Figured I'd make a thread regarding the trials and tribulations with my FX-8350. This is a new area for me, so expect some noobness to my efforts, feedback and pointers will be much appreciated as I attempt to scale the new learning curve :thumbsup:
A trip down memory lane:
The last time I had the opportunity to really dive into AMD hardware it was circa 1998 or 1999 and I had just acquired 24 800MHz K7 systems with a government grant in the name of assembling what was called, at the time, a "beowulf class supercomputer". It was basically a cluster of desktop PC's, connected by a high-speed interface (100Mbit ethernet at the time) that ran multi-threaded applications much as we all do today with our multi-core processors.
^ 12-node cluster of 800MHz K7's.
^ 24-node cluster of 800MHz K7's, built by me and my lab mate Jason.
The primary function of this cluster was to run a parallelized version of a computation chemistry application called Gaussian98. (quantum chemistry modeling) We were designing and building molecules which would be used to harvest/absorb sunlight and split water into hydrogen and oxygen gas. The modeling helped us decide which molecules would be worth attempting to engineer in the lab, which we also did right next door.
At any rate this cluster had long been my benchmark for "performance in a single CPU" because I longed for the day when I would have the same performance as this cluster but in a single cpu that could be purchased at commodity prices.
Enter the FX-8350! 4GHz of 8 core goodness (or 4 cores, depending on how you count ALU vs FPU).
From a straight dumb-flops computation (thanks to guskline for motivating me to think about this) the old cluster was capable of 24*800MHz = 19,200 MHz of K7 IPC. The FX8350 would be capable of 8*4000MHz = 32,000MHz of K11 IPC. (yes I know piledriver is not a K11, but there is no harm in making the analogy IMO)
Surely K7 IPC < K11 IPC, so from a brute flops standpoint the FX-8350 probably has at least 2x the capability of my previous benchmark, the 24-node beowulf cluster.
Not to mention the cost and power-consumption reduction that comes with an FX-8350 over the old cluster, and just from a span of ~12yrs.
All right, enough nostalgia, lets talk about the FX-8350 itself.
First thing is the testbed, I bought the Crosshair V Formula-Z for one specific reason - the ProbeIt belt.
I learned with my MIVE-Z in testing the 2600k and 3770k that relying on CPUz, or any other software voltage reporting tool for that matter, to document the applied CPU voltage was a pursuit of one blind man by another. If you truly want to know what voltage your CPU is experiencing you simply have no choice but to get a physical voltmeter connected to the circuit.
Now one obvious issue with the probe-it belt is the matter of relative dimensions. Those probe pabs are silly tiny and close together. I also did not like my prospects of being able to maintain electrical isolation between the probes themselves (the alligator clips in my case).
So I realized first thing I needed to do here was to electrically isolate the pads, and the clips, lest I make some sparks by accident and let out the blue smoke
I turned to the oldest trick in the book, black electrical tape and masking off the work area:
^ here I've isolated the two probe pads of interest.
^ then I masked off the alligator clips themselves, and using a razor blade I opened just one metal tooth at the very front of the clip, but taking care to keep the sides of the tooth still covered with the electrical tape. (I don't want the alligator clips to short out each other)
And the final assembly:
OK, all is well and good, no risk of shorting now.
And the result?
For my mobo, which is updated with the latest BIOS, CPUz under-reports the idle voltage by a hefty 0.108V! That is a rather large miss.
And at full load (LinX, 8 threads):
The gap between actual and CPUz reported has markedly narrowed but it is still the case that CPUz is under-reporting the actual Vcore. In this case it is off by ~0.010V.
I don't have the screen grab comparison to show you here, but in my various runs of LinX so far the delta between CPUz and actual Vcore tends to run about 2x this - i.e. ~0.020V error on behalf of CPUz at load volts.
Is this gap specific to my mobo? Or is it common to everyone's Crosshair V Formula-Z? I can't say, but it just goes to show that relying on CPUz values is pretty much a crapshoot.
