Post your AMD FX-8350 Overclock

[guskline]

jvroig: My BIOS does allow for seperate cores (modules) to be enabled or disabled. I guess my question would be why? Pehaps to test seperate cores or modules. My simplistic thinking is if I only want to run 4 cores (2 modules) buy the 4300 Vishera or 6 cores (3 modules) buy a 6300 Vishera.

 

[inf64]

I think jvroig is thinking of resource sharing penalties when it comes to modules and ST/MT workloads. It's useful to test it out and see how performance scales with threads evenly distributed and threads put on the separate modules.

 

[guskline]

I think jvroig is thinking of resource sharing penalties when it comes to modules and ST/MT workloads. It's useful to test it out and see how performance scales with threads evenly distributed and threads put on the separate modules.

Sorry: I just noticed that! That type of testing is way above my "pay scale!"

 

[jvroig]

jvroig: My BIOS does allow for seperate cores (modules) to be enabled or disabled. I guess my question would be why? Pehaps to test seperate cores or mudules. My simplistic thinking is if I only want to run 4 cores(2 modules) buy the 4100 Vishera!

Just to clarify, do you mean individual cores and not necessarily an entire module?

As for the why, just curious. Aside from inf64's feedback (which is mostly spot on, it would indeed be something to try, given how cheap FX processors are now), in my X4 and X6 rigs, they have different profiles in the BIOS depending on their duties for the period (day, week, or month, as the case may be). For example, I am currently using my X4 for nothing more than basic office tasks and non-intensive software development. Given the need for little power, I am running it on the lowest power profile designed to save the most power while maintaining reasonable everyday performance: 1.092V, 4 cores, 2.6Ghz. (it used to be 1.092V, 2 cores, 2.7Ghz, but a recent bios update somehow removed the option to disable cores, very frustrating).

My X6 also has a similar low-power profile: 1.152v, 2 cores, 3.1Ghz.

When either (or both) of them need to be involved in some rather CPU-heavy multi-threaded tasks, I just go into the bios and load the max performance profile, all cores enabled at max stable OC, then let them do whatever work they are scheduled to do. When it's done, I stick them back into the low power profile.

 

[guskline]

I'm not at home right now, but I believe my BIOS on my Sabertooth FX990 allows module disabling, not seperate cores. Probably due to the Bulldozer/Piledriver architecture. I'll try to test it tonight and report back tomorrow. BTW jvroig, what mb are you presently using for AMD chips?

 

[Idontcare]

Idontcare: Here are my AIDA 64 sensor specs for my 8350 running at 4.53Ghz (21 x 215).

Is that the preferred method of overclocking AMD black cpu's? Pushing the HT speed and not increasing the multiplier?

 

[guskline]

Is that the preferred method of overclocking AMD black cpu's? Pushing the HT speed and not increasing the multiplier?

From what I read probably not. They prefer higher multiplier. However, for me this required more voltage. I kept the multiplier lower and found a fsb of 215 to run solid. Everything I read about the 8320/8350 appears to have OCs all over the lot. What's great about your selection of your Asus MB is in it's BIOS you can let it do the automatic OC. I can too but I have to do more manual tweeking. When I allowed my mb BIOS to automatically OC with all stock settings it set the multiplier at 20 and the fsb at 215! So, I thought, HMMM! Let's keep the fsb at 215 and raise the multiplier. I had to do a lot of tweeking but arrived at my present settings of 21 x 215. Your mb will probably do better at auto tweeking. Would not be suprised to see a higher OC. My key for stability is the ability to run with acceptable temps and pass 20 x IBT; 1 hr AMD OC stability test with all components stressed and 1 hr OCCT stress. This setup does it ROCK SOLID.

 

[Idontcare]

From what I read probably not. They prefer higher multiplier. However, for me this required more voltage. I kept the multiplier lower and found a fsb of 215 to run solid. Everything I read about the 8320/8350 appears to have OCs all over the lot. What's great about your selection of your Asus MB is in it's BIOS you can let it do the automatic OC. I can too but I have to do more manual tweeking. When I allowed my mb BIOS to automatically OC with all stock settings it set the multiplier at 20 and the fsb at 215! So, I thought, HMMM! Let's keep the fsb at 215 and raise the multiplier. I had to do a lot of tweeking but arrived at my present settings of 21 x 215. Your mb will probably do better at auto tweeking. Would not be suprised to see a higher OC. My key for stability is the ability to run with acceptable temps and pass 20 x IBT; 1 hr AMD OC stability test with all components stressed and 1 hr OCCT stress. This setup does it ROCK SOLID.

