CPU Multithreading Topic, User runtime

[blckgrffn]

Updated Core i3 POV-Ray score. It appears that scaling is a touch better than Original/4 - and that is with HT!

If you have a Core 2 Quad, it doesn't look like you need to hold out until Haswell. Cash out that investment for an i3 and enjoy the better power consumption while you wait

Come next year, it is likely the i3 will be easier to move than a Core 2 Quad.

 

[AtenRa]

I will post my Core2 Q9450@3.2GHz later and see if it will be faster than the Core i3.

 

[Maximilian]

Stock Core i5 2400s - 2.6ghz turbo

Cinebench 11.5

7 zip 32MB

TrueCrypt 500MB

x264

POVray 3.7 RC5

 

[blckgrffn]

I will post my Core2 Q9450@3.2GHz later and see if it will be faster than the Core i3.

Hah, it seems likely that it will be, but at maybe a huge power deficit. Will be interesting to see, anyway

You definitely have a nicer Core 2Q than I, as well.

 

[Ferzerp]

Pfft, the original post says any CPU, and when I start posting my CPUs of choice, we get that disclaimer.

Pffffft.

I also want to point out that I'd never in a million years want to game with that system, as the single threaded performance is pure crap due to the low clock speeds necessary when putting 10 cores on one chip.

edit: and yes, I know I'm cheating a bit by posting benchmarks from $9k or $18k of processor (each of the 4 retails for around $4.5k, but most of the benchies can only use 2 of 4).

 

[Ferzerp]

Nice! All my big machines are Linux / Unix, wish some of them were Windows.

Hopefully you are running DC on those

esx.

 

[AtenRa]

Pfft, the original post says any CPU, and when I start posting my CPUs of choice, we get that disclaimer.

Pffffft.

By all means continue posting more results with the E7. As i have said, it is nice to have it as a reference.

I just dont want every server guy come here and post, it will make it irrelevant (not to say depressing as well) for the rest of us

 

[blckgrffn]

esx.

You are are talking the Westmer-EX boxes, right? Not the legion of 2P Westmere boxes you mentioned? Those sound like they would be light on the ram density otherwise - but if you need the CPU power, you need the CPU power...

 

[Hatisherrif]

A very interesting thread this is indeed. I guess that as far as cinebench goes, these CPUs would more or less be tied clock-for-clock at the same clock speed:

AMD Phenom II X6 1090T
Intel Core i7 920
Intel Core i5 2500K
AMD Bulldozer FX 8150
Intel Core i5 3570K

I am only comparing the cinebench results, I'm not stating either of these is a better chip or that they are the same, just saying that they are pretty comparable when it comes to this particular benchmark. This is also interesting, as we can compare their other features and see which is "the best" right now. For example, the bulldozer may be more power hungry, while the "old" i7 920 may fair better with memory because of its triple-channel controller.

All in all, I say this thread is a win! No bias, no crapping, I think it is great. Informative as well

 

[Ferzerp]

You are are talking the Westmer-EX boxes, right? Not the legion of 2P Westmere boxes you mentioned? Those sound like they would be light on the ram density otherwise - but if you need the CPU power, you need the CPU power...

They're blade servers, and when the purchase was made, 16GB dimms were a bit pricy, and VMware hadn't upped the vSphere 5 allocation per socket to 96GB yet. Purchased today, they would be 192GB each, but we didn't predict the ram allocation would change, and it was cheaper at the time to buy twice the servers with 8GB dimms per server, and 2 vSphere licenses than it would have been to buy twelve 16 GB dimms and 4 vSphere licenses per server. If the vram allocation per socket had been changed prior to the order, it would have tipped it in the favor of denser memory, but we can go back if we run out and replace as needed. Add that to a need to spend the available funds for the project back in June, and though they weren't needed until now, the money was use it or lose it.

edit: basically, we got the same RAM volume, and twice the everything else that we would have gotten at the time if we had went with 16GB DIMMS.

