Observations with an FX-8350

[Idontcare]

Maybe it's just that app you are using? Have you tried some other similar software?

I already have a license for this app, it cost me $100.

I use this app not because it is the fastest or the cheapest but because TMPGEnc is known for having the best IQ/bitrate when it comes to MPEG2 compression.

If the race was to find the best FPS out there, sans concern for IQ or bit-rate, then I'm sure there is no shortage of ways to get there (including transcoding with a GPU - very fast and totally crap results can be had )

There have got to be people in the transcoding industry who know the best optimizations that would tease the same (or better) IQ/bitrate out of the FX8350 at much better FPS. I just wish they'd get with the TMPGEnc people and give them the secrets

Is it 40%? I was looking at the last dot of the Q6600, (3.5?) from there it looks like 40FPS vs 30FPS, so on a clock-to-clock basis it's 33%, although it (Q6600) does seem to be scaling poorer as the clock increases.

Out of the box, of course, hands down the FX8350 performs better. This is not news, because AMD has been beating the Q6600 in out-of-the-box performance since higher-clocked Denebs (Q6600 is clocked pretty low), then Thubans, then Bulldozer. I therefore assumed Homeles' comment was in some context other than out-of-the-box, although I admittedly may have settled unfairly on an IPC / per-core context that was unwarrated, which led me to conclude it's still not a major ass-whooping architecturally against the Core2Quad, since it has 100% more cores, but only manages 33% more performance clock to clock.

The best fit equations are on the graph, I just used those for scaling to 4GHz, but I compared the actual stock results (red dots in the graph) for the stock comparison.

(10.223*4+5.3464)/(6.4247*4+7.1302) = 1.408 = 41% faster

46.15/22.24 = 2.074 = 107% faster

Like you say though, given 2x the cores, faster memory subsystem, SSD vs HDD, etc - is it really all that surprising the FX8350 is 41% faster than the Q6600 rig at 4GHz?

 

[xilience]

Well done!

 

[lau808]

i have a question u can answer as u run your tests, does the added cooling to the vrms allow for a higher oc?

 

[grimpr]

I'm willing to bet that TMPGenc is compiled with Intel tools and libraries, thats why it sucks on AMD, the classic cpuid check. Other mpg2 encoders like HCEnc dont do this.

http://www.intel.com/content/www/xr/en/benchmarks/resources-pegasys-tempgenc-xpress44-guide.html
http://www.bitburners.com/hc-encoder/

 

[guskline]

I'm still optimizing my 8350's OCs with the stock HSF. The good news to report there is that my previously observed Vmin's (minimum voltages needed at various clockspeeds for LinX stability) were way too high for one very specific reason - not enough air cooling was being provided to the motherboard's passive heatsinks that are atop the vrms and chipset chips.

I added two strategically located case fans to blow air more directly onto the surface of the mobo and now my 4GHz Vcore has dropped from ~1.32 to a much lower 1.244V. This enabled me to hit 4.3GHz with 1.347V, LinX stable.

I would have never expected the minimum Vcore could be lowered so dramatically on the basis of just getting the mobo cooled.

So my advice to anyone who is OC'ing these 8350's: before you settle on a given Vcore for your OC at any given clockspeed, adds ridiculous amounts of air cooling to the mobo just to see if it makes a materially significant difference to your OC like it does to mine :thumbsup:

Could you be more specific on the type of fan and the relative placement?

 

[Idontcare]

Could you be more specific on the type of fan and the relative placement?

I hacked this together with photoshop just because it is vastly clearer to see what I am talking about in the image sans the power cables and so forth present on my own:

If you can imagine the stock HSF sitting over the AM3+ socket, I wedged the 120mm fan between the side of the HSF itself and the edge of the mobo.

I did this because the vrm heatsink was just silly hot to the touch. Once I did it they are barely warm to the touch now.

Initially I suspected the extra fan was just helping augment the stock HSF and was helping to cool the CPU more than anything else, but the CPU temp only changed by 1°C with the additional fan.

The sizable change in Vmin for stable operation really comes from cooling down whatever is underneath that large passive heatsink on my mobo. (admittedly I assumed it is the VRMs since it effected Vcore so much, but it could be the extra cooling of the NB that is making the difference I suppose).

 

[Idontcare]

i have a question u can answer as u run your tests, does the added cooling to the vrms allow for a higher oc?

With the stock HSF I would certainly say it does. At 4.3GHz my temps peak at 61C now, before I was hitting above 70C and was not able to get a voltage that was LinX stable.

I don't want to over-sell my results though. This could be something that is truly only relevant to other crosshair V formula-z mobos, or it could be that it is only relevant to people like myself who are otherwise running with what may well turn out to be less-than-adequate cooling for the mobo.

