FX-8120 BOINC Benchmarks - Real World Stuff

Well, I am doing what I did with the 1090T... start from stock, bench, Prime-95... tweak, do it over. Learning some things! 8120 is more sensitive to CPU/NB speeds/voltages... trying to get it 100% stable at 3.6GHz now... working my way up to 4.3 I hope...

At 3.6, it passmarks the same speed as the 8150 chip...

Progress I guess after 50 reboots... always only change one thing at a time, slowly...

Sigh...



Update: Turns out my Uber-Cheapo 16Gig (4x4g) DRAM setup ($100 Blue Vengence 1600MHz kit) won't go so fast with 2 modules loading the lines (I think!). Maybe that is why the hyper-sensitive CPU/NB settings. Turns out, just slowed it down a bit to around 1425Mhz and I can go tons faster... at 3.9GHz+ now with 265 CPU/NB clock... making progress. (yes, I ran D.O.C.P. too, no help...)



Updated:

Okay, start off STOCK fixed clocks, no turbo or anything enabled... stock volts etc...




Now, this is what I ended up with... want to run it this way a while and see how it does....





And CPU-Z validate of course...


http://valid.canardpc.com/show_oc.php?id=2048217



For those that will ask, yes it's water cooled, yes it ran Prime-95 Blend 30 minutes fine, max Temp 44c, and now it runs Linux Glitch free and solid...







Updated:

8120 OC'ed Linux Bench marks: 2997,13371 Float and integer.

1090T OC'ed Linux Bench marks: 3707,17965 Float and integer.

I'm betting I have some corrupted install data or files.... but still looks like the 8120 is somehow WAY slower... have to investigate this and see if something can be discovered... maybe from scratch install now that it's solid.

Verified with fresh Ubuntu 11.10 install... it's slow... PERIOD... We will see how it does running overnight doing BOINC....

Update:

Ran all night at 4.24GHz solid without error. It is taking LONGER to do Optima WU's than the 3.94GHz 1090T system even with that higher overclock. So, clearly, the FPU construction (Shared between two cores) is having an affect (about 25% slower per WU). I'll try to push it faster, then back to house work.

I added the current figures for Optima to the CPU spreadsheet... the average PPS per core 1090T vs. 8120 is:

1090T:
1959 32.7 0.54424.580.013451084
11025183.83.063138.280.013451084

8120:
3642 60.7 1.01234.480.00934818
12236203.93.399115.830.00934818

At these values, the 8120 makes 8x818=6544 PPD total while the 1090T makes 6x1084=6504 PPD total.

So far, the 1090T is doing a lot better PER CORE running Optima WU's and validating the different Linux benchmarks 2997 vs. 3707 FLOPS.

Keep in mind the OC spread... 1090T went from 3.2GHz to 3.94GHz (23% increase) vs. FX-8120 3.1GHz to 4.24GHz (36% increase). It is only slightly better than the 4.156GHZ 955BE system at this point... Have to say, I am disappointed. That split FPU thing really hurts crunching floating point tasks...



FINAL Overclock Update:
Well, I knew I was close, but had no idea how close. Using the other posted 8150 overclock results, I figured I would be able to match them MINUS the rated 500MHz speed difference. That is why I stopped yesterday at 4.24GHz or so.

After spending HOURS and HOURS (yuck) trying, I've come to the following conclusions about this particular FX-8120 CPU.

1) It hates any multiplier over 16... even 16.5 is problematic. (maybe that why binned 8120??)
2) Manually set CPU voltages over 1.435 Volts make no difference in max speed, but cause a lot more heat.
3) I was on the edge of the max settings when I stopped yesterday. Sure, I can get it to boot 4.6GHZ, but every time I try to get a CPU-Z Validate, it crashes... LOL!

Sabertooth Mobo Notes:
The Asus Sabertooth board has many settings to compensate for VDROP and VDROOP etc. I watch CPU-Z voltage values and have it set to Optimize mode... it adds voltage, typically over 1.51 readings at max clock speeds. I've also tried various combinations on Load Line Calibration, HT speeds etc. to no real benefit. The BIOS seems to increase the voltage with sensed load and also seems to supply predetermined (or calculated?) voltages when things set to AUTO like CPU and CPU/NB voltages. I ended up setting most things to the EXTEME mode... LOL! Well, it is OCN after all....


Did My Best:
Best results were obtained by first manually setting the CPU and CPU/NB voltages to the BIOS displayed ones, then dropping them down slowly to find a minimum. Then, running Prime-95 Max Heat tests, see what could be done. I ended up doing a sort of teeter-totter thing with the CPU vs. CPU/NB voltages trying to find the lowest values and best combo. I did the same teeter-totter thing trying to find max clock speeds with Multiplier vs. BCLK as well and could never get anything over 4.3GHz stable with the Multiplier over 16... stinking brick wall with that multiplier (and maybe why this chip binned an FX-8120 and not FX-8150!).

