Return of FX : GA-990FXA-UD7 - OC'd AMD FX 8150 / 6990 Performance Comparison Review

[soccerballtux]

single -- 1.16

 

[soccerballtux]

have you done any 4-thread testing on Bulldozer, so we can see "without module contention, bulldozer is this much faster per clock than Ph2"?

 

[MentalIlness]

I'd be willing to throw a few scores into the mix if need be. Let me know what to run and how.

 

[sandorski]

Haha, I clicked the second link and guess what ad came up?

:biggrin:

 

[sandorski]

I don't expect Trinity/Piledriver to completely turn things around, but if they can improve performance by 10ish% and reduce Power Consumption by 20-25%(all numbers guesstimates), then at least people might seriously consider getting one.

I suspect Trinity will do well regardless, especially the Low Power ones whether in Laptops or some kind of Mini-HTPC configurations. It really seems up in the air about Piledriver and the Enthusiast Market though. That segment requires at least the improvements mentioned in the above paragraph.

 

[Don Karnage]

OT but I just found this video. Hilarious

https://www.youtube.com/watch?v=nwF1zGrNriI&feature=youtube_gdata_player

 

[TakeNoPrisoners]

I thought it was ironic that you would comment on the OC'd bulldozer's power usage like 568W was something crazy when your own setup probably uses double that at load.

That is, it's an ironic statement when you make it sound like 500W is a tremendous amount of power, when reality is that it's far less than your system's power consumption.

Yea well his PC is actually fast...

BD isn't.

 

[Chiropteran]

Incorrect. I guess you suddenly forgot that Hyper-Threading is a 20% improvement in MT.

"Ivy" doesn't mean a CPU is hyper-threaded. There are many ivy bridge CPU without hyper-threading, and I specifically said 4 cores vs 8 cores. Why would you assume I was talking about an i7?

Now, the truth, according to Anandtech, reveals that my post was completely correct and you are wrong:

http://www.anandtech.com/bench/Product/288?vs=434

i5 2500k is 48% faster than FX-8150 at a single thread.

i5 2500k is 0.63% faster than the FX-8150 at the multi-threaded cinebench test. This is propaply within margin of error, might as well say they are equally fast.

Now my earlier reply was based on the poster saying "ivy" (I have no idea if he was talking about an i3, low end i5, higher end i5, i7, or what) was 40% faster with a single core compared to FX-8150. It seems to me my response was 100% correct.

If i5 2500k, with a 48% per core advantage, is only equal to FX-8150 at multithreaded, then it stands to reason that a slower CPU with only a 40% per core advantage would actually be slower overall when multicore performance is measured.

And Cinebench is mostly floating-point code, and Bulldozer has four FPUs.

Oh really? That has to be incredibly embarrassing then, if FX-8150 can match 2500k while limited to 4 FPU.

 

[pelov]

Now my earlier reply was based on the poster saying "ivy" (I have no idea if he was talking about an i3, low end i5, higher end i5, i7, or what) was 40% faster with a single core compared to FX-8150. It seems to me my response was 100% correct.

If i5 2500k, with a 48% per core advantage, is only equal to FX-8150 at multithreaded, then it stands to reason that a slower CPU with only a 40% per core advantage would actually be slower overall when multicore performance is measured.

You're forgetting CMT tax in your hypothetical scenario. As soon as you factor that in (-20% per filled module), along with the 4 FPUs (though they perform well enough individually there just isn't enough of them) you get this

It performs roughly like 6 individual Thuban cores despite the higher clock speed and more ALUs. The architecture needs a lot of work so let's hope they've done enough.

As was mentioned already, Trinity should be great in mobile even if it can't compete with IB/SB on the CPU level. Where HTPCs and lappies are concerned, better perf-per-watt, far better GPU performance and cheaper chips should certainly sell well. Vishera, on the other hand isn't likely to challenge Intel on the desktop other than maybe the low end.

 

[Chiropteran]

It performs roughly like 6 individual Thuban cores despite the higher clock speed and more ALUs. The architecture needs a lot of work so let's hope they've done enough.

