Some early six core + HT numbers..

T

Tobit

[H]ard|DCer of the Month - March 2010/May 2011
Joined
Dec 12, 2005
Messages
5,145
BrentPresley over at FoldingForums just posted some numbers from of his new six core with HT build.

Got a goodie to test for a little while. Latest Linux kernel (2.6.33) has better scaling w/ 12 threads.

Project #: 2683
Average time/frame: 27:09 Mins
CPU: X5670 B0 @ 3.3 GHz (yes, I know, not an impressive OC, but the B0s have some significant bugs limiting OCs)
# of CPU sockets: 1
# of cores: 12 (6 "real", 6 "virtual")
# of fahCore_A2 processes running: 12

RAM installed: 6GB
RAM used by FAH: 3.4GB
OS / Linux kernel: Ubuntu 9.10 - Kernel 2.6.33
Click to expand...
 
27 TPF = 34,200 PPD

Not too bad at 3.3 Ghz, imagine it at 4.0....
If it scales well... that means around 22:30 TPF , if that's right, 45k PPD!

What would the draw be at 4.0? It's gotta be less than the GPU equivalent of 45k, that's 4 GX2's!
This is why it may be worth it to get the 6cores, replace an entire GPU box with one CPU

front page has some AMD multi core goodness
...drool
 
Last edited:
Some nice numbers there...Nitro any word on what cooling was used and temps at those speeds?
 
Axdrenalin said:
Some nice numbers there...Nitro any word on what cooling was used and temps at those speeds?
Click to expand...

It was claimed to be on air, but no proof so it makes you wonder.

Although i wouldnt be too terribly surprised cause a lot of people take their 920 to 4.2 on air, and with the die shrink it is definitely a possibility.

Especially if you turn HT off. You still max the CPU without HT, so for folding purposes there is really no point in running with HT on.
 
nitrobass24 said:
Especially if you turn HT off. You still max the CPU without HT, so for folding purposes there is really no point in running with HT on.
Click to expand...
I don't see how/why some of you keep saying to turn off HT and still get the same performance. No way in heck can I turn off HT on my i7 920 and still run 30m TPFs. On a six core, how can one turn off HT and still maintain 27m frames? :confused:
 
Tobit said:
I don't see how/why some of you keep saying to turn off HT and still get the same performance. No way in heck can I turn off HT on my i7 920 and still run 30m TPFs. On a six core, how can one turn off HT and still maintain 27m frames? :confused:
Click to expand...

Have you tried it?

HT is just takin the 1 core and creating two Logical Cores....its not like you are getting any more CPU Power

The A3's are a perfect example. I get the exact same TPF with HT on and -smp 16 as i do with HT off and -smp 8
 
nitrobass24 said:
Have you tried it?
Click to expand...
Oh, I see what you are saying. Turn off HT but still launch eight threads. Gotcha. No, I haven't tried this. However, I will make it a point to when my new motherboard arrives. :rolleyes:
 
Thats the case on my setup that i have observed but i dont have an i7 to test with.
 
Very COOL :cool:

but just like the i7 920... wait a while tell they get the stepping corrected so you can OC and not melt your cooler
 
nitrobass24 said:
Have you tried it?

HT is just takin the 1 core and creating two Logical Cores....its not like you are getting any more CPU Power

The A3's are a perfect example. I get the exact same TPF with HT on and -smp 16 as i do with HT off and -smp 8
Click to expand...

You're basing all of this on your Dual Xeon setup though, where you have 8 physical cores to play with. Using a single i7, you would have to enable HT in order to be able to use the -smp 8 flag, otherwise it would only detect four cores total, correct? With this logic, there is no way to be able to process -bigadv units on just the four cores so HT must be enabled.
 
Axdrenalin said:
You're basing all of this on your Dual Xeon setup though, where you have 8 physical cores to play with. Using a single i7, you would have to enable HT in order to be able to use the -smp 8 flag, otherwise it would only detect four cores total, correct? With this logic, there is no way to be able to process -bigadv units on just the four cores so HT must be enabled.
Click to expand...

your missing the point though.

Yes it would be -SMP 4, but its not like you are disabling half the cores.
i7s only have 4 cores, and when you turn on HT it just creates logical cores(think disk partitions).

You are just running 8 threads to max out 4 cores, when all you really need is to run 4 threads to max out 4 cores.
Yes i am basing this off of my Xeon setup, but wouldnt something like this be even more exaggerated on a larger setup like mine? And I am not seeing a difference at all.
 
Axdrenalin said:
Using a single i7, you would have to enable HT in order to be able to use the -smp 8 flag, otherwise it would only detect four cores total, correct?
Click to expand...
By using the -smp 8 flag to begin with, you are telling the client to ignore the automatic core detection built into the client and force it to spawn 8 threads. There is no technical limitation with the client preventing 8 threads to be spawned on a quad core.
 
nitrobass24 said:
your missing the point though.

Yes it would be -SMP 4, but its not like you are disabling half the cores.
i7s only have 4 cores, and when you turn on HT it just creates logical cores(think disk partitions).

