AMD's Bulldozer Architecture: Overclocking Efficiency Explored

[geekapproved]

Windows 8 will be able to park cores on BD, bringing it within 9w of i5-2500k at idle. Haven't seen the load tests done with Win8 though.

 

[rebel1280]

AMD used to be fast, then they took an arrow to the knee....
yup
an
arrow

 

[Guest]

Why does CPU-Z report the processor as an "FX-8130P" when Tom's Hardware clearly states they are testing an FX-8150?

And why does CPU-Z report the TDP to be 223 watts?

 

[ammaross]

Since people are too busy bashing AMD vs Intel to bother commenting on the article itself, I figured I'd venture forth....

When you run the system with a lower CPU voltage setting, performance differences compared to default are minimal. In fact, strangely enough, the system is actually faster.

One would think that if the thought passes of "strangely enough" that you would delve into what is going on, rather than simply stating utter bafflement and not pursuing the thought further. However, one would posit it is the FX CPU calculating its TDP using the voltage (which has been underclocked) as a factor, and thus rightly assuming it has more headroom to stay at higher boost frequencies for longer/more frequently. Shame boost duration/frequency wasn't explored to give insight on why it may be good to attempt undervolting stock CPUs in an attempt to gain performance and efficiency out of them in a similar fashion.

 

[alidan]

[citation][nom]mikenygmail[/nom]Wrong again, and this article is about a CPU, not a GPU - you don't seem to be aware of that. Actually you CAN credit AMD for keeping CPU's and graphics cards cheap because they don't try to squeeze every last drop of money out of consumers the way other companies do. AMD brings prices down for us all.It's ridiculous to show AMD CPUs on the front page covered with thermal paste, rust and generally looking dirty and unattractive. This is disappointing and you guys should be ashamed of yourselves![/citation]

AMD is shown us in the past that when they're ahead and have little to no threat, they jack their prices up to. That's why say can give them credit when they have a clear competitor, or the underdog, because they can't jack the prices of somewhere else can beat them or at least match their performance, but when you're the underdog you don't have the option of jacking your prices.

[citation][nom]Tomfreak[/nom]Agreed, Pentium 4 is also a high frequency CPU, it has low IPC, large cache(for ex. Prescott has 1M-2M L2 vs Athlon64 L2 512k-1M) same as bulldozer now.the first gen P4 willamate is suck(just like what Bulldozer now) then the intel fix the prob, Northwood do better, then they hit the Frequency TDP wall Prescott. I presume they will follow this same footstep in the future. Just dont know when they gonna hit the TDP wall again.Adding 2 more cores on the Existing Phenom II is like extra 25% die size. on a smaller 32nm thats just about equal the die size of Phenom II X6 45nm. Since a lot of multi-threaded benchmark shows 1100T is just as fast as intel's 1155 quad core, that extra 2 cores would have easily beat intel's quad core cpu. besides with 32nm die strink AMD could also clock Phenom II higher GHz too.as explained by me earlier Star architecture is base on K7 it is clearly superior to Core 2 duo which is base on P6 ones. If only they tweak the star architecture to a 4 issue like core 2 quad. They would easily have the IPC of Nehelem already. Add up the 2 extra cores + higher frequency on 32nm. They can complete with Intel's SB-E. IMO, if bulldozer is really design or server, they should have just limit it as server CPU like Intel's Itanium.[/citation]

I question if you happen to know, can a bulldozer core activist single core? I mean by turning off the second core/thread thing. If they figure a way to do that and only activate the second core when the applications threaded they should be able to match phenom II performance in single thread, and get higher than the I7 and multi thread. If you could intelligently choose which mode to go into I think it would have a large one up on the I series.

 

[ammaross]

[citation][nom]alidan[/nom]I question if you happen to know, can a bulldozer core activist single core? I mean by turning off the second core/thread thing. If they figure a way to do that and only activate the second core when the applications threaded they should be able to match phenom II performance in single thread, and get higher than the I7 and multi thread. If you could intelligently choose which mode to go into I think it would have a large one up on the I series.[/citation]
In case you didn't read, this is what the Windows 7 patch was supposed to do synthetically. Rather than randomly picking cores for threads, it would load 1 or 2 cores to cap, leaving the spare modules unutilized to allow the CPU to Turbo Boost properly. I'm sure part of this is scheduling core threads in separate modules and light threads as the "second" thread in a core. Putting one thread per core and essentially hobbling Turbo only makes sense if the fetch/decode stage is what is getting bottlenecked.

 

[kikireeki]

I think AMD should learn from Intel and team up with IBM.