Anyway, finding the max stable clocks and then lowering the CPU and CPU/NB voltages took hours... I even covered the fan exhaust on top of the 800D box to let it heat up faster for testing... then uncover when it hit 52C and let it rise on its own thereafter.faster for testing... then uncover when it hit 52C and let it rise on its own thereafter.

I tried several CPU/HT and HT Link clock values as well. Not much choice in that area... and not much performance difference...


This is MAX CLOCK ... 4.334GHz stable...


But, look at the Price I had to pay.... in terms of temperature... went up 10c+ to get that 100MHz extra!!! The temp was still climbing too... eventually settled at 58c...

Fact is, this particular chip is really max'ed out at 4.3GHz... which isn't bad considering that represents a 39% increase in the overall speed...

This is where I ended up setting it for normal (read OverClocked BOINC) use 24/7 @ 4.27GHz...




Runs a LOT cooler not having to bump all the voltages to MAX... but getting that little bit extra is costly in terms of POWER and HEAT!!!

I'm happy where it is now at 4.27GHz and will leave it there... That is about exactly 500MHz lower than what others have achieved with the 8150... so about exactly what I expected.

Conclusion:

The FX-8120 isn't the smoking fast BOINC cruncher I was hoping for... barely better than the CHEAPER 1090T chip for sure... but that is only on floating point tasks. There are some Integer based tasks and memory manipulation based tasks yet to test...

Question: Would I have purchased this chip if I had known what I know now?

Answer: Well, I wanted that Sabertooth board anyway to really push 3 GPU cards in one 800D box... and needed a new MOBO for that anyway. I had that planned for the 1090T already... BUT, I was expecting a LOT more from Bulldozer chip... So, the answer is, I would have purchased the MOBO anyway to support Dual GTX460 and a Single 6990... but I would NOT have purchased the FX-8120...

But the FX-8150 would have done better... but TG and NE ran out so fast it wasn't funny... still I had to find out for myself... so it's a lesson learned and maybe in the future things will get better... It's barely better than the 1090T, but I bet I could have got some more out of that 1090T on the Sabertooth board...

Cheers!

Okay, the house got warm, the 800D got warmer (inside and out)... got a glitch... so backed it down slightly to help stabilize it. Now running with these settings....



Did the Sisoft Sandra thing... got this result...




Now to compare raw power, check out what the Dry-Stones (Integer) and Wet-Stones (Floating Point) (lol) of the 8120 @ 4.22GHz are compared to an i7-2600K@4.6GHz... Kicks Azz on the 8120... sigh...



So far, looking okay... still not the screamer I (and most folks) were hoping for...

I also wrote a nice review about this chip on TigerDirect and NewEgg... neither review was very good... NewEgg didn't post it yet... maybe never...

http://www.tigerdirect.com/applicati...958&CatId=1946
http://www.newegg.com/Product/Produc...82E16819103961




(Oh, text file of all the SiSoft Sandra Arithmetic stuff...)

Well, numbers don't lie. The FX-8120 is pitiful... The two BOINC tests, especially the NFS test show the huge performance differences clearly. NFS has 1.3GByte sized tasks and uses a LOT of memory I/O. Optima use much less memory at 2MBytes and seems more INTEGER intensive, but more of it stays in cache is my guess.

THE OPTIMA TESTS RUN using Ubuntu Linux-64b.

THE NFS TESTS RUN using Windows-7 (Version 1.09 Tasks ) and one set on the 1090T using Ubuntu Linux-64b (Version 1.07 Tasks).


NFS@Home running 16e Lattice Sieve

3.94GHz Phenom II X6 1090T:

(Windows-7 64b)


(Linux-64b) NOTICE the Linux task is much faster for whatever reason...


4.22GHz AMD FX-8120:

(Windows-7 64b)


As one can see, the FX-8120 takes twice as long per task on average running and the extra two cores don't make up the difference in total output by a long shot!!!

OPTIMA@Home running 4.00 Molecular Conformation Problem

3.94GHz Phenom II X6 1090T:

(Linux-64b)


4.22GHz AMD FX-8120:

(Linux-64b)


Again we see the FX-8120 way behind!!!

Not good...

Conclusion:

The FX-8120 (and probably the faster FX-8150) is just not made to do BOINC tasks at all... I would get the Phenom II X6 1090T or 1100T.

You may have seen this on the web; it has happened to me a LOT!



Not only that, but stable tasks like Freehal and Rosetta and others are giving calculation errors constantly for no reason. I even downclocked the CPU today to 3.4GHz and still get random errors!!! Keep in mind this chip otherwise passed Prime-95 and other tests fine, but when crunching tasks all night, weird errors happen.

Something is wrong with this chip... the FX-8120 is NOT working properly.

At this point, I removed the CPU and replaced it with a Phenom II X6 1090T and now the system works perfectly....

Sigh... total flop in FLOPS and normal stuff...

Nahh, I will report most benchmarks when I get them in the Spreadsheet. The two reserved posts are plenty for comments and updates... No worries! Post away!!! Don't move a thing!