Observations with an FX-8350

[Arkaign]

Wow, that is truly a guzzler we have here. Even if it isn't all the CPU and you can adjust settings, for the poor fellows who buy this, a 500W+ PSU and aftermarket cooling is practically a necessity.

True, and imagine if you pair it with a guzzler video card like a 480SC or 6990. Talking about serious PSU at that point.

 

[Markfw]

True, and imagine if you pair it with a guzzler video card like a 480SC or 6990. Talking about serious PSU at that point.

And did you bother to read my post ? I think more power drain research is in order. See this REPEAT:

Well, only somewhat related, but I have 2 opteron 6234's, 12 cores each, and 2 GTX 460's and with both cards @ 100% and all 24 cores @ 100%, they take 470 watt's total. It's the same cores, just 12 instead of 8 and 2 chips, except they are at 2.4 ghz. But there are 3 times as many cores !!! And with the 2 460's taking at least 200 watts. not bad total.

 

[BD231]

lol @ 200 watt 460's, not even close.

 

[AtenRa]

lol @ 200 watt 460's, not even close.

and 2 GTX 460's and with both cards @ 100% and all 24 cores @ 100%, they take 470 watt's total.

And with the 2 460's taking at least 200 watts. not bad total.

Each GTX460 will consume more than 100W at 100% load.

 

[Sable]

lol @ 200 watt 460's, not even close.

http://www.legitreviews.com/article/1360/12/

They reckon 160W for the 1GB version. 200W+ for two cards sounds right to me.
You really hate AMD don't you.

 

[guskline]

Idontcare:

My observations of powewr usage with stress being IBT
Stock 8350 with all power saving enabled
Idle: 80 Watts
Max 231 Watts

OC to 4.532 Ghz
Idle 129
Max 365

These do use power! Compared to Bulldozer @50 watts less at max but compared to Intel's Ivy Bridge NO Contest!

 

[BD231]

Each GTX460 will consume more than 100W at 100% load.

Thought he said 200 watts per card.

 

[BD231]

http://www.legitreviews.com/article/1360/12/

They reckon 160W for the 1GB version. 200W+ for two cards sounds right to me.
You really hate AMD don't you.

Nope, just bulldozer/vishera. It should be discontinued.

 

[inf64]

Vishera is selling well and according to real buyers (people who actually bought it) on newegg it gets great user reviews.

Also "BD" you can edit your post instead of spamming IDC's topic , we get notifications and I hoped it was OP who did more tests.Unfortunately it was spam.

 

[guskline]

I own both a Bulldozer 8150 and Vishera 8350 as well as 2 2500k rigs. BD231 has made his feelings well known. "It" should be discontinued.

For what? If AMD had a more powerful replacement it would be out. For the short time since the 8150 came out and limited R&D that AMD has, the PileDriver was a decent improvement over the Bulldozer.

Now let's get back on track.

Idontcare, anymore observations about the 8350?

 

[Arkaign]

And did you bother to read my post ? I think more power drain research is in order. See this REPEAT:

Well, only somewhat related, but I have 2 opteron 6234's, 12 cores each, and 2 GTX 460's and with both cards @ 100% and all 24 cores @ 100%, they take 470 watt's total. It's the same cores, just 12 instead of 8 and 2 chips, except they are at 2.4 ghz. But there are 3 times as many cores !!! And with the 2 460's taking at least 200 watts. not bad total.

I think that's the answer.

IDC's setup is simple, I don't think it's a fluke that he's seeing power spikes that vastly exceed the 125W rating.

There is a huge rampup in power consumption, and @ 4Ghz, there should be a a massive difference between 2.4Ghz power consumption and 4Ghz power consumption, enough even for a single CPU package to drink a considerable amount more power than two 2.4Ghz CPU packages.

 

[inf64]

@ Arkaign

Unless he got a chip that's different from every other sample in 8350 reviews out there, it should not be drawing more than maximum of 125W from the socket. I already posted a link on previous page from hardware.fr review and the measurements they did show a maximum of 116W for CPU alone.

 

[Arkaign]

@ Arkaign

Unless he got a chip that's different from every other sample in 8350 reviews out there, it should not be drawing more than maximum of 125W from the socket. I already posted a link on previous page from hardware.fr review and the measurements they did show a maximum of 116W for CPU alone.

Fritz Chess =! LinX.

To be fair, LinX is tougher than the worst-case regular usage scenario anyone could envision.