That is why I temper my reporting of these results as "try adding a fan or two just to experiment and see if it nets you anything like it did me, if it does then you can invest the time at that point in finding a more aesthetic way to arrive at the same mobo cooling".

 

[Abwx]

I did this because the vrm heatsink was just silly hot to the touch. Once I did it they are barely warm to the touch now.

Initially I suspected the extra fan was just helping augment the stock HSF and was helping to cool the CPU more than anything else, but the CPU temp only changed by 1°C with the additional fan.

The sizable change in Vmin for stable operation really comes from cooling down whatever is underneath that large passive heatsink on my mobo. (admittedly I assumed it is the VRMs since it effected Vcore so much, but it could be the extra cooling of the NB that is making the difference I suppose).

The VRMs (in fact power mosfets) cooling is poorly implemented
in this MB.

Assuming 10W dissipation from the VRMs and a cooler that seems
to be about 5°C/W thermal resistance , the components will be
at Tamb + 50°C temperature , that is 70°C if the ambient temp
is 20°C , that is with non forced cooling.

You could think that once the CPU fan start to roll it will help
cool the thing since the air flux will move the air around the said
heatspreader but alas , things will only get worse as the air coming
from the CPU heatspreader will be at higher temperature than ambiant air
once the CPU is loaded , rejecting a 40°C hot air to cool the VRMs ;
hence the VRMS heatsink will be forced at an equivalent 40°C ambiant temp that will add to the 50°C due to the VRM dissipation , the whole
thing will reach untouchable temp at high loading.

A better solution would have been for Asus to give the VRMs and NB heatsinks separate 40x40mm fans.

 

[grimpr]

Just checked TMPGencs use of encoding, using a 2GB DV .avi file and transfer to mpg2 dvd-video, the encoder uses only about 45% of the 8350s resources, the mpg2 encoding engine is clearly not well multithreaded in contrast to mpg4-avc and tmpgencs licenced x264 encoder.

 

[AnandThenMan]

I have not read this entire thread, so apologies if this has been asked and answered. But why are you using the stock HSF? I chucked the boxed cooler on all of my rigs, the extra $30 it's about the best money you can spend on a system.

 

[AnandThenMan]

Just checked TMPGencs use of encoding, using a 2GB DV .avi file and transfer to mpg2 dvd-video, the encoder uses only about 45% of the 8350s resources, the mpg2 encoding engine is clearly not well multithreaded in contrast to mpg4-avc and tmpgencs licenced x264 encoder.

What happens if you run 2 instances at once, does the software allow it? I have not used TMPGenc in years, totally forgot about it actually.

 

[lau808]

he is playing with the stock hsf before he plays with his other toys

 

[Idontcare]

Gaussian is a computational chemistry program that I have used for many years and is an industry standard within the ab initio realm.

While I have more modern versions of it, I stick with using Gaussian98 ver A.7 because that is the one which I have the largest amount of my benchmarking data on. Newer version do run faster but in a way that is just like the tide that lifts all boats.

The windows version of the software is single-threaded, but you can run multiple instances in parallel which how one typically uses the application.

I can't find my Athlon K7 results but it was king of this application back in the day which is why I ended up with 24 K7's to create a cluster as noted in the OP of this thread.

I was expecting the FX-8350 to do rather well in this app, what I was not expecting was that it is basically on-par with having the same IPC as a 6yr old QX6700 Blows my mind.

But, thanks to the efficiency of CMT over that of HT, when it comes to multi-tasking single-threaded instances of this application the piledriver microarchitecture really shines here :thumbsup:

That's what we like to see, the ability to load up multiple instances of the software and just have real compute throughput push through on all your projects. Generally when operating this software you might have upwards of 20 or 30 different projects to run, so there is no shortage of opportunity to keep the cores fully loaded.

There is a multi-threaded version of the application but the price has markedly increased (now $5200) and regardless the price it just doesn't make sense to run through a batch job in series instead of farming out the jobs to run in parallel (avoiding the Amdahl tax by doing so).

 

[2is]

What exactly does this program do for one's life for it to cost $5200?

 

[grimpr]

Great job IDC, although that old version of Gaussian you're using is most propably running x87 code which modern AMD processors are very bad at (you can use AMDs excellent CodeXL to profile apps), your multitasking gauassian graph shows what this chip is capable of, its a multitasking throughput monster for the price.

 

[Jovec]

your multitasking gauassian graph shows what this chip is capable of, its a multitasking throughput monster for the price.

It's the difference between HT and CMT.

It shows that when working with INT heavy workloads, where the shared frontend isn't starving the cores, and where the slow cache isn't a big factor, that FX runs as an actual 8 core.

 

[Homeles]

It's the difference between HT and CMT.

It shows that when working with INT heavy workloads, where the shared frontend isn't starving the cores, and where the slow cache isn't a big factor, that FX runs as an actual 8 core.