 

[blckgrffn]

They're blade servers, and when the purchase was made, 16GB dimms were a bit pricy, and VMware hadn't upped the vSphere 5 allocation per socket to 96GB yet. Purchased today, they would be 192GB each, but we didn't predict the ram allocation would change, and it was cheaper at the time to buy twice the servers with 8GB dimms per server, and 2 vSphere licenses than it would have been to buy twelve 16 GB dimms and 4 vSphere licenses per server. If the vram allocation per socket had been changed prior to the order, it would have tipped it in the favor of denser memory, but we can go back if we run out and replace as needed. Add that to a need to spend the available funds for the project back in June, and though they weren't needed until now, the money was use it or lose it.

Ah, makes perfect sense. :thumbsup:

Clearly these three 1TB boxes we have are the last we'll order like that now. We were shafted hard by VMware on that, and let them hear our displeasure. ~200 VMs on a server (with an average size of ~2 vCPU and 4GB ram) is a solid investment.

Especially annoying about the situation is that, like you, we could have done ~6 512GB servers for the price of those three but were trying to optimize for Windows DC and VMware Ent+ licensing footprint. Dammit.

Also, having the memory model based on allocated memory makes me Hulk-inducing angry. We are development centric and are currently working at ~130% ram allocation. Paying for RAM you don't even have? Freaking classy, especially when Hyper-V and Xen Server both have burstable ram modes. Heck, even our AIX vms have that.

Bah. Enough for another thread. Sorry to clutter this one up.

Good to know there are x86 admins/architects posting though :thumbsup: I wish we had an appropriate sub-forum like [H] has for discussing this type of thing.

 

[Makaveli]

Neat, if I have some time this weekend, I'll try to see why cinebench would only use 32 threads (it should have access to 40), and maybe run a few of the other benchmarks. Also, technically, there are 4 procs in that thing, each 10c/20t, so really that score was 1.6 of one of those procs (as it was a 32 thread run). That makes the OC'ed 970 look impressive indeed.

What is the consensus on the processor group issue? Run two concurrently or what? I'll be honest, I couldn't get the x264 benchmark to utilize even half a processor group's worth of cores. I ran one instance, and got in the low 60's on pass two, then I ran 4 instances, and all 4 instances also got the low 60's on pass two at the same exact time. Not sure what to do there. Not sure how many instances I can run before the max throughput is met.

I should be able to do POVray as well, but I have to be careful not to step on any of the real work that system is doing.


edit: got the current release of cinebench, was able to do 40 threads, maybe I did before as well and forgot. It gets 20.48 with 40 threads. Cinebench won't run twice concurrently though, so I can't run two at the same time to add the result together (one per proc group).

Thanks i've updated your score on the chart. I had a feeling those numbers were from some kinda blade server that is in production so I do appreciate you posting the numbers from a enterprise class system. I also figured you would have run into issues with how many threads the app would be able to push on machine like that but it looks like the newer version of cinebench helped. That processor i'm not too familiar with so thanks for the 10c/20t break down.

Pfft, the original post says any CPU, and when I start posting my CPUs of choice, we get that disclaimer.

Pffffft.

I also want to point out that I'd never in a million years want to game with that system, as the single threaded performance is pure crap due to the low clock speeds necessary when putting 10 cores on one chip.

edit: and yes, I know I'm cheating a bit by posting benchmarks from $9k or $18k of processor (each of the 4 retails for around $4.5k, but most of the benchies can only use 2 of 4).

I did see the changes he made but I think we understand why.

Everyone showed up to race day in a civic and you came with a ferrari lol

 

[Ferzerp]

Everyone showed up to race day in a civic and you came with a ferrari lol

I would equate it to everyone coming with pickup trucks that could cruise along at 80 mph, but I brought a:

 

[blckgrffn]

Thanks i've updated your score on the chart. I had a feeling those numbers were from some kinda blade server that is in production so I do appceiate you posting the numbers from a enterprise class system. I also figured you would have run into issues with how many threads the app would be able to push on machine like that but it looks like the newer version of cinebenched helped. That processor i'm not too familiar with so thanks for the 10c/20t break down.



