way, this hotfix patch wouldn't have been necessary!

I think he said (which he was either told by engineers or marketing) that filling up modules with threads first to allow higher turbo would offer more performance than the penalty from sharing. Either someone was mislead, they couldn't turbo high enough, or the sharing penalty turned out to be too great.

 

[Chiropteran]

A patch that supposedly "optimizes" the scheduler for Bulldozer architecture, yet the performance gain is only 1-2%.

Ok AMD, why not just let Bulldozer be a failure and move on?

It was a 4-5% gain on a couple games. The only reason the "average" gain was 1-2% was because many of the tests were (admittedly) on workloads that either use only 1 thread or all threads, in which moving threads around won't change anything.

Would it be more impressive if Anand said "since single and 8 threaded applications shouldn't benefit from this patch, we won't test them" and the average gain was 5%?

 

[dma0991]

well, with intel content....with Haswell.

I don't think $1000 is a matter to the die hard enthusiasts who wants more performance regardless of the core count.

AMD may have the upper hand with the core count but we're dealing with issues like BD's inability to reach its optimal clockspeed, heat and power consumption. AFAIK strong ST performance will benefit towards a better MT performance as well so a cluster of cores with weak ST performance will not result in stellar MT performance, which also complies to the fact that the performance scaling for more cores is not linear.

If the market is trending towards the mobile and there is more money to be made with laptops and tablets, Intel is not targeting towards more performance but less power consumption. What does the development of the IGP benefits most? It is of no concern for most desktop users but it definitely appeals a lot to mobile users. Then again it is not all that bad when QuickSync proves to be a great feature that is absent from the LGA2011 processors.

I would definitely like a Core i7 3930K for its 6 cores and the sheer size of the processor but I'm pretty sure there will be a 6 core mainstream processor in the future versions after Haswell, you'll just have to live long enough to see it 😛.

 

[Gikaseixas]

The problem with BD is what it is up against with. Intel is just on a roll and has a excellent line up right now.

 

[TheWay64]

It was a 4-5% gain on a couple games. The only reason the "average" gain was 1-2% was because many of the tests were (admittedly) on workloads that either use only 1 thread or all threads, in which moving threads around won't change anything.

Would it be more impressive if Anand said "since single and 8 threaded applications shouldn't benefit from this patch, we won't test them" and the average gain was 5%?

If you looked at anand benchs you would notice that 5% is the best case scenario, not average.

 

[Anarchist420]

I actually think it's done more harm than good. Chances are AMD had to pay MS to develop this "hotfix" as it's their hardware which changed things.

If the hotfix had given a big bump, like 15% or more than it certainly would have been worth while. When their internal testing came out around 1-2%, someone should have thought that this isn't worth it and people will just laugh at it when it's released.

15% of nothing is nothing.

 

[Joseph F]

15% of nothing is nothing.

No, a 15% boost would make it a hell of a lot more competitive.