An actual 8 core would still be quite a bit faster. But yes, this does show the CMT approach besting the SMT approach in this one scenario.

 

[Jovec]

An actual 8 core would still be quite a bit faster. But yes, this does show the CMT approach besting the SMT approach in this one scenario.

A slow 8 core is still an 8 core CPU. The definition of a true core can be argued. While the net effect of CMT and HT speed up performance, they are different things. Even better would be a CPU with both CMT and HT - 4 module / 8 "core" / 16 thread CPU.

 

[Homeles]

A slow 8 core is still an 8 core CPU.

What does that have to do with what I said?

While the net effect of CMT and HT speed up performance, they are different things.

I know this. That's kind of the point of me claiming that CMT approach bests the SMT approach here.

 

[SPBHM]

What does that have to do with what I said?

I know this. That's kind of the point of me claiming that CMT approach bests the SMT approach here.

it's not that simple, you need to consider Die size and power usage I think,

but the CPUs are so different, that, it's difficult to compare I think,

Intel is not the only one using SMT

 

[2is]

it's not that simple, you need to consider Die size and power usage I think,

but the CPUs are so different, that, it's difficult to compare I think,

Intel is not the only one using SMT

I'm not sure what this has to do with the post you quoted...

All he said was one approach is better than the other in THIS case... It seems simple enough to me if you don't take it out of context.

 

[Homeles]

it's not that simple, you need to consider Die size and power usage I think,

but the CPUs are so different, that, it's difficult to compare I think,

Intel is not the only one using SMT

Yes, I know all of this as well. IBM, for instance, uses SMT.

But you're going off on a tangent. Of course there are tradeoffs — if adding another core was superior in every way to SMT and CMT, those approaches wouldn't exist. But that has nothing to do with Jovec's claim and my rebuttal.

Jovec claimed that an 8 core FX runs as a true 8 core in that application. If you had an 8 core Phenom II or 8 core Ivy Bridge with SMT disabled, you'd find that both would outperform the FX by a significant margin.

Therefore the FX is not running as "a true 8 core." It's running better than a hyperthreaded quad core, but not at the level of an 8 core processor.

 

[dkm777]

I just wanted to chime in about VRM cooling affecting overclocks. Even though I don't have an FX (undecided), but when I lowered the fans on my Cogage Arrow so that some air would flow over the large VRM heatsink I could reduce both the CPU-NB voltage (which reduced temps dramatically) and core voltage. I left my PC folding for the day and if I don't find that it has crashed then I think we can confirm the theory that certain Asus AMD motherboards need better cooling for the VRMs (I have a Sabertooth 990FX).

 

[AtenRa]

Therefore the FX is not running as "a true 8 core." It's running better than a hyperthreaded quad core, but not at the level of an 8 core processor.

You have to compare it against a 8 (legacy) core based Bulldozer to see how it performs. Most of the time, CMT performs at 80-90% of a legacy Multi Processor(MP) design.

 

[Jovec]

Jovec claimed that an 8 core FX runs as a true 8 core in that application. If you had an 8 core Phenom II or 8 core Ivy Bridge with SMT disabled, you'd find that both would outperform the FX by a significant margin.

Therefore the FX is not running as "a true 8 core." It's running better than a hyperthreaded quad core, but not at the level of an 8 core processor.

It's hard for us to define what it means to run as a "true 8 core." Although the module design by definition means that the FX line cannot operate as a "true 8 core" for all workloads, for specific ones it can. It's not as fast as it could be, but then even that 8 core Phenom II or 8c/8t Ivy Bridge could be faster but for the tradeoffs AMD and Intel make - to the cache, front end, execution units, back end, die size, power consumption, etc.

AMD seemed to miscalculate the bottlenecks in the Bulldozer and Vishera design. Steamroller really looks to address these. Note that it doesn't necessarily mean besting Intel (unlikely IMHO), but rather just maximizing the FX design.

You have to compare it against a 8 (legacy) core based Bulldozer to see how it performs. Most of the time, CMT performs at 80-90% of a legacy Multi Processor(MP) design.

Perhaps you meant to say compare it Phenom or Thuban (i.e. a non-CMT CPU)?

As consumers apples-to-apples comparisons matter less than real-world performance, the question is how much of that 80-90% is due to the FXs higher clocks? Presumably someone has compared a Phenom II quad to a Vishera 2M at the same clockspeed and turbos off.

Edit. Here is one from Tom's showing 2 benchmarks at 4GHz. Far from thorough, but the single-threaded bench has Thuban beating Bulldozer and Vishera and has 6c/6t Thuban beating 4m/8t Vishera in the multithreaded one. I'm sure there are more (and more thorough) clock for clock benches out there. We can extrapolate an 8c Thuban at 1m 48s compared to 4m/8C Vishera at 2m 35s