How does thuban have anything at all to do with comparing bulldozer and "ivy"? The poster made a post bragging about how per thread ivy was 40% faster, I was simply pointing out that measuring performance "per thread" when one CPU has twice as many threads isn't really telling the whole story.

For example, despite the "per thread" advantage of i5 in that graph, the 8150 actually beats.

 

[BD231]

How does thuban have anything at all to do with comparing bulldozer and "ivy"? The poster made a post bragging about how per thread ivy was 40% faster, I was simply pointing out that measuring performance "per thread" when one CPU has twice as many threads isn't really telling the whole story.

For example, despite the "per thread" advantage of i5 in that graph, the 8150 actually beats.

That 40% figure still stands as you need twice the hardware to do what four single Intel cores can at stock speeds. Nobody wants crap single threaded performance, and nobody wants sh tty power envelopes. Unless of course they enjoy ignoring the fact that there are cheaper options on the table that are twice as fast and efficient in the majority of software available right now.

Most free encoding programs have thread limitations, and if you don't fully utilize those 8 threads (which 99% of typical users don't) you're not going to get anywhere near the performance of an Ivy. If you'd like to be as close minded as AMD's PR team and lie to people based on one performance variable at least have the decency to say that in the pursuit of low power high performance computing Bulldozer is the literally the sh ttiest CPU ever made within its time.

If anyone sends me an 8150/8120 bulldozer CPU free of charge I will personally make a youtube video of myself smashing it with a hammer, or, have a friend roll over it repeatedly in an H2 hummer with 30 inch chrome rimz. Seriously, there's no good reason to own bulldozer.

 

[LOL_Wut_Axel]

"Ivy" doesn't mean a CPU is hyper-threaded. There are many ivy bridge CPU without hyper-threading, and I specifically said 4 cores vs 8 cores. Why would you assume I was talking about an i7?

Now, the truth, according to Anandtech, reveals that my post was completely correct and you are wrong:

http://www.anandtech.com/bench/Product/288?vs=434

i5 2500k is 48% faster than FX-8150 at a single thread.

i5 2500k is 0.63% faster than the FX-8150 at the multi-threaded cinebench test. This is propaply within margin of error, might as well say they are equally fast.

Now my earlier reply was based on the poster saying "ivy" (I have no idea if he was talking about an i3, low end i5, higher end i5, i7, or what) was 40% faster with a single core compared to FX-8150. It seems to me my response was 100% correct.

If i5 2500k, with a 48% per core advantage, is only equal to FX-8150 at multithreaded, then it stands to reason that a slower CPU with only a 40% per core advantage would actually be slower overall when multicore performance is measured.




Oh really? That has to be incredibly embarrassing then, if FX-8150 can match 2500k while limited to 4 FPU.

No, because Sandy/Ivy Bridge only has four FPUs too. And the i5-3570K is faster than the 8150 in Cinebench, all while being clocked lower. Clocked the same, the 2500K would match the 8150 in Cinebench--a rare sight, given the 2500K seems to slaughter it in most tasks.

Nice try, though.

 

[Chiropteran]

No, because Sandy/Ivy Bridge only has four FPUs too. And the i5-3570K is faster than the 8150 in Cinebench, all while being clocked lower. Clocked the same, the 2500K would match the 8150 in Cinebench--a rare sight, given the 2500K seems to slaughter it in most tasks.

Again, you are pulling out models to fit your argument. I was responding to a guy saying that whatever "ivy" cpu was 40% faster than bulldozer when testing a single thread. 40%. We both know from above that i5 2500k is 48% faster when testing a single thread. Unless the i5-3570k is slower than the 2500k, it isn't the CPU in question.

Let me spell this out since everyone seems to have problems:

If a hypothetical CPU is 40% slower at a multithreadable task, but can handle twice as many threads, it will generally be FASTER when allowed to run all of those threads. The i5-2500k vs FX-8150 bench results support this completely.

If you are going to bring up some other CPU that is not "40% faster" than 8150 at a single thread, then you are NOT talking about the same situation at all.