You are just running 8 threads to max out 4 cores, when all you really need is to run 4 threads to max out 4 cores.
Yes i am basing this off of my Xeon setup, but wouldnt something like this be even more exaggerated on a larger setup like mine? And I am not seeing a difference at all.
Click to expand...
I know that when people were testing normal A2 units on i7 CPUs when they first came out, they were seeing much higher PPD numbers from running -smp 8 with HT enabled than they were while running -smp 4 with HT off.
 
nitrobass24 said:
your missing the point though.

Yes it would be -SMP 4, but its not like you are disabling half the cores.
i7s only have 4 cores, and when you turn on HT it just creates logical cores(think disk partitions).

You are just running 8 threads to max out 4 cores, when all you really need is to run 4 threads to max out 4 cores.
Yes i am basing this off of my Xeon setup, but wouldnt something like this be even more exaggerated on a larger setup like mine? And I am not seeing a difference at all.
Click to expand...

Tobit said:
By using the -smp 8 flag to begin with, you are telling the client to ignore the automatic core detection built into the client and force it to spawn 8 threads. There is no technical limitation with the client preventing 8 threads to be spawned on a quad core.
Click to expand...

I understand all of this, I'm just wondering how well this would do on a -bigadv unit if you were only using the -smp 4 flag. For some reason I was under the impression that just using the 4 physical cores wouldn't be enough to complete it on time. ;)
 
Zero82z said:
I know that when people were testing normal A2 units on i7 CPUs when they first came out, they were seeing much higher PPD numbers from running -smp 8 with HT enabled than they were while running -smp 4 with HT off.
Click to expand...

Yea that may be the case. Thats what im trying to figure out;)
 
nitrobass24 said:
Yea that may be the case. Thats what im trying to figure out;)
Click to expand...
I'll test your theory for you tomorrow when my new board arrives. :cool: Thankfully, I have two other SMP clients up and crunching A3's against my passkey to keep my 80% return rate up in the event I feck up a couple bigadvs. :rolleyes:
 
Tobit said:
I'll test your theory for you tomorrow when my new board arrives. :cool: Thankfully, I have two other SMP clients up and crunching A3's against my passkey to keep my 80% return rate up in the event I feck up a couple bigadvs. :rolleyes:
Click to expand...
Keep us posted, I am curious if HT on is faster.
 
Air Cooled 420 said:
Keep us posted, I am curious if HT on is faster.
Click to expand...

When running the A2/A1 cores, HT does help out a whole lot.

When running the A3, it becomes a little harder to answer.
Currently, A3 2.14 / 2.15, there is a benefit on certain projects.

Each time a new Version of A3 is released, we will need to re-evaluate that to determine the benefit.
 
If I remember correctly, the whole theory behind hyperthreading is that the chip creates a logical processor core for each physical core to enable a second pipeline of instructions and data that can be inserted into the processing pipeline wherever there is idle time (where an instruction is fetching data from memory, for instance, instead of performing a computation) to make the processor be able to more efficiently pipeline the instructions to do the most useful work in a given clock cycle.

If the F@H computation engine is already highly efficient and doesn't leave much idle time per cycle, then HT won't be able to increase the efficiency much. However, on more standard sets of instructions that aren't highly optimized (or do more memory-intensive operations where there are processing delays while the data is fetched from cache or main memory) you may see more benefit from HT.

As far as bigadv goes, I thought you had to have at least 8 logical cores of some type to even run the units. Is that not the case?
 
mjannusch said:
As far as bigadv goes, I thought you had to have at least 8 logical cores of some type to even run the units. Is that not the case?
Click to expand...
No, the -bigadv and -smp 8 flags is the only thing that tells the server to go ahead assign the WUs. As I have stated before, the -smp 8 flag bypasses all automatic core detection built into the client. There have been many documented cases of quad core users attempting to run bigadv work. Some of them, with extremely high overclocks (4.25GHz and faster), can actually complete them in time as well.
 
Tobit said:
Some of them, with extremely high overclocks (4.25GHz and faster), can actually complete them in time as well.
Click to expand...

...and by saying this, do you mean they are completing them on time (within the specified parameters than Stanford dictates), or are they completing them with plenty of time left over to gain a healthy bonus from it? My i7-860's OC'ed to 3.7 are pulling frame times of 33:50 with HT active. I'm wondering (out loud ;) )just what kind of changes in speed I can expect if I run the MB with HT disabled and just a straight 4 cores. I'm also thinking the -smp 8 flag can't be used if there aren't at least 8 (logical) cores detected. I know this contradicts what you mentioned before, but I'm thinking I tried this once and it failed. If that's the case, then how are they managing to get a bigadv unit downloaded in the first place to run on a straight quad without HT enabled?

On the surface I understand the premise...it's the finer details that are a little vague to me...
 
You can download a bigadv on a regular quad, theres nothing preventing that.
Yes you can run 8 threads on 4 cores.....but why would you?
 
Axdrenalin said:
...and by saying this, do you mean they are completing them on time (within the specified parameters than Stanford dictates), or are they completing them with plenty of time left over to gain a healthy bonus from it?
Click to expand...
Let me try and find some numbers for you. I know of one guy running bigadv on a Q9650 but I will have to dig through some threads to find his frame times.