I see.

And what about Vdram settings, with Intel you are well advised to stick with 1.5V memory or you'll burn up your IMC. Where is that threshold for my 8350?

One thing I was quite surprised to see was how poor the performance is in LinX. I'm wondering if I have something set wrong.

At stock 4GHz clocks running 8 threads and problem size 43122 (~14.2GB), my 8350 is turning in a paltry 35GFlops. At the same settings, same OS and so forth, but running only 4 threads my 3770k turns in ~114GFlops at 4GHz.

Is the LinX gap supposed to be that large? Am I missing some key Win7 update or something?

I understand LinX performance is not indicative of general performance because LinX is a specialized synthetic benchmark, but I am worried that my 8350's performance is low because I have something setup incorrectly. I expected a score that was at least 2x-3x the current value, up around 80 or 90 GFlops.

 

[jvroig]

I understand LinX performance is not indicative of general performance because LinX is a specialized synthetic benchmark, but I am worried that my 8350's performance is low because I have something setup incorrectly. I expected a score that was at least 2x-3x the current value, up around 80 or 90 GFlops.

I believe your expectation of higher performance is correct, since my X6 is 60Gflops at stock.

Also, although I may be remembering this incorrectly, shouldn't AVX support mean an even higher performance? I think I remember reading here that AVX on/off results in widely different GFLOPS in LinX for Intel CPUs (am I truly whacked out here and remembering something else completely? How/why is AVX support toggled?) Since Vishera has AVX (unlike Thuban), I thought the scores would no doubt shoot up.

 

[blckgrffn]

I believe your expectation of higher performance is correct, since my X6 is 60Gflops at stock.

Also, although I may be remembering this incorrectly, shouldn't AVX support mean an even higher performance? I think I remember reading here that AVX on/off results in widely different GFLOPS in LinX for Intel CPUs (am I truly whacked out here and remembering something else completely? How/why is AVX support toggled?) Since Vishera has AVX (unlike Thuban), I thought the scores would no doubt shoot up.

Intel's AVX implementation is much wider @ 256 bit and can be done in one cycle, AMD has to fuse two 128 bit operations (?) and it takes multiple cycles (?) - resulting in Intel-like AVX workloads running just as fast on AMD with or without AMDs AVX implementation being used.

In DC projects like PrimeGrid, AVX is not enabled for AMD CPUs because there is little point optimizing for it, so far as I understand it.

This is one reason I think an AMD console CPU would be really interesting, as we'd some actual optimization for what AMD brings to the table.

 

[Solomutt]

I see.

And what about Vdram settings, with Intel you are well advised to stick with 1.5V memory or you'll burn up your IMC. Where is that threshold for my 8350?

One thing I was quite surprised to see was how poor the performance is in LinX. I'm wondering if I have something set wrong.

At stock 4GHz clocks running 8 threads and problem size 43122 (~14.2GB), my 8350 is turning in a paltry 35GFlops. At the same settings, same OS and so forth, but running only 4 threads my 3770k turns in ~114GFlops at 4GHz.

Is the LinX gap supposed to be that large? Am I missing some key Win7 update or something?

I understand LinX performance is not indicative of general performance because LinX is a specialized synthetic benchmark, but I am worried that my 8350's performance is low because I have something setup incorrectly. I expected a score that was at least 2x-3x the current value, up around 80 or 90 GFlops.