 

[Guest]

AMD Bulldozer is still the only 8 core CPU available for us, and it has yet to be used to its full potential.
AMD APU's are an amazing product, combining CPU and GPU combined power that no company can match for anywhere near the same price.
AMD 7970 is the fastest single-GPU graphics card, and it will be many months before Nvidia has any chance of matching or beating it.

AMD is a true innovator that makes CPU's and GPU's cheaper for us all.

Huge Fanboy or PR Department plant?
8 Core CPU that performs worse then a 4 core CPU while consuming considerably more power.
APU's delivering the performance of a 3 year old Intel CPU with the power of a $50 GPU.
The 7970, which is only 15-30% faster then a 580, is not even available right now, for all we know Nvidia releases GTX 700 series parts to retailers around the same time you can actually get your hands on a 7970. And if the leaked reports are true, they will be 80-120% faster then a 580, trouncing easily the 7970.

Oh and AMD is releasing the 7970 at $550, they jacked up the price! But hey they are keeping GPUs cheaper for us all by increasing the cost of the 6970 it replaced by $180! And replacing faster performing Phenom II parts with slower FX chips at the same cost!

 

[dalethepcman]

I appreciate the effort that went into making this article, but once again Tom's has made an huge error against AMD. Testing with 2 x 4 GB DDR3-1333 is a disgrace.

The FX-8 series CPU's are supposed to be used with PC-1866. Running all of these test with memory that's rated at 75% of the recommended speed reflects poorly on the testing team.

I'm not saying that it would have made these cpu's shine and suddenly outperform everything, but at least it would have been testing them running their recommended configuration, instead of the minimum supported configuration.

This post describes everything I'm talking about with lots of links to other review sites testing with proper memory.

 

[g4114rd0]

 seriously Learning To Share, Of course!

+1 Blender inside bench.

 

[spp85]

Good to see everyone supports AMD. We all need a proper competition for the sake of consumers. My principle is that, if i can only afford a CPU or GPU of this performance and that product is available at AMD then i buys from AMD and neglects intel or nvidia.............

 

[tomfreak]

[citation][nom]ammaross[/nom]In case you didn't read, this is what the Windows 7 patch was supposed to do synthetically. Rather than randomly picking cores for threads, it would load 1 or 2 cores to cap, leaving the spare modules unutilized to allow the CPU to Turbo Boost properly. I'm sure part of this is scheduling core threads in separate modules and light threads as the "second" thread in a core. Putting one thread per core and essentially hobbling Turbo only makes sense if the fetch/decode stage is what is getting bottlenecked.[/citation]I am still not sure how bulldozer work, but on a 2 thread application, would it be much better to load the second thread on the 3rd core?(on second module) since the second core share the same units from the module on first core.

So which one is better, turbo boost or adding second thread to 3rd core(on a second module).

 

[ammaross]

[citation][nom]Tomfreak[/nom]I am still not sure how bulldozer work, but on a 2 thread application, would it be much better to load the second thread on the 3rd core?(on second module) since the second core share the same units from the module on first core. So which one is better, turbo boost or adding second thread to 3rd core(on a second module).[/citation]
Depends on how concurrent-friendly the fetch/decode stages for the thread is (likely not a problem, or else one or the other may have been replicated as well, just like the true heart of the cores). But you are right, for the most part it would perform better if intensive threads were placed on their own modules, interlaced with light threads for the module. My Win7 PC reports 81 processes with 1291 threads at the moment, so you must understand, we're not talking about just 2 or 3 threads in the CPU at a time. Likely, you'll only have 16 or so heavily active at any given time, but may only need a fraction of the CPU time, allowing the Win7 patch attempt (yes, it fails for now, until the second half of patch is released...) to stack all the active threads on one or two cores, allowing the other cores to shut down and the remaining to turbo and get the job done quicker (Hz is cycles per second. More cycles per second, faster the work is done, easy as that. Esp if a CPU has a per-cycle disadvantage vs Intel's architecture). It also helps if the OS has core-affinity for threads, preventing the need for the thread state and pre-fetched L1/L2 data from having to be shipped out to L3 and synced back to L1/L2 of another module to continue working. Currently, Win7 is only optimized to work with Intel CPUs by stacking heavy threads on the "real" core and trying to leave the HT "cores" free. You can see this in action (if you have a HT-enabled Intel CPU) by watching your task manager's Performance tab. You'll notice cores 1,3,(5,7) will have higher CPU utilization than 2,4,(6,8). AMD isn't as fortunately, since up til now, they've been full-fledged real cores, so no fancy footwork needed.