Axdrenalin said:
I'm also thinking the -smp 8 flag can't be used if there aren't at least 8 (logical) cores detected. I know this contradicts what you mentioned before, but I'm thinking I tried this once and it failed.
Click to expand...
I don't know why it failed for you in the past but it works as I had previously stated.
 
Axdrenalin said:
...and by saying this, do you mean they are completing them on time (within the specified parameters than Stanford dictates), or are they completing them with plenty of time left over to gain a healthy bonus from it?
Click to expand...
I found the numbers I was thinking of and, compared to A3 and even A2 PpD, I wouldn't call it healthy. No. However, it can be done. The person I am referring to ran bigadv on his Q9650 @ 4GHz and received a total of 49,434 points. This roughly equates to a TPF of 45m 38s or 15.5K PpD. Not really efficient at all but it does meet the preferred deadline. :rolleyes: I think he only ran it once or twice to prove that it could be done. He is now back running A3's and A2's.
 
So... Back OT, anyone know of a street date for these yet? Street price? Any other tidbits?
Who's gonna be the first to do a dually with these?
Posted via [H] Mobile Device
 
Well I'm getting 27:54 TPF on -bigadv now on Nexus.

This weekend when I hope to get a different distro on that might support turboboost I hope to make it a flat 27 or better.

But from one chip, this is impressive. And it should have more headroom for OC-ing.

It's just not going to be cheap.
 
Wasn't estimates at like 550 for the 970?
Posted via [H] Mobile Device
 
Tobit said:
The person I am referring to ran bigadv on his Q9650 @ 4GHz and received a total of 49,434 points. This roughly equates to a TPF of 45m 38s or 15.5K PpD. Not really efficient at all but it does meet the preferred deadline.
Click to expand...
15.5k PPD is pretty good for a C2Q running at 4GHz. Certainly better than you would see from A3 units.
 
Zero82z said:
15.5k PPD is pretty good for a C2Q running at 4GHz. Certainly better than you would see from A3 units.
Click to expand...
Yeah, you are right. I must have picked up the wrong crack pipe this afternoon. :confused:
 
*waits for numbers on AMD's new six core CPU* :D
 
There's a guy selling six of them on eBay for 225 a pop, not a bad deal considering you can make lower offers and shipping is free!
Posted via [H] Mobile Device
 
I read a little about the new 6100 series AMD server chips. 8 and 12 core variants in Multi-chip-modules in up to 4P sockets. Obviously you can't run VMs because the cores are clocked around 2 Ghz, but theres so damn many of them. the lowest 12-cores are rumored to be around ~$940 each and are clocked at 1.9 ghz with 4 channels of DDR3 available to each chip.

$1880 2x AMD 6168 CPUs 115W TDP
$400 2P motherboard
$600 8x 2GB memory modules
$100 500W PSU

=$3000 for a 24-core system that should run on less than half a kilowatt of power. Based on a 2.2 Ghz dual 6-core istanbul-core making 34:28 minute frame times, this rig at 1.9 Ghz should make ~20 minute frame times if it all scales perfectly across cores. 50,000 ppd from 500W? Do these numbers make sense to the rest of you or am I on the wrong crack-pipe as well?

ont he other hand if you compare to the 32-core 8P system using Opteron 8347s at 1.9 Ghz, it is getting frame times of 22:41 with 8 more cores according to http://foldingforum.org/viewtopic.php?f=55&t=11314. I would assume that you gain some scaling when you are only dealing with 2 sockets instead of 8.
 
Last edited:
no, I've been having some serious memory issues, but I think I may have tracked down the problem in another forums, I think it's related to a bios issue, since I've got a few hours on my hands right now I'll go downstairs and see what I can get done on it. I don't think I'm going to get the add on board, the add on is 250% more expensive than the damn mobo itself!!

What I think I'll do is sell of the other 4 and then see about upgrading to some Istanbul's in the future, once AMD releases their 8 and 12 core CPUs I'm assuming the 6 core market will free up a bit. Right now I can get them for under 225 a piece, but I don't have that kind of money to throw into a project that I can't even get stable yet!

What I'm thinking is by selling the rest of my spare mobo's (I've already sold all but 3 of my GPU's and I'm keeping the rest) and my spare opterons I'll have enough for another bigadv box, that should give me 2 i7 boxes, but not buy the CPU yet. Once Fermi and 6 cores hit, sell the last of my GPU's and my dual 295's, buy one Fermi card and one 6 core i7. By the summer I hope to have

1 Quad i7 (Pure SMP rig)
1 6core i7 (Pure SMP rig)
and my gamer rig, a trusty old Q66 with a Fermi card and my last GPU (something small, probly my 8800 GT, I'll sell the last GX2s at this point) for Physx

Should be able to get 100k-115k PPD from a simple farm like that, and keep power draw just below where I'm at now.

I've already got cases, PSU's, 8400 GS GPU's, and a rack, I'm just a few components away from having a decent sized farm.
 
Vaulter98c is sounds like your the man with the plane. what you just discribed is also where im hoping to be in 6 months or soo
 
You must log in or register to reply here.
Back
Top