Not at home, atm, but the things I recall in playing w/8320:

CPU voltage:
1.31
CPU-NB
1.25 (This seems to be the upper end if you don't want to risk overheat/electro-migration over time)
FSB:
220
Multiplier:
19.5 or 20

{Resulting Speed was 3.32-ish}

IBT run on high setting resulted in 51 deg celsius (PEAK) on a hyper 212+ w/Artic 5 paste

Idle temps are ~33C in 70 deg F room, Antec 300 case w/broken top fan

Typical load temps are <42 deg celsius

CPU Fan control is set at 60% for >20 C, and 95% for>42C

CPU and CPU-NB LLC are both at 1 level from highest.
CPU max current:130%
CPU-NB max current is at 120%

CPU-NB 1nd HT are both ~2350, synched manually.

CPU VRM cpeed setting was 1 from fastest, and the speed setting was 420 (300-500 range)

I used T. Probe instead of current for regulation in the relevant setting.
c1E, Turbo Core, and other upscaling / down-clocking modes diabled.

Secure VM is enabled, as it is not relevant to OC enough to lose the hardware acceleration of this function when using VM.

I think my temps were actually lower by 1C or 2, but I want to be conservative.

I am considering dropping 1C per module, to reduce temps under load, and give each core more cache and a 256 bit floating point instead of 128. It will also reduce throughput latency, and NB load, is my hypothesis.

Final numbers for my optimal settings will come later.

 

[Solomutt]

Intel's AVX implementation is much wider @ 256 bit and can be done in one cycle, AMD has to fuse two 128 bit operations (?) and it takes multiple cycles (?) - resulting in Intel-like AVX workloads running just as fast on AMD with or without AMDs AVX implementation being used.

In DC projects like PrimeGrid, AVX is not enabled for AMD CPUs because there is little point optimizing for it, so far as I understand it.

This is one reason I think an AMD console CPU would be really interesting, as we'd some actual optimization for what AMD brings to the table.

May be correct, if you are not disabling modules, and/or the MS scheduler fixes are not implemented, and the AVX load ends up trying to traverse across eight "expected" cores.

In non- benchmarking scenarios, I find that my 8320 is just plain faster than a higher clocked {3.7 vesus 3.3 on 8320} 960T unlocked to six cores, even on the same hardware.

My machine runs a lot of background processes, and I multi-task a good bit as the net admin of my {8 total PC, multi-switch and router} home network. I normally 1 or 2 multi-gig ISO torrents running, and typically use games + voice chat, Voice chat server, and multiple firefox windows/tabs open for researching items for friends/clients.

I am sure that Intel will always have the fastest overall platform, at a cost, and will even have superior benhchmarks in the same price range. I don't need e-peen points, and I use my PC for just about anything BUT benches.

 

[blckgrffn]

May be correct, if you are not disabling modules, and/or the MS scheduler fixes are not implemented, and the AVX load ends up trying to traverse across eight "expected" cores.

It's not a scheduler thing, the FPU can only handle 128-bit AVX. When Intel sells a couple orders of magnitudes more worth of chips that can do 256-bit AVX, you aren't going to see much time and effort spent extracting the performance from the AMD setup and its relatively tiny install base.

 

[NTMBK]

It's not a scheduler thing, the FPU can only handle 128-bit AVX. When Intel sells a couple orders of magnitudes more worth of chips that can do 256-bit AVX, you aren't going to see much time and effort spent extracting the performance from the AMD setup and its relatively tiny install base.

As far as I was aware, BD executes a 256bit op in a single cycle. However the bottleneck comes from that 256bit FPU being shared between two cores.

 

[sequoia464]

Is that the preferred method of overclocking AMD black cpu's? Pushing the HT speed and not increasing the multiplier?

Just a rookie at this, but if I understand correctly HT is already set at 2600 stock - it is usually left alone or close to stock.

FSB is often raised as Guskline mentioned, but from what I have read it isn't as important to raise it on FX as it was on Denebs/Thubans.

Decent guide here on overclocking Bulldozer, I would imagine most of this applies to PD as well.

http://www.overclock.net/t/1140459/...nce-scaling-charts-max-ocs-ln2-results-coming

 

[blckgrffn]

As far as I was aware, BD executes a 256bit op in a single cycle. However the bottleneck comes from that 256bit FPU being shared between two cores.

http://www.extremetech.com/computing/100583-analyzing-bulldozers-scaling-single-thread-performance/4

http://www.extremetech.com/computing/100583-analyzing-bulldozers-scaling-single-thread-performance (picture!)