 

[esrever]

its the crappy transisters

 

[eddieroolz]

Excellent analysis of the design of Bulldozer. To me it seems like Bulldozer is pretty much hitting the clock ceiling for its design at 3.9GHz stock. So, that leaves clock raises out as one of the choices in raising performance over the next year.

Given that, I really wonder how AMD plans on delivering that raise.

 

[Homeboy2]

[citation][nom]deadon2[/nom]Fehh... did my build on a 990fx platform with a 955be CPU. Runs plenty fast, and I can upgrade the AM3+ in a year when AMD gets it right.Although I appreciate the work done on this article...Nothing to see here folks, move along...[/citation]

meh, by then Intel will have Haswell and be even further ahead

 

[mikenygmail]

AMD 4GB (1x4GB) DDR3-1600hz PC3-12800 Desktop RAM - $12 shipped today via OutletPC.
Wow, AMD gets into the RAM market and already there's an awesome deal for us!

The AMD Bulldozer is still the only 8 core CPU available for us, and it has yet to be used to its full potential. AMD APU's are an amazing product, combining CPU and GPU combined power that no company can match for anywhere near the same price. The AMD 7970 is the fastest single-GPU graphics card, and it will be many months before Nvidia has any chance of matching or beating it.

AMD really is a true innovator that makes CPU's and GPU's cheaper for us all.

 

[mikenygmail]

Originally posted by Anonymous: "Did AMD specifically ask Tom's and Anand to show their CPUs on the front page covered with thermal paste and generally looking dirty and unattractive, or did Intel specifically request it? I have a hard time believing that both of you came up with the same stupid idea.

In other news, your desktop PC has relatively low power requirements compared to every other appliance in you house, if you think you need to worry about a lousy few watts, you'd probably be better off switching to more efficient light bulbs and finding a lower wattage coffee pot first."

 

[gamerk316]

So which one is better, turbo boost or adding second thread to 3rd core(on a second module).

Neither, as doing one hurts performance because of the other. Hence why I don't expect much from the Windows patch, and actually expect some software to get worse. Scheduling is a red-herring thrown out to try and keep sales going, nothing more.

Also, I suspect Intel HTT would also benifit from this patch; would be interesting to see those results side by side...

 

[Guest]

Do we really need to click through 36 "next" button pages to read these articles? Can't you line them all up on longer pages? Is it an advertising thing? I love the detail you provide in these articles, but some hotels have terrible internet connections and I bail after a few pages of "loading"...

 

[iLLz]

Still baffles my mind why AMD went this route. Efficiency is terrible.

 

[kingnoobe]

@Freggo to get more bang for the buck. I've got my i7 920 oc'd to 3.5 (mind you that's not a huge oc or anything but still pretty decent). And I was also to lower the voltage a tad. It runs prefect no issues what so ever. And that's as an everyday gaming machine.

So why wouldn't you oc if there is very little risk (as long as you don't do anything to stupid and run the proper tests you should be fine).

 

[peevee]

[citation][nom]Tomfreak[/nom]What the HELL AMD engineers are thinking?![/citation]

Managers.

 

[peevee]

[citation][nom]mikenygmail[/nom]The AMD Bulldozer is still the only 8 core CPU available for us[/citation]

It is not truly 8-core, it uberhyperthreaded 4-core. In terms of transistor budget, the extra hardware used by AMD to make their 2-threaded "module" vs similar core is very little compared to what is shared - instruction decoder, caches, vector ALU etc. It is a good step beyond Intel's HT, low hanging fruit which had to be picked. The problem is that the core itself was not made right, it is way too slow compared to modern Intel's cores. And besides, sharing instruction decoder, without having something like Intel's L0 microcommand cache (which is probably patented) is crazy - it's the obvious bottleneck right at the beginning of the pipeline.

 

[peevee]

[citation][nom]gamerk316[/nom]Neither, as doing one hurts performance because of the other. Hence why I don't expect much from the Windows patch, and actually expect some software to get worse. Scheduling is a red-herring thrown out to try and keep sales going, nothing more.Also, I suspect Intel HTT would also benifit from this patch; would be interesting to see those results side by side...[/citation]

Actually, most of the real software either uses just one thread, or as many threads as available for most of its workload (if you created a complex multithreaded algorithm anyway, there is no point to limit it to just some fixed number of threads). So this scheduling trick works only for those hordes of sleeping threads which wake up only occasionally and do almost no work. In case only one worker thread working hard (single-threaded load), it makes sense to schedule them to the same core/module just to avoid waking up the other cores, which takes both time and energy.