Even so, I still take IDC's numbers over almost any so-called 'tech' websites. His professionalism and attention to detail trump anything I've seen from the usual sites.

 

[inf64]

I still think there is something in his BIOS that's set to Auto and is causing such a power draw. You can see this is affecting his total system power draw which is also much higher than any user review(guskline above included) or website review out there.


BTW I don't doubt his actual measurement results,he did everything fine.

edit:
I have found a review by xbitlabs that measured total system power draw of 8xxx/6xxx CPUs under Linx. They got considerably lower total system power draw during full load Linx test on 8350 machine: 218W for full load on 8 cores and (again much lower) idle power of 67W. So full load delta between their result and IDC's is 63W or 281/218=1.29x or 29%. That's a huge delta to be within a margin of error or even be attributed to a platform difference (ie. motherboard,ram etc.). To make things more interesting, XBitlabs was using the exact same motherboard ASUS Crosshair V Formula.
Also they used the LinX 0.6.4-AVX utility which taxes the Vishera's FP units to the max producing the improved FLOPs numbers versus the standard(non-AVX) version.

 

[Lepton87]

Also they used the LinX 0.6.4-AVX utility which taxes the Vishera's FP units to the max producing the improved FLOPs numbers versus the standard(non-AVX) version.

Nope, to get max flops you need FMA on Vishera not AVX, AVX performance should be similar to non-AVX. Maybe IDC is using FMA optimized Linx (73Gflops) and they are not. How much GFLOPS do they get? the same number as IDC?

 

[Markfw]

lol @ 200 watt 460's, not even close.

OK all at full power, I get 470 watts. cpu's only I get 300 watt. So 170 watt for the GPU's sounds close. They are 768 meg 460.s, so less than 200 sounds right based on what I see below. And I know you can't exactly extrapolate, but if I had 12 cores @4.8 instead of 24 @ 2.4 then 200 for the CPU (not counting the 4 hard drives and memory/motherboard/cd drive) sounds right. So 125 at stock sounds right.

I am just saying that more information is required here.

 

[BD231]

OK all at full power, I get 470 watts. cpu's only I get 300 watt. So 170 watt for the GPU's sounds close. They are 768 meg 460.s, so less than 200 sounds right based on what I see below. And I know you can't exactly extrapolate, but if I had 12 cores @4.8 instead of 24 @ 2.4 then 200 for the CPU (not counting the 4 hard drives and memory/motherboard/cd drive) sounds right. So 125 at stock sounds right.

I am just saying that more information is required here.

I thought u said 200 for each card, but yea even 100 each sounds high for gp-gpu use only. If u were stressing the frame buffer in the slightest itd be a different story.

Vastly different numbers floating in this thread though, if a simple oc of 5 to 600 mhz gets u over 200 watts on vishera I dont see how 4ghz only eats 125.

 

[Idontcare]

I'm still feeling my way around this system, not fully up to speed yet.

I checked and yes, in the BIOS all the power-saving options are enabled.

Three things come to mind that go some distance to explain the elevated levels of power-consumption.

First, I realized today that the power numbers for my Intel CPU's are from optimized voltages and with superior cooling (so less static power losses occurring), whereas the values I am observing with my FX-8350 are all from the "stock" configuration.

Stock cooling is going to lead to higher temperatures and higher static power losses compared to my numbers with the Intel chips, and stock voltage may be way too much for the clockspeed. I haven't delved into undervolting yet to find the minimum voltage needed for 4GHz stock clockspeeds.

That said, stock is what you get with the retail box, and the retail box claims stock means a 125W TDP. And that shouldn't require non-stock cooling or intentional undervolting to stay within the TDP.

A note about all the extra hardware, it is the same as what I used when measuring the power-use with my 2600k and 3770k. Same PSU, same kill-a-watt, same ram, same SSD, same GTX460, same OS, and same LinX settings.

The only differences between all my tests are the motherboard itself and the CPU of course. Both motherboards are the top-of-the-line ROG boards from ASUS: the Maximus IV Extreme-Z for Intel chips and the Crosshair V Formula-Z for AMD.

I just don't see the Intel mobo using 100W less than the AMD mobo when running the exact same tests. That defies reason.

That said, I'm not nearly done with this, so it will be intriguing to see what turns up in the forthcoming rounds of testing.

For now, anyone wanting to generate apples-to-apples power-consumption numbers with their AMD chips will want to use the latest version of LinX (11.0.1.005) that can be downloaded with this link. (taken from here on XS)

I run it with problem size 43122 which ends up being ~14.2GB.