Examples: 3570k is presumable faster than the 2500k, a CPU which is 48% faster, so it doesn't fall into the category of "CPU which is 40% faster".

This isn't even about bulldozer vs ivy or any other architecture, I was simply noting that "single threaded IPC" is a terrible measurement of a CPU when you are referring to a test that is normally VERY threadable. Some posters seem to be so blinded with anti-AMD rage that they can't see the point for what it is, and try to twist the argument into one that makes AMD look worse. I know bulldozer is a a cheap low end CPU. I like it because it's fast enough for me, and cheap, but I am not trying to trick myself into thinking it can compete with intel CPU which cost 50% to 100% more dollars.

 

[polyzp]

"Ivy" doesn't mean a CPU is hyper-threaded. There are many ivy bridge CPU without hyper-threading, and I specifically said 4 cores vs 8 cores. Why would you assume I was talking about an i7?

Now, the truth, according to Anandtech, reveals that my post was completely correct and you are wrong:

http://www.anandtech.com/bench/Product/288?vs=434

i5 2500k is 48% faster than FX-8150 at a single thread.

i5 2500k is 0.63% faster than the FX-8150 at the multi-threaded cinebench test. This is propaply within margin of error, might as well say they are equally fast.

Now my earlier reply was based on the poster saying "ivy" (I have no idea if he was talking about an i3, low end i5, higher end i5, i7, or what) was 40% faster with a single core compared to FX-8150. It seems to me my response was 100% correct.

If i5 2500k, with a 48% per core advantage, is only equal to FX-8150 at multithreaded, then it stands to reason that a slower CPU with only a 40% per core advantage would actually be slower overall when multicore performance is measured.




Oh really? That has to be incredibly embarrassing then, if FX-8150 can match 2500k while limited to 4 FPU.

Still there are numerous cases where an FX 8150 with 8 threads will beat the i5 2500k.

The i5 2500k is only more appealing to gamers because it has strong 4 core performance, where many dx11 games still only use 2-4 cores effectively.

Multi threaded encoding performance for example will show you how the i5 2500k's 4 threads cant keep up with more cores. By 2013 cores will be needed more and more, hence why Haswell Extreme will have 16-20 threads. Even intel knows.

 

[AtenRa]

That 40% figure still stands as you need twice the hardware to do what four single Intel cores can at stock speeds.

One FX Module is almost equal in size to a single Intel SB core. There are MT applications where Bulldozer's Module is equal in performance to a Intel SB Core. PileDriver will close even further the gap and i believe in some MT apps a PileDriver Module will be even faster than a Single SB core.

Nobody wants crap single threaded performance, and nobody wants sh tty power envelopes.

First of all, IPC is getting harder and harder to raise, AMD and Intel knows that. AMD have chosen to pursue the MultiThreaded road sooner than Intel. If users need a strong single thread performance it means they dont need a highly parallel CPU and they using old Single-Threaded applications. The future is MultiThread and Multi-Task and both AMD and Intel knows it.

There are Multithreaded apps that FX8150 is more power efficient than Core i5 2500K. You need not to only measure the power consumption but also the time needed to complete the task.

Measuring only the highest power consumed i could come to a wrong conclusion that ATOM 2700 is the most efficient CPU ever, but if it takes one hour to finish what a 10 times more power hungry CPU will take a minute then things change

Unless of course they enjoy ignoring the fact that there are cheaper options on the table that are twice as fast and efficient in the majority of software available right now.

What Intel CPU is cheaper and twice faster ?? plz state CPUs and apps involved.

 

[pelov]

If i5 2500k, with a 48% per core advantage, is only equal to FX-8150 at multithreaded, then it stands to reason that a slower CPU with only a 40% per core advantage would actually be slower overall when multicore performance is measured.

Chiropteran, I was replying to this statement. If you neglect the -20% CMT tax then your argument would be correct, but current architecture/software and the lack of IPC and clock speed dictate that this isn't the case, which is why in many of the reviews the 2500K and Thuban both compare well to Bulldozer in multi-threaded scenarios. It's a perfect case for why moar coars isn't better as both IPC and clock speed both will decrease as you increase core count and that both of those factors (among others) play a huge role in multi-threaded performance as well.