I did see the changes he made but I think we understand why.

Everyone showed up to race day in a civic and you came with a ferrari lol

lol, Ferzerp put it better That was what I was seeing in my mind's eye, but could get the analogy right....

 

[Makaveli]

agreed his analogy was better now that I think about it.

Guys i'm gonna try graph the 7-zip results next.

However I have a question from the screenshots what value's are the important ones.

Is it just the Rating (Mips) or is it the speed kb/s?

i'm thinking of showing the data for both Compressing and Decompressing.

 

[Hatisherrif]

I think this may suffice

I didn't include the crazy server scores though Also, I didn't find the Phenom II X6 score so it's missing for now.

 

[blckgrffn]

@ Maximilian -

A ~2.2Ghz SB Quad delivering a handy defeat to a 3.0ghz Core 2 Quad? Very nice. I wonder if that translates right into gaming too? It must...

 

[Maximilian]

@ Maximilian -

A ~2.2Ghz SB Quad delivering a handy defeat to a 3.0ghz Core 2 Quad? Very nice. I wonder if that translates right into gaming too? It must...

Yeah its a nice chip. Its clockspeed is 2.5ghz stock and it turbos up to 2.6 when all cores are loaded so the clock speed difference from the C2Q isnt too much, plus it dosent overclock, or at least its not supposed to. I was gonna settle for an i3 but this was cheaper surprisingly Its probably decent at gaming but ive never tested it.

 

[Makaveli]

I think this may suffice

I didn't include the crazy server scores though Also, I didn't find the Phenom II X6 score so it's missing for now.

Nice graph i'm still in the middle of making mine.

I was going to include the seperate scores for the compress and decompress but then your graph has the total.

Does this mean you will take over graphing duty and will do the rest

 

[Hatisherrif]

Nah, too much work for me. Not that I'm too busy or anything, I'm just lazy as hell. But I really love this thread, we get to see legit what we can expect from our CPUs on various benchmarks (which aren't exactly the best indicators of real-world performance, but what the heck, anybody can run them and that's cool).

Just use the same pattern for every graph and change the numbers, it would be great. I think this thread's destiny is either to become a sticky or die just about now (after we get the graphs). Great thread, again.

 

[exar333]

Hoping we can get someone with an IB to run the suite too.

 

[exar333]

I think this may suffice

I didn't include the crazy server scores though Also, I didn't find the Phenom II X6 score so it's missing for now.

:thumbsup:

Great looking graph.

 

[Maximilian]

Hoping we can get someone with an IB to run the suite too.

IB would be interesting to see.

We've seen ultra high end multicore insanity (thanks ferzerp), it would be cool to see some of the older less successful CPU's like the phenom 1 (i never see anyone with these chips) or that pentium D which had HT tech. Just for comparisons sake.

 

[blckgrffn]

Because we are always saying "Core2Quad == Phenom 2"

CPUz:

Cinebench - 3.4

x264 HD BENCHMARK 4.0 RESULTS

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Results for x264.exe r1913
==========================

Pass 1
------
encoded 1442 frames, 98.36 fps, 3912.32 kb/s
encoded 1442 frames, 98.64 fps, 3912.32 kb/s
encoded 1442 frames, 98.64 fps, 3912.32 kb/s
encoded 1442 frames, 97.84 fps, 3912.32 kb/s

Pass 2
------
encoded 1442 frames, 19.68 fps, 3961.45 kb/s
encoded 1442 frames, 19.75 fps, 3961.31 kb/s
encoded 1442 frames, 19.67 fps, 3961.56 kb/s
encoded 1442 frames, 19.69 fps, 3960.98 kb/s

7zip:

TrueCrypt:

POV-Ray:

Looks like the Phenom 2 has a slight edge of first gen Core2 Quad at the exact same clock speed. I wonder if Penryn completely closes the gap? I am guessing it does, along with its more typical DDR3.

 

[Makaveli]

Just finished the 7zip one I went for a slighly different look with my graph as it has the compress and decompress listed instead of the totals.