Again, and I don't have time to go find the articles on this (you know, the big one that dug into the cache structure, etc on BD) but there is something about the FPU being starved for load and store ports when it comes to this as well.

If AMD had four FPUs that were as good as 4 FPUs from Intel, we'd get similar performance between the two, despite the share nature. I maintain that it isn't the "core sharing" - its just that the FPU itself cannot handle the AVX load that the current Intel one can.

I stand corrected on the number of cycles it takes to operate, but I believe that due to the cache structure for the FPU, it can't really do a 256-bit AVX operation every cycle.

 

[jvroig]

What would truly help IDC here is if the FX owners we have here post their LinX/IBT Gflops at Max setting. If all of you guys say you get around 35Gflops also, then all is well, IDC's cpu is then performing as expected. I am also very curious to find out myself.

BTW jvroig, what mb are you presently using for AMD chips?

My older mb for the X4 is an MSI 785GT E63. The X6 is running on an MSI 990FXA-GD80, but only because when the gigabyte mobo it used to run on died last year, that was the only quality mobo I could get my hands on.

 

[inf64]

I see.

And what about Vdram settings, with Intel you are well advised to stick with 1.5V memory or you'll burn up your IMC. Where is that threshold for my 8350?

One thing I was quite surprised to see was how poor the performance is in LinX. I'm wondering if I have something set wrong.

At stock 4GHz clocks running 8 threads and problem size 43122 (~14.2GB), my 8350 is turning in a paltry 35GFlops. At the same settings, same OS and so forth, but running only 4 threads my 3770k turns in ~114GFlops at 4GHz.

Is the LinX gap supposed to be that large? Am I missing some key Win7 update or something?

I understand LinX performance is not indicative of general performance because LinX is a specialized synthetic benchmark, but I am worried that my 8350's performance is low because I have something setup incorrectly. I expected a score that was at least 2x-3x the current value, up around 80 or 90 GFlops.

Linx has no support for FMA,at least the version you are running. FMA optimized you would see 2x of that,roughly.

 

[Abwx]

I understand LinX performance is not indicative of general performance because LinX is a specialized synthetic benchmark, but I am worried that my 8350's performance is low because I have something setup incorrectly. I expected a score that was at least 2x-3x the current value, up around 80 or 90 GFlops.

Peak FP throughput cant be reached if the soft doesnt take
advantages of FMA wich can double the FP exe rate....

Edit : didnt see Inf64 post above......

 

[guskline]

Idontcare: I upped my dram voltage to 1.538. Helps with stability.

 

[Idontcare]

I'm now getting ~80GFlops after downloading the most recent version of LinX that is posted on XS.

Ver 11.0.1.005

 

[inf64]

I'm now getting ~80GFlops after downloading the most recent version of LinX that is posted on XS.

Ver 11.0.1.005

Yep, now it runs FMA codepath . Not too bad for 4 shared units . Still less efficient than IB's 4 cores though so they have to work on that in future iterations of the core.

 

[Idontcare]

CPU/NB 1.350 V

Any idea why my crosshair mobo wants to run the CPU/NB at 1.40V when I set the ram to DDR3-1866? (versus running ~1.19V when I set ram to DDR3-1600)

Seems like a rather large jump in the voltage. DRAM voltage is 1.5V.

 

[AtenRa]

Sorry: I just noticed that! That type of testing is way above my "pay scale!"

http://atenra.blog.com/2012/02/01/amd’s-bulldozer-cmt-scaling/

 

[guskline]

Any idea why my crosshair mobo wants to run the CPU/NB at 1.40V when I set the ram to DDR3-1866? (versus running ~1.19V when I set ram to DDR3-1600)

Seems like a rather large jump in the voltage. DRAM voltage is 1.5V.

Perhaps to maintain stability it needs more power. From everything I've read, your mb is the Cadillac and has an advanced BIOS so I'd go with it if your temps are acceptable. Also I see from my post on the last page that my CPU/NB setting is 1.35 V and I have my ram at 1600 (the fsb drives it to @1738 but it is rated for 1866 so I'm ok).