If your hypothetical argument were right then BD would stomp over everything with with 4 cores or less when tasked with >5 threads (mainly integer), but that just isn't the case. Sharing the front end, cache and the decrease in IPC and hampered clock speeds show that this just isn't true.

Remember that as you go up in core count there's a diminishing return on clock speed and IPC both, as well as larger die size (generally, though CMT skirts that rule a bit). Unless you're compiling and optimizing for specific architectures with higher core counts then it just doesn't make sense. BTW, we don't compile or optimize for desktop CPUs outside of ISAs, really. The OS plays a role but mainly as far as scheduling goes.

Moral of the story is that moar coars can be good for server but moar coars on the desktop/laptop doesn't make any sense.

 

[AtenRa]

Moral of the story is that moar coars can be good for server but moar coars on the desktop/laptop doesn't make any sense.

Depends on the application used. More and more programs become Multithreaded in desktop. Not to mention you cannot have a 40% increase of IPC from one CPU generation to the next anymore. But you can have 30-50% more performance through higher core/thread count.

Im not saying that IPC doesn't count, but it all depends on the applications we are using.

 

[Maximilian]

Stay Tuned! I am free to answer any questions and also am willing to take any requests you guys might have!

Send BD231 your chip.

If anyone sends me an 8150/8120 bulldozer CPU free of charge I will personally make a youtube video of myself smashing it with a hammer, or, have a friend roll over it repeatedly in an H2 hummer with 30 inch chrome rimz. Seriously, there's no good reason to own bulldozer.

 

[pelov]

Depends on the application used. More and more programs become Multithreaded in desktop. Not to mention you cannot have a 40% increase of IPC from one CPU generation to the next anymore. But you can have 30-50% more performance through higher core/thread count.

Im not saying that IPC doesn't count, but it all depends on the applications we are using.

Definitely and in a sense hardware has always been years ahead of software but with a tick-tock schedule now adapted by both CPU manufacturers there's no need to make something that forward-thinking if it means your chip isn't as appealing as it could be for today's workload. If AMD/Intel want to introduce AVX2 they can take another 5 years and we still wouldn't know the difference because the hardware depends heavily on the software to stretch its legs. It's one of my gripes with Haswell. Intel isn't really introducing anything revolutionary on the CPU at all other than ISAs so I'm not sure why anyone would want a Haswell chip if it means you'll get roughly equal performance outside of the on-die GPU. New instruction set? That's the same thing AMD said when they announced an 8 and 10-core processor. Great. Call me when it matters.

Core count has increased because it's an easy substitute to actual marked improvements; you can market your chip as an "8-core processor" or "4 cores is better than two" when in reality the majority of users still won't benefit from anything over 2 and they'd be far better off with power savings and cheaper CPUs than they would be with 8 core monsters (and it's not 8 cores but rather 4 modules).

I can't shake the feeling that AMD thought Bulldozer would bring about an x64-type revolution in computing that happened years ago. But unlike x64, which was a natural progression at the right time, the Bulldozer design and architecture is hopelessly too radical to make good use of and its current implementation is lackluster at best. Escalation in core count is the future, I've no doubt about that. But if it comes at the cost of IPC, clock speed, die size and power consumption and much too early to matter then please put it on the shelf and bring it out when it makes more sense.

edit - Just to clarify, by die size I meant overall die size and the size of the individual cores/modules, which are quite thin. The die size of the BD chip in total is far too big, though.

 

[Chiropteran]

Chiropteran, I was replying to this statement. If you neglect the -20% CMT tax then your argument would be correct, but current architecture/software and the lack of IPC and clock speed dictate that this isn't the case, which is why in many of the reviews the 2500K and Thuban both compare well to Bulldozer in multi-threaded scenarios.

I was not speculating. Look at my post, follow the links. I was posting actual real performance data.

In the cinebench, which was the application in question, what I said follows reality perfectly.

