PrimeGrid Distributed Computing & AVX - When an ISA Extension Matters

[blckgrffn]

(and why aren't you racing? )

Currently there is a race on for the DC project PrimeGrid. The Work Units (WUs) are huge and take tens of hours to complete. Local thread here: http://forums.anandtech.com/showthread.php?t=2243779

Points are a representation of the difficulty of the WU, so a WU with the ~same difficulty will score the same. This means while no two WUs are identical, it is possible to compare them via dividing the time it took to complete by the score to get the time it takes in seconds to score a point. Points per second would be more intuitive, but the values so low as to meaningless.

Some data that might interest you:

2600k, 3.8, AVX, w/HT: 57.2
2500k, 3.4, AVX: 48.0
i3-2100, 3.1, no AVX, no HT: 117.55
1045t, 3.04: 74.9
960t, 6C, 3.2: 64.8
8120, 3.65 (No AVX support from PrimeGrid): 58.2

Now, dividing that into the number of seconds the challenge ran, 432,000 (5*24*60*60) and you can see how many points a core will score over that period:

The HT tax on the 2600k is apparent vs the 2500k. You can see how the i3 struggles, deprived of AVX and HT (now, this may not be good data - I've tweaked the VM config but it takes hours/days to see how that effects WU completion time right now) while we see the Thuban cores doing fine. BD is actually holding its own here... clock for clock it is roughly equal to the Thubans when the FPU is being pounded on. Imagine if it had eight!

The real story is how many cores each CPU "has" - whether they are pure "cores" with full ALU/FPU, etc. or shared via CMT or HT:

Obviously, the HT tax is worth it with so much more work being done over the period of time. BD shows that given 40% or so more power vs the 2600k, it can stay in the game.

I am really looking for some Nehalem/Westmere/Lynnfield numbers, I'd take them from random racers but Prime Grid doesn't report actual clock speed

I need more AVX vs non-AVX data on the Intel side, the i3 numbers look off.

Thanks to Uppsala for donating his FX data. May his 960t rig RIP!

Finally, if you think your OC is stable on your SB, SB-E or IVB, I invite you to participate in the rest of his race. TeAm Anandtech has one rig that has died in a fire, maybe yours will do better! Intstructions in the linked thread above.

Comments/trashing of my data/conclusions welcome. Not an ace graph maker

If there is enough interest, I'd attempt to normalize for each CPU by 100Mhz or similar to give an idea how things stack up clock for clock. Sorry, I am not gathering power data as much as I would like to. The 2600k has the kill-a-watt and is reading 140W under this load. I estimate the Thubans @ ~200-220W based on previous experience.

 

[blckgrffn]

 

[petrusbroder]

Thanks for the data, blckgrffn. The AMD-CPU-data are quite close to mine.
However, there is a big difference between the FX8150 (which is similar to the 8120, a bit faster though) and the FX6100 which - in my hands (no OC, stock all) is approx half as powerfull as the 8150. I would have expected 2/3 ... but it is actually a lot less.
I am still asking myself why ... and I have no idea.

 

[blckgrffn]

Thanks for the data, blckgrffn. The AMD-CPU-data are quite close to mine.
However, there is a big difference between the FX8150 (which is similar to the 8120, a bit faster though) and the FX6100 which - in my hands (no OC, stock all) is approx half as powerfull as the 8150. I would have expected 2/3 ... but it is actually a lot less.
I am still asking myself why ... and I have no idea.

Hmm, is there an OS difference? I am not sure if AVX is working for BD, but it could be that one has operating system AVX support and one does not? A 20-40% per thread difference is what I am expecting from AVX based on past experience with PrimeGrid.

Also, I will post the raw data shortly in the placeholder post above, and if you could provide the matching data (all of three numbers!) then I'll update the .xls and the graphs should nearly take care of themselves. That would be interesting...

I am also seeing some variation across my Thubans... I think there are some cherry WU's out there that don't increase work time in proportion to their scores, and in a project like this were you get big swaths of time consuming work it doesn't even out like it can when you get large numbers of shorter duration WUs more frequently...

 

[biodoc]

I don't have data for WUs completed with hyperthreading enabled with my 3930K and 2600K machines since turned it off on both rigs after about 12-18 hours into the race. I also got a late start since I was travelling and didn't get home until Saturday evening EST.

So I have 12 "hybrid" WUs on the 3930K (part HT on and part HT off) and 8 hybrid WUs on the 2600K. Both machines are on 64-bit linux and both were clocked down before the start. The 3939K is running at 3.8 GHz and the 2600K is at 4 GHz.

I'l stick with HT off for the remainder of the race but I suspect it may have been a mistake on my part based on the nice data you've collected.

Thanks for posting this data!

 

[Uppsala9496]

Thanks for the data. Really surprised to see how the 8120 compares to the 2600K. Not sure what the power draw is on the box (no killawatt).

EDIT: Oh and no AVX.

 

[VirtualLarry]

Hmm, i3-2100 lacks AVX? I didn't realize that. So much for using a dual-core for crunching.

 

[grimpr]

Hmm, i3-2100 lacks AVX? I didn't realize that. So much for using a dual-core for crunching.

It does support AVX, all the Core i3 lineup. Prime95 beta loves it.

http://www.mersenneforum.org/showthread.php?t=16779

 

[Plimogz]

Finally, if you think your OC is stable on your SB, SB-E or IVB, I invite you to participate in the rest of his race. TeAm Anandtech has one rig that has died in a fire, maybe yours will do better! Intstructions in the linked thread above.

So is Primegrid really so tough on unstable overclocks? Because I got a BSOD a while into running it, but I think it could be on account of some shabby power-management options.

