Observations with an FX-8350

[SammichPG]

It would clock infinitely higher, have 0W power draw and be infinitely slower.

(It only has 4 modules. )

Derp, meant cores.

 

[sequoia464]

Slowspyder has a good point. AMD's 4 stick memory in my machine required a bump in the Dimm voltage to 1.53 from 1.5 to assure enough juice for stability. I also agree that 1T timing is preferred but 2T might be needed. Finally, despited the claim that PileDriver supports 1866 memory I believe the fine print reads that it supports 2 sticks at that rated speed not 4 sticks.

Right now I have 2 sticks of Samsung 1600 memory @1866 1T (9-9-9-24-45).

Just ordered a couple more, be interesting to see how things work with 4 of the Samsung - I know that my GSkill and Kingston would not run @ 1T even with just 2 sticks.

 

[frozentundra123456]

It seems that the 8320/8350 are vastly improved over BD, and a fair option for anyone with an existing mobo that is compatible, but they remain of questionable value compared to Intel options. I believe I'd personally rather have a 2500K or 3570K, to be honest.

What should a fair market price be for these, considering the uneven performance and relatively large power draw / cooling requirement? Maybe 8320 @ $109, 8350 @ $139? I bet AMD would move plenty at those prices.

Price would make some difference, especially in a mid-range situation. However, I think more critical is what your intended use is of the chip. For a higher end gaming rig I would not consider AMD at any price. The performance trade off is just too much in gaming. If otoh you use some of the heavily multithreaded apps that PD is best at, AMD could well make sense.

In the mid range, the FX6300 already competes fairly well with the i3, but obviously if the price of an 8 core came down to that range it would be a better choice, especially for productivity. For primarily gaming, I am not sure the other 2 cores would be of much benefit.

I guess what I am saying is that if you want the absolute best gaming performance, you would go intel regardless of the cost. If you are primarily interested in productivity apps or are willing to accept less well rounded gaming performance, AMD makes sense at the current prices. And even though it may not cost that much extra money, the higher power usage does still bother me.

 

[Idontcare]

IDC, do you still have your 2700K running? I know your Ivy died (RIP buddy!). I know this wouldn't be scientific with different components, but I would be curious to see how eight threads scales on Intel's Hypterthreading. I know the overall score will likely be higher on a 2700K, but I'm more interested in the scaling, just for comparrison and for the sake of curiosity.

Yeah I have the 2600K (easily clocked to 2700K specs), and I actually do have a 3770k (again!)...and I think you make an excellent suggestion. I will setup the sister rigs and run some thread-scaling tests in parallel.

IDC: the "8 core" PileDrive really is closer to 6.7 true cores due to the sharing. If only the software would be tailored for this sharing.

True, the CMT tax was expected to come in at 0.8x, and 8 * 0.8 = 6.4, so we are seeing this reality born out in the benches.

Steamroller is supposed to address much of this CMT tax, it will be quite interesting to see how much of it they can eliminate.

My Intel 2500ks don't have hyperthreading but what is added to the 2600/2700/3770 series to allow hyperthreading?

They are same chips, hyperthreading is simply disabled at the factory (along with 2MB of the 8MB L3$...your 2500k has 6MB whereas my 2600k and 3770k have 8MB)

IDC, Give this nice "simulator" a try.

[URL="http://www.dimkovic.com/node/1"]http://www.dimkovic.com/node/1[/URL]

Cool! I love new benches, and science-related ones are the pinnacle of awesome :thumbsup:

One question - I downloaded the SpikeFun_latest application and installed it, which created a shortcut called "Benchmark...". I'm running the benchmark, it has been running for a long time now, when does it end? (ever?)

I'd be curious to see how that cpu behaves with 4 modules disabled, I guess that it would clock higher, consume less power and be generally slower while running highly threaded applications.

Derp, meant cores.

I can't disable individual cores, the BIOS only allows the disabling of entire modules.

But I can "park" individual cores, which essentially shuts them off as far as the OS is concerned. They won't consume dynamic power, but of course they are still consuming static power from leakage when the Module is powered up because power-gating is a module-level granularity, not a core-level granularity.

Right now I have 2 sticks of Samsung 1600 memory @1866 1T (9-9-9-24-45).

Just ordered a couple more, be interesting to see how things work with 4 of the Samsung - I know that my GSkill and Kingston would not run @ 1T even with just 2 sticks.

The thing is I know these 4 GSkill sticks will do 1T at these timings and volts because they do it on my Intel rig.

So it can't be the ram that is the issue in preventing 1T operation for my specific ram sticks, can it?

 

[AtenRa]

Thanks for the data :thumbsup:

Do you happen to know anything official regarding TJmax for the FX-8350? Know anyone that would know?

If JFAMD was still at AMD then I would ask him...but he's moved on to greener pastures. In fact every single employee of AMD that I use to know have since moved on from AMD.

I dont know the Official TJmax for the BD/PD (it was supposed to be 60c for BD) and as you and many more in the industry have lost 99% of the people i knew in AMD. I still have a single contact, ill try to see if i can find anything and report it here.

