So I recently picked up a couple of Opteron 6168 (Magny-Cours) "extra spicy" editions from another folder on my other team (37412) in hopes of being able to do a little comparo between overclocking Xeons and overclocking Opterons.
While I realize it's nigh impossible to get these Opterons in the wild, I still wanted to report on my findings of what I think are a pair of pretty damn good CPUs. Just because this is the [H]ardforum right?
I plugged 'em in on the weekend and let 'em run some F@H BigAdv units and at stock speeds they are, well, exactly the same as regular 6168s.
But, these chips are special because they come with an unlocked multiplier, thus the "black edition" label on this thread.
The 6168's stock speed is rather low at 1.9Ghz, but with 12 real cores and quad channel RAM access, there is a lot of performance even there.
The lowest BigAdv (2684) results are 40+K PPD for a pair of 6168s, and in general, they are competitive at stock with a pair of Xeon x5650s. In fact, for once I'm actually looking forward to getting a 2684 to see how it chews through one of these.
So anyway, up 500mhz from stock speed @ 2.4Ghz with a slight voltage bump from 1.075 to 1.10v on VCore, they are folding rather nicely. Stock on a 2685 the 6168 pair gets somewhere around 53K PPD with a 23 minute TPF. Right now at 2.4Ghz they are getting a TPF of 19:40 and 70K PPD.
All this with the chips being only 49 - 50 celsius, so I'm thinking there is a fair bit o' head room here. My aim is to see if I can hit the magical 3.0Ghz mark, but 2.7 - 2.8 should also be pretty cool.
So far they've been rock-solid, and strangely enough, having run LinX on 'em for a couple of hours last night I find it strange that the GFlops reported are over 145 Gflops. Not that it's a benchmark, but rather strange that this is nearly double what my two overclocked Xeon x5650s get.
Alternatively this unit's PPD on my SR-2 config is around 135K PPD, but I'm sure that when I settle on a final OC on these chips, they will likely be able to pump out nearly 100K PPD on the fastest BigAdv work units. And I highly doubt the power usage will be anywhere close to the 500+ watts the SR-2 configuration is using. Kill-a-Watt to the rescue.
The process for overclocking these chips is far different from the Xeons on SR-2. Since the bus speed cannot be moved up past 200mhz, and the multiplier cannot be changed via AMD OC tools, a combination of tools is needed.
Using a program called RW-Everything and some WRMSR calls (Write MSR or "memory specific registers" the multiplier, VID and NB voltage can be changed per chip, and the actual multiplier can be changed per core.
I was able to find out the MSR bits that needed to be changed from the fine gentleman who sold them to me, and I also used K10Stat's Debug mode to find out how these bits relate to multi, VID and NB voltage. Taking this output and copying them to a read-write script/text file for RW-Everything was easy. I just had to create a text file with a .rw extension to get that done, and it uses three entries per CPU core. The first entry puts the chip in P-State 1, the second modifies P-State 0 (full performance) VID and multi, and the third returns the core to P-state 0, which activates the changes, as they need to be refreshed.
I just thought I'd post this to help everyone remember that the hardware part is half the fun of folding. I'll update my findings as I update my overclocks
While I realize it's nigh impossible to get these Opterons in the wild, I still wanted to report on my findings of what I think are a pair of pretty damn good CPUs. Just because this is the [H]ardforum right?
I plugged 'em in on the weekend and let 'em run some F@H BigAdv units and at stock speeds they are, well, exactly the same as regular 6168s.
But, these chips are special because they come with an unlocked multiplier, thus the "black edition" label on this thread.
The 6168's stock speed is rather low at 1.9Ghz, but with 12 real cores and quad channel RAM access, there is a lot of performance even there.
The lowest BigAdv (2684) results are 40+K PPD for a pair of 6168s, and in general, they are competitive at stock with a pair of Xeon x5650s. In fact, for once I'm actually looking forward to getting a 2684 to see how it chews through one of these.
So anyway, up 500mhz from stock speed @ 2.4Ghz with a slight voltage bump from 1.075 to 1.10v on VCore, they are folding rather nicely. Stock on a 2685 the 6168 pair gets somewhere around 53K PPD with a 23 minute TPF. Right now at 2.4Ghz they are getting a TPF of 19:40 and 70K PPD.
All this with the chips being only 49 - 50 celsius, so I'm thinking there is a fair bit o' head room here. My aim is to see if I can hit the magical 3.0Ghz mark, but 2.7 - 2.8 should also be pretty cool.
So far they've been rock-solid, and strangely enough, having run LinX on 'em for a couple of hours last night I find it strange that the GFlops reported are over 145 Gflops. Not that it's a benchmark, but rather strange that this is nearly double what my two overclocked Xeon x5650s get.
Alternatively this unit's PPD on my SR-2 config is around 135K PPD, but I'm sure that when I settle on a final OC on these chips, they will likely be able to pump out nearly 100K PPD on the fastest BigAdv work units. And I highly doubt the power usage will be anywhere close to the 500+ watts the SR-2 configuration is using. Kill-a-Watt to the rescue.
The process for overclocking these chips is far different from the Xeons on SR-2. Since the bus speed cannot be moved up past 200mhz, and the multiplier cannot be changed via AMD OC tools, a combination of tools is needed.
Using a program called RW-Everything and some WRMSR calls (Write MSR or "memory specific registers" the multiplier, VID and NB voltage can be changed per chip, and the actual multiplier can be changed per core.
I was able to find out the MSR bits that needed to be changed from the fine gentleman who sold them to me, and I also used K10Stat's Debug mode to find out how these bits relate to multi, VID and NB voltage. Taking this output and copying them to a read-write script/text file for RW-Everything was easy. I just had to create a text file with a .rw extension to get that done, and it uses three entries per CPU core. The first entry puts the chip in P-State 1, the second modifies P-State 0 (full performance) VID and multi, and the third returns the core to P-state 0, which activates the changes, as they need to be refreshed.
I just thought I'd post this to help everyone remember that the hardware part is half the fun of folding. I'll update my findings as I update my overclocks