Might Primegrid crash a system which can crunch WCG endlessly without errors or hiccups? (no, no, don't tell me, let me bask in my denial)

Opinions?

PS. Specs are 2500K - GA-P67A-UD5 - WD Black 1TB - 4890 - PC P&C 750W - WIN7 64

And I've successfully run the basic tests outlined in the sticky at the top of the sub-forum (ie. Linpack, P95, HCI) though I don't have time for the gold standards ATM.

 

[soccerballtux]

(and why aren't you racing? )

Currently there is a race on for the DC project PrimeGrid. The Work Units (WUs) are huge and take tens of hours to complete. Local thread here: http://forums.anandtech.com/showthread.php?t=2243779

Points are a representation of the difficulty of the WU, so a WU with the ~same difficulty will score the same. This means while no two WUs are identical, it is possible to compare them via dividing the time it took to complete by the score to get the time it takes in seconds to score a point. Points per second would be more intuitive, but the values so low as to meaningless.

Some data that might interest you:

2600k, 3.8, AVX, w/HT: 57.2
2500k, 3.4, AVX: 48.0
i3-2100, 3.1, no AVX, no HT: 117.55
1045t, 3.04: 74.9
960t, 6C, 3.2: 64.8
8120, 3.65 (AVX?): 58.2

Now, dividing that into the number of seconds the challenge ran, 432,000 (5*24*60*60) and you can see how many points a core will score over that period:

The HT tax on the 2600k is apparent vs the 2500k. You can see how the i3 struggles, deprived of AVX and HT (now, this may not be good data - I've tweaked the VM config but it takes hours/days to see how that effects WU completion time right now) while we see the Thuban cores doing fine. BD is actually holding its own here... clock for clock it is roughly equal to the Thubans when the FPU is being pounded on. Imagine if it had eight!

The real story is how many cores each CPU "has" - whether they are pure "cores" with full ALU/FPU, etc. or shared via CMT or HT:

Obviously, the HT tax is worth it with so much more work being done over the period of time. BD shows that given 40% or so more power vs the 2600k, it can stay in the game.

I am really looking for some Nehalem/Westmere/Lynnfield numbers, I'd take them from random racers but Prime Grid doesn't report actual clock speed

I need more AVX vs non-AVX data on the Intel side, the i3 numbers look off.

Thanks to Uppsala for donating his FX data. May his 960t rig RIP!

Finally, if you think your OC is stable on your SB, SB-E or IVB, I invite you to participate in the rest of his race. TeAm Anandtech has one rig that has died in a fire, maybe yours will do better! Intstructions in the linked thread above.

Comments/trashing of my data/conclusions welcome. Not an ace graph maker

If there is enough interest, I'd attempt to normalize for each CPU by 100Mhz or similar to give an idea how things stack up clock for clock. Sorry, I am not gathering power data as much as I would like to. The 2600k has the kill-a-watt and is reading 140W under this load. I estimate the Thubans @ ~200-220W based on previous experience.

congradulations. You have sort of found one reason to buy a Bulldozer.

 

[paperwastage]

congradulations. You have sort of found one reason to buy a Bulldozer.

similar results, but twice as much power consumption for bulldozer? maybe....

interesting results.... thanks OP and others who are adding to this thread... after my finals, maybe I'll help out a bit

 

[blckgrffn]

It does support AVX, all the Core i3 lineup. Prime95 beta loves it.

http://www.mersenneforum.org/showthread.php?t=16779

No AVX for me as it is running ESXi 4.1. Once I update to ESXi 5, it should get passed through to the VM, no problem. I am also running it with only 2vCPU, hence the lack of HT.

@ soccerballtux, any chance you could trim down that massive quote? It borks the readability of the thread. Thanks

FWIW, this isn't about justifying any CPU. It's just to illustrate some information. The last couple of races, AVX increased Intel performance by 30-40%. Previous to that, Thubans really made a case for them selves by virtue of their core count and likely performed more work in a linear fashion with their power consumption. Now, Intel appears to do as much (or more!) work with fewer cores in a smaller power envelope.

To me, that illustrates how not having a certain ISA, like AVX, can suddenly make a CPU non-competitive. It will be interesting to see if AMD essentially implements the same AVX capabilities from SB and IVB into PD - and if there will be game changers like this with Haswell too.

 

[blckgrffn]

So is Primegrid really so tough on unstable overclocks? Because I got a BSOD a while into running it, but I think it could be on account of some shabby power-management options.

Might Primegrid crash a system which can crunch WCG endlessly without errors or hiccups? (no, no, don't tell me, let me bask in my denial)

Opinions?

PS. Specs are 2500K - GA-P67A-UD5 - WD Black 1TB - 4890 - PC P&C 750W - WIN7 64

And I've successfully run the basic tests outlined in the sticky at the top of the sub-forum (ie. Linpack, P95, HCI) though I don't have time for the gold standards ATM.

Yes, this is super brutal. It is expected that if anything will bring down your machine, this will be it. Part of it is just the sustained, week long run. It also uses AVX, which draws more power and thus puts more stress on the power delivery system, etc.

A bit of the fun in these races is that rarely you just kick the tires and everything makes it through the end (kind of like any type of auto racing) - there is always something to take care of. In Uppsala9496's case, his 960t rig may have nearly caught on fire!

Thanks for the data. Really surprised to see how the 8120 compares to the 2600K. Not sure what the power draw is on the box (no killawatt).

EDIT: Oh and no AVX.

Thanks again for the data

Another quick note, yes, the i3 really is that slow. It's hard to believe AVX is making that big of a difference, given that AMD is lacking it yet is that much more competitive? I'll work on find other non-AVX Intel data...