Another anecdotal observation, not captured in these static images but notable by the eye in real-time, is that CPUz likes to shake around the reported Vcore like a cat playing with a ball of string; whereas the measured values are much more stable and consistent.
A trip down memory lane:
The last time I had the opportunity to really dive into AMD hardware it was circa 1998 or 1999 and I had just acquired 24 800MHz K7 systems with a government grant in the name of assembling what was called, at the time, a "beowulf class supercomputer". It was basically a cluster of desktop PC's, connected by a high-speed interface (100Mbit ethernet at the time) that ran multi-threaded applications much as we all do today with our multi-core processors.
^ 12-node cluster of 800MHz K7's.
^ 24-node cluster of 800MHz K7's, built by me and my lab mate Jason.
The primary function of this cluster was to run a parallelized version of a computation chemistry application called Gaussian98. (quantum chemistry modeling) We were designing and building molecules which would be used to harvest/absorb sunlight and split water into hydrogen and oxygen gas. The modeling helped us decide which molecules would be worth attempting to engineer in the lab, which we also did right next door.
At any rate this cluster had long been my benchmark for "performance in a single CPU" because I longed for the day when I would have the same performance as this cluster but in a single cpu that could be purchased at commodity prices.
Enter the FX-8350! 4GHz of 8 core goodness (or 4 cores, depending on how you count ALU vs FPU).
From a straight dumb-flops computation (thanks to guskline for motivating me to think about this) the old cluster was capable of 24*800MHz = 19,200 MHz of K7 IPC. The FX8350 would be capable of 8*4000MHz = 32,000MHz of K11 IPC. (yes I know piledriver is not a K11, but there is no harm in making the analogy IMO)
Surely K7 IPC < K11 IPC, so from a brute flops standpoint the FX-8350 probably has at least 2x the capability of my previous benchmark, the 24-node beowulf cluster.
Not to mention the cost and power-consumption reduction that comes with an FX-8350 over the old cluster, and just from a span of ~12yrs.
All right, enough nostalgia, lets talk about the FX-8350 itself.
First thing is the testbed, I bought the Crosshair V Formula-Z for one specific reason - the ProbeIt belt.
I learned with my MIVE-Z in testing the 2600k and 3770k that relying on CPUz, or any other software voltage reporting tool for that matter, to document the applied CPU voltage was a pursuit of one blind man by another. If you truly want to know what voltage your CPU is experiencing you simply have no choice but to get a physical voltmeter connected to the circuit.
Now one obvious issue with the probe-it belt is the matter of relative dimensions. Those probe pabs are silly tiny and close together. I also did not like my prospects of being able to maintain electrical isolation between the probes themselves (the alligator clips in my case).
So I realized first thing I needed to do here was to electrically isolate the pads, and the clips, lest I make some sparks by accident and let out the blue smoke
I turned to the oldest trick in the book, black electrical tape and masking off the work area:
^ here I've isolated the two probe pads of interest.
^ then I masked off the alligator clips themselves, and using a razor blade I opened just one metal tooth at the very front of the clip, but taking care to keep the sides of the tooth still covered with the electrical tape. (I don't want the alligator clips to short out each other)
And the final assembly:
OK, all is well and good, no risk of shorting now.
And the result?
For my mobo, which is updated with the latest BIOS, CPUz under-reports the idle voltage by a hefty 0.108V! That is a rather large miss.
And at full load (LinX, 8 threads):
The gap between actual and CPUz reported has markedly narrowed but it is still the case that CPUz is under-reporting the actual Vcore. In this case it is off by ~0.010V.
I don't have the screen grab comparison to show you here, but in my various runs of LinX so far the delta between CPUz and actual Vcore tends to run about 2x this - i.e. ~0.020V error on behalf of CPUz at load volts.
Is this gap specific to my mobo? Or is it common to everyone's Crosshair V Formula-Z? I can't say, but it just goes to show that relying on CPUz values is pretty much a crapshoot.
Another anecdotal observation, not captured in these static images but notable by the eye in real-time, is that CPUz likes to shake around the reported Vcore like a cat playing with a ball of string; whereas the measured values are much more stable and consistent.