 

[Ajay]

I dont know the Official TJmax for the BD/PD (it was supposed to be 60c for BD) and as you and many more in the industry have lost 99% of the people i knew in AMD. I still have a single contact, ill try to see if i can find anything and report it here.

60C! I guess AMD uses a much different definition than Intel. Sorry about losing contact with folks at AMD. Same happened to me with Network IHVs in 2002-2003. Suddenly they just all up and vanished.

 

[AtenRa]

60C! I guess AMD uses a much different definition than Intel. Sorry about losing contact with folks at AMD. Same happened to me with Network IHVs in 2002-2003. Suddenly they just all up and vanished.

Yea, a lot have changed in AMD the last year. They have closed the PR and their main Distributor in Greece(and other EU countries) since Summer.

 

[thilanliyan]

But I can "park" individual cores, which essentially shuts them off as far as the OS is concerned. They won't consume dynamic power, but of course they are still consuming static power from leakage when the Module is powered up because power-gating is a module-level granularity, not a core-level granularity.

How are you able to "park" individual cores?

 

[ShintaiDK]

How are you able to "park" individual cores?

Windows affirmity I assume.

 

[Idontcare]

How are you able to "park" individual cores?

There are utilities that can set the parking for cores, this is one for example.

It is handy when you want to leave HT enabled but you basically want to force the windows scheduler to park the virtual cores and only assign threads to the physical cores.

I haven't tried it but presumably there is a way to make that work with CMT. This program for example.

 

[Idontcare]

Ran some Fritz Chess benchmarks today.

13,040 kilo-nodes/sec represents a 5.6x scaling over the single-threaded performance of 2,338 kilo-nodes/sec.

The lower thread scaling was due to a higher CMT tax with this benchmark:

^ notice the thread scaling is nearly perfect for 2 and 4 threads when loading just the modules (M1/M2 and M1-M4 tests) but the penalty for loading both cores within a module is 0.77x scaling.

Fritz Chess scales well with ram bandwidth so don't be surprised if you get markely worse, or better, results than me.

This is one test that I think will best highlight the impact of changing the NB frequency too.

 

[grimpr]

Cool! I love new benches, and science-related ones are the pinnacle of awesome :thumbsup:

One question - I downloaded the SpikeFun_latest application and installed it, which created a shortcut called "Benchmark...". I'm running the benchmark, it has been running for a long time now, when does it end? (ever?)

Check out "benchmark.pdf" in Spikefuns directory for a explanation of Spikefuns modes. Yes it runs forever but writes a log file named "SpikeBench.txt" in its home directory with the cpu information, program version and performance numbers.

 

[C.C.]

I am not sure if it is the same with BD/PD, but I could not get my 1T rated high end Corsair Dominators to run @ 1T on my old 1055t/890GPA setup..From what I was told at the time, it didn't make any difference to run them @ 2T vs 1T..

 

[SPBHM]

could you try two super PI running at the same time on a single module compared to using one per module?

 

[Idontcare]

I am not sure if it is the same with BD/PD, but I could not get my 1T rated high end Corsair Dominators to run @ 1T on my old 1055t/890GPA setup..From what I was told at the time, it didn't make any difference to run them @ 2T vs 1T..

I'm suspecting this is the case for me too. I've only done some preliminary testing but here is what I have so far:

4x4GB DDR3-1866 10-10-10-28-1T @1.5V on i7-2600k/3770k = OK

4x4GB DDR3-1866 10-10-10-28-2T @1.5V on FX-8350 = OK

4x4GB DDR3-1866 10-10-10-28-2T @1.4V on FX-8350 = OK (haven't tested lower Vdimm yet)

4x4GB DDR3-1866 10-10-10-28-1T @1.65V on FX-8350 = Fails

Someone mentioned increasing the voltage of the IMC - I have not found this option yet in the BIOS. What does that mean, the voltage of the IMC? Is that the "CPU/NB"?

could you try two super PI running at the same time on a single module compared to using one per module?

Will do Thanks for the request/suggestion :thumbsup:

 

[parvadomus]

What about trying with some games? If you have any ofc.

 

[jvroig]

Someone mentioned increasing the voltage of the IMC - I have not found this option yet in the BIOS. What does that mean, the voltage of the IMC? Is that the "CPU/NB"?

Yes, CPU-NB V is for the IMC.

 

[Idontcare]

Yes, CPU-NB V is for the IMC.

Thanks, I will play with that then. I wish ASUS would unify their MemTweakIt across their entire ROG portfolio, it works great for tweaking/optimizing memory timings on the fly from within windows but it apparantely has been coded to only with with the Intel-based ROG mobos

 

[Idontcare]

could you try two super PI running at the same time on a single module compared to using one per module?

Here are the results!

This is for just one core loaded, affinity locked:

^ 4GHz, turbo-core disabled, DDR3-1866 10-10-10-28-CR2

^ Two cores loaded, on separate modules, times are basically the same as when running one core alone. I expected bus contention for the ram to effect the scores more, but apparently it isn't an issue.