I'd agree if you were talking about other software that you often can't scale as well (or at all), but the post I responded to was specific in that it was referring to cinebench performance, one case where the software scales very well on FX's multicore architecture.

 

[pelov]

Except that it doesn't. The decrease in single-threaded hurts it as well as the CMT-tax (20% per filled module), therefore you get a 2600K with only 4 cores and 8 threads still beating it and a Thuban scoring just below it (as does that 2500K).

Scales well? If it sacrifices single-threaded and can't make up that gap in multi-threaded then regardless of how well it "scales" it still isn't good enough.

CMT doesn't do well in Cinebench and it's a benchmark that AMD ignore altogether therefore I'm not sure your persistence on using that benchmark to show the scaling of modules is a good idea. Truecrypt and 7-zip yes, but Cinebench?

 

[sandorski]

Cores will continue to increase in number. Sorry, but IPC and Hz have been butting up against a ceiling for close to a decade now.

 

[Mallibu]

Depends on the application used. More and more programs become Multithreaded in desktop. Not to mention you cannot have a 40% increase of IPC from one CPU generation to the next anymore. But you can have 30-50% more performance through higher core/thread count.

Im not saying that IPC doesn't count, but it all depends on the applications we are using.

Knowing that:
~50% of the people have dual cores, and another ~40% have quad cores (saw that some days ago from steam surveys),so we are in the phase of mainstream software transition from 2 cores/threads to 4 cores/threads. And we're not even in the middle of it (how many programs the everyday user has, that utilise more than 4 cores?). There's also a limit to how much you can write code efficiently to utilise more threads as the thread count increases, and it's certainly not linear improvement, except workloads like video conversion.
In the other hand, by IPC we're not talking about "old single threaded programs" as you make it seem so. In programs, that use 2 to 4 threads, aka 99% of today's and at least the next 2-3 years software you gain a lot with 4 strong cores, than 8 weaksauce ones(FX).

Play some games (not 3-4 cherry picked gpu-bound ones :whiste and observe that in almost every game,there's usually the first one or two threads that are getting >80-90% usage, with the rest hovering below 40%. With weak IPC/Core power , you are always waiting for your slowest thread so more IPC is never bad. Play some Shogun 2/ Civ 4 / Starcraft 2 and observe it yourself.
I went from C2Duo E8400-> Phenom II 965@3.8 ghz, and with about the same IPC and twice the cores,I saw some gains in selected games but still meh in comparison to what i was expecting. I went from the Phenom to i5 2500@3.8 ghz and with ~+40% IPC I got, i saw a lot more difference in everything, not selected few programs that people use one time per month.

 

[Chiropteran]

Except that it doesn't. The decrease in single-threaded hurts it as well as the CMT-tax (20% per filled module), therefore you get a 2600K with only 4 cores and 8 threads still beating it and a Thuban scoring just below it (as does that 2500K).

As they should, since the FX-8150 is a cheaper less expensive part compared to those other CPU. Thuban die size is larger than 8150. Intel CPU are on a smaller more advanced process.

Fact remains, it goes from being 48% behind at single threaded to DEAD even in the multithreaded test, compared to 2500k. If you are going to call that "bad scaling", than what is wrong with Intel's scaling such that they lose so much ground when going to the multicore test?

 

[AtenRa]

@Mallibu

The problem is that single thread performance is almost stagnant for the past 3-4 years. We only getting a roughly a 5% Single thread increase every year (Clock to Clock or for people that like to call it IPC) in legacy codded apps.

Have a look at iTunes, a single threaded application. From Core i7 Nehalem back from 2008 to Core i7 Ivy Bridge 2012. We don't even getting more than 15-20% higher performance Clock to Clock after four CPU generations.

Now have a look at a multithreaded application like x264 HD.

From dual cores to 6 cores/12 threads in just 2-3 years the performance has skyrocketed.

Yes i know we cant have an infinity number of threads but it seams that 12-16 threads are the sweet spot for desktop/workstation for the years to come.

High IPC is nice to have but alone it will not get you anywhere anymore. We need more threads and more apps codded to take advantage of SIMDs (ISAs).