Is there a better stress tester for these Piledriver chips? LinX only stresses the FPU if I understand correctly, so a program that stresses the ALU's would probably be better at truly stress testing the chip I would think.

 

[blckgrffn]

Idontcare:

My observations of powewr usage with stress being IBT
Stock 8350 with all power saving enabled
Idle: 80 Watts
Max 231 Watts

OC to 4.532 Ghz
Idle 129
Max 365

These do use power! Compared to Bulldozer @50 watts less at max but compared to Intel's Ivy Bridge NO Contest!

That's what my 3.8Ghz 2600k pulls under AVX load - with a ~30W HP raid card

Also, that's what my A8-5600k pulls crunching with three cores and the onboard GPU!

 

[Pilum]

First, I realized today that the power numbers for my Intel CPU's are from optimized voltages and with superior cooling (so less static power losses occurring), whereas the values I am observing with my FX-8350 are all from the "stock" configuration.

Which doesn't matter that much, as I don't think I've heard of Intel chips @stock exceeding their TDP. BTW, the guys from Heise/c't have noticed the TDP excursion as well; for a 8350 they measured up to 168W power draw at the 12V supply lines between PSU and MB (no indication of the used software, so this might be a less powerhungry benchmark than Linpack). This includes the draw of the VRMs, but it's still clear that the 125W TDP is exceeded. Linky (in German).

That said, stock is what you get with the retail box, and the retail box claims stock means a 125W TDP. And that shouldn't require non-stock cooling or intentional undervolting to stay within the TDP.

Yep. But it seems that this only happens with some chips, many reviews from reputable sites didn't notice these power excursion. Which is actually worse than if the thing had a 150W TDP. AMD obviously doesn't have its production/QA under control, they don't know what kind of CPUs they are delivering to their customers. If you'd be an OEM, would you use a CPU which has a random TDP, requiring overdesigning cooling & PSU? Don't think so.

I sure hope they don't have this problem with their server CPUs; if they should, they'll be quickly removing themselves from the server market. I doubt there's any server vendor which would accept this kind of problem.

I just don't see the Intel mobo using 100W less than the AMD mobo when running the exact same tests. That defies reason.

Especially considering that MB components would start to die from overheating quite soon (what components could dissipate 100W on a MB?).

Is there a better stress tester for these Piledriver chips? LinX only stresses the FPU if I understand correctly, so a program that stresses the ALU's would probably be better at truly stress testing the chip I would think.

I don't think so. The BD architecture doesn't have that many INT resources, and they're likely held back by the decoder anyway. And it's not just about the execution backend; Linpack streams the data, so it also puts maximum stress on the cache hierarchy and IMC. You can't achieve that with integer code working with 64-bit accesses.

 

[Ancalagon44]

It would be awesome if you could run Gaussian98 on this 8350 and that 12 node cluster to compare. I mean, its gonna be faster of course, but how much?

 

[BrightCandle]

Just bare in mind TDP is about cooling not CPU power consumption. Its perfectly possible for a CPU to draw my power than its TDP up until its raised its temperature sufficiently that it needs to slow down. The cooling system must be able to at least remove tdp watts to be sufficient but a CPU can and will exceed that for periods of time. What a processor must do however is ensure that over seconds and minutes that it doesn't exceed TDP on average.

 

[guskline]

Idontcare: I ran the same test as you did on my rig with stock settings and only got 224Watts Max. When I ran OC at my settings for 4.53Ghz my max jumped to 283Watts. There might be a setting problem with your mb at stock.

 

[inf64]

There we have confirmation from guskline running the same linpack binaries. It has to be a some voltage setting problem in the BIOS(being much over recommended value). Or a dud CPU...

 

[Idontcare]

Idontcare: I ran the same test as you did on my rig with stock settings and only got 224Watts Max. When I ran OC at my settings for 4.53Ghz my max jumped to 283Watts. There might be a setting problem with your mb at stock.

Interesting. Just to ensure we are talking apples-to-apples, your power numbers are with a kill-a-watt or software measure program utility? And did you run with problem size 43122 using 8 threads? What GFlops did you get? And lastly, what is your CPU voltage at load?

With DDR3-1866 10-10-10-28-T2, I'm currently peaking at ~85 GFlops w/8 threads. Kill-a-watt reports pulling ~287W, and my CPU voltage is ~1.377V. Loaded CPU temps are ~58C with 19C ambient.