^ 2 cores loaded onto the same module, resource contention becomes apparent.

^ all the times were averaged for both cores within each test group, converted to seconds and plotted, y-axis is log-plot which compresses the appearance of the computation times accordingly.

The CMT "tax" is observed to be roughly 10% with Super PI. Much better than observed with the other programs I have benched to date.

 

[Rvenger]

Isn't SuperPI irrelevant for AMD cpus since its not optimized for x87?

 

[Zodiark1593]

Isn't SuperPI irrelevant for AMD cpus since its not optimized for x87?

I'm somewhat skeptical as well, considering my 2.53 GHZ i5-460M scored 16 seconds on SuperPi 1M and 39 seconds on 2M. I haven't used an AMD cpu myself, but I really hope that this isn't representative of it's single thread performance. :hmm:

 

[Idontcare]

Isn't SuperPI irrelevant for AMD cpus since its not optimized for x87?

I'm somewhat skeptical as well, considering my 2.53 GHZ i5-460M scored 16 seconds on SuperPi 1M and 39 seconds on 2M. I haven't used an AMD cpu myself, but I really hope that this isn't representative of it's single thread performance. :hmm:

The benchmark is relevant for making comparisons regarding those portions of the ISA which are used by the program in question. That is true of any benchmark.

Obviously when it comes to performance with applications that critically depend on the instructions for which superPI depends, if performance is your metric of concern then you are going to want to go with an Intel CPU that is clearly more optimized for that particular suite of instructions.

However, I think a more generalized version of the same question is in order: "Isn't SuperPI irrelevant?"

This I would probably find myself agreeing with. I don't know too many non-legacy applications for which an individual would find themselves flush with enough cash as to afford a new $500-$1000 desktop and yet they don't have enough cash to afford to update their program to a more recent version that has likely been compiled to take advantage of the more recent microarchitectural and ISA enhancements that either AMD or Intel bring to the table in their recent CPUs.

I view superPI as an tool in my diagnostic toolbox. Yet another way to interrogate the microarchitecture of bulldozer and sandy bridge to tease out the strengths, weaknesses, and trade-offs inherent to both.

For example, I fully expect superPI to be invaluable as a quick feedback benchmark in my efforts to co-optimize the CPU/NB clockspeed and voltage, alongside the efforts to do the same with the ram bandwidth versus subtimings for latency.

But does superPI tell me anything about how bulldozer is going to perform in a real-world application of interest to me like TMPGEnc? I very much doubt it will.

 

[guskline]

IDC: Bulldozer or PileDriver? They are similar but different. In your post above you refer to "Bulldozer vs Sandy Bridge...". BTW, nice explanation of the CMT tax. "Sharing" resources via a module concept has its cost.

 

[Idontcare]

IDC: Bulldozer or PileDriver? They are similar but different. In your post above you refer to "Bulldozer vs Sandy Bridge...". BTW, nice explanation of the CMT tax. "Sharing" resources via a module concept has its cost.

I refer to "bulldozer" as the microarchitecture family in this sense, same as folks refer to "netburst" as a microarchitecture family within which northwood and prescott microarchitectures were members.

"Bulldozer" could be "bulldozer" could be "piledriver" could be "steamroller" in this context.

Consider that technically none of us are really talking about bulldozer or piledriver, if anything we are talking about zambezi and vishera.

Same with "Sandy Bridge"...in this context "Sandy Bridge" could be either 32nm Sandy or 22nm Ivy products, they both have a common microarchitecture and the features only differ in minor details across the members of the family.

Intel generalizes this by referring to them as "generations", which makes sense in the context of the "family" vernacular used to capture successive iterations of the main microarchitecture trunk.

So consider my "bulldozer vs Sandy" comments as being akin to what people mean when they nowadays refer to "netburst vs K8" in the generalized family sense.

I would be surprised if my results here regarding CMT tax would be any different if I ran the same analysis on a bulldozer 8150 chip (just as one would not expect the performance hit from HT to differ between Sandy Bridge and Ivy Bridge). These aspects of bulldozer are the same in piledriver and won't stand a chance of changing until steamroller is released.

In this context, do my interchangeable references to bulldozer <-> piledriver and sandy <-> ivy make any more sense?

Regardless, your question alone makes a very important point - it doesn't matter what I am thinking when I write what I write, if it isn't clear from the outset to the reader then there is no value added by me using such short-hand references. I will endeavor to keep it more clear going forward, piledriver == piledriver, Sandy == Sandy, etc. No need to have me making it even more murky considering it is very easy for me to keep it clear from the outset

I thank you for bringing this to my attention, no doubt I confused more people than just you, so thanks for speaking up and making me aware of it. :thumbsup:

das ist klar? ja, gute

 

[hdfxst]

your super pi times look similar to my 955 when i first got it,i found out it wouldn't ramp up right away and stay at full speed for the entire run.I used a program to set the load threshold and my times at 1M went from 23-24 to 18.3 seconds.By comparison my 8M time is 3m 52.08.I don't think super pi loads the cpu enough to keep it at full speed