backtop


Print 84 comment(s) - last by saratoga.. on Feb 2 at 6:27 PM

More "Penryn" details emerge





Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: Until there is proper software, HT is over rated
By Phynaz on 1/31/2007 11:51:06 AM , Rating: 3
quote:
On a multicore chip, I can't see HT doing very much for performance until CSI and Nehalem is released in 2008/9 (remember that the bottleneck nere is the FSB).


Sigh...

I'ts been shown again and again that there is no FSB bottleneck.

It would be really refreshing if you quit spreading your misinformation.


By Griswold on 1/31/2007 2:08:55 PM , Rating: 2
In multi-socket systems, thanks to cache coherency traffic that has to go over FSB, there seems to be your non-existant bottleneck. It materializes in mediocre scalability.

See also:
http://www.anandtech.com/IT/showdoc.aspx?i=2897&p=...

This may not be relevant to somebody like you, but to others it is.


RE: Until there is proper software, HT is over rated
By Phynaz on 1/31/2007 2:13:12 PM , Rating: 2
Backup what you are saying. State how much bandwidth cache coherency traffic takes up on the FSB.


By saratoga on 1/31/2007 3:23:30 PM , Rating: 2
quote:
Backup what you are saying.


He did . . .

Let me guess, you didn't actually read the link?

quote:
State how much bandwidth cache coherency traffic takes up on the FSB.


That depends on the load chosen. Theres no one number, so your question doesn't even have a specific answer. You might as well ask how much memory bandwidth is enough, or many cores benchmarks use and how much cache is idea. You need to define your load, since the answer depends on what you're doing.


By Viditor on 1/31/2007 8:55:06 PM , Rating: 2
Phynaz, I don't know if you realize it but you've asked this question before in another thread.
There, we were talking about scalability...
Here was my reply to your request for proof on that thread:

"Here is a first indication that quad core Xeon does not scale as well as the other systems. Two 2.4GHz Opteron 880 processors are as fast as one Xeon 5345, but four Opterons outperform the dual quad core Xeon by 16%. In other words, the quad Opteron system scales 31% better than the Xeon system"
http://tinyurl.com/2cgnj8

Please note that as you add cores and sockets that use the FSB, in other words scale, the relative performance of the Opteron gains dramatically.
If you think about it you'd realize that since the cores haven't changed, only the load on the FSB could account for this.


By coldpower27 on 2/1/2007 9:03:55 AM , Rating: 2
Interesting, this problem doesn't seem to manifest itself till you reach the 2P System with Dual Clovertown's.

Now I can understand how AMD can claim the 40% improvement over Clovertown, with Barcelona "in a wide variety of workloads".

It will be interesting to see how much this increase diminishes, assuming AMD's numbers are currently correct for Bareclona vs Clovertown performance, when you make the comparison with the Agena Quad Core to the Kentsfield Quad in Single Socket Desktop systems.

The FSB issue isn't a issue on a Single Socket, however it seems Clovertown suffers from poorer scaling as you increase the number of Sockets.

So overall is the FSB an issue, not on the Single Socket Arena, when you are talking 2 Sockets or more there is something that is weakening Clovertown's scaling ability. It could also be because Clovertown has to work with FB-DIMM technology which is of higher latency compared to Unbuffered DDR2 on the desktop.


By Viditor on 2/1/2007 12:33:32 PM , Rating: 2
Good points CP.
I agree that AMD probably chose Cloverton as a comparison very carefully, and 40% really isn't unbeleivable on this specific comparison .
One thing I've been saying all along is that if the K10 core is only equivalent to C2D, then AMD will have a much better spec because of the platform. Of course I was most incomplete in my comments...
I should have qualified that I was speaking of servers...mainly because of the scaling.

As to FBDs being an issue, there is something to that...but it seems to me that the latency doesn't come close to accounting for the large difference in scaling.
I am still wondering why Intel went with a high latency/high bandwidth model for memory instead of a low latency/low bandwidth one...


By Phynaz on 2/1/2007 1:14:59 PM , Rating: 2
K10?

Ummmm....This is K8L we're talking about, not some fanciful chip that's always five years away.


By Viditor on 2/1/2007 1:30:11 PM , Rating: 2
quote:
Ummmm....This is K8L we're talking about, not some fanciful chip that's always five years away


Ummm...there is no K8L. The next-gen chip coming out next quarter (Barcelona) is a K10 chip...

"Again, AMD has explicitly told me its native quad-core chips will be K10, not K8. That's from their Technical Director - Sales and Marketing EMEA, so isn't likely to be wrong"
http://forums.hexus.net/showthread.php?t=92137&pag...

But a "Rose by any other name"...let's just call it Barcelona.


By Phynaz on 2/1/2007 10:33:07 AM , Rating: 2
Yes, I have asked this question before.

I hardly consider one article from Anandtech to be proof of anything.

For example, where in the article is empirical evidence that the FSB has become saturated or is a bottleneck?

Answer: There isn't any, it's all conjecture.


By Viditor on 2/1/2007 12:17:51 PM , Rating: 2
quote:
where in the article is empirical evidence that the FSB has become saturated or is a bottleneck?
Answer: There isn't any, it's all conjecture


Well no actually, it's a hypothesis backed by scientific data that corroborates the conclusion.

It might be helpful if instead of us taking your word for it, you could offer some indication of substance for your assertion that:
"I'ts been shown again and again that there is no FSB bottleneck"


By Phynaz on 2/1/2007 1:13:30 PM , Rating: 2
Yeah, prove a negative, that will work.

Please point me to this scientific evidence you speak about.

Even though I'm not the one making the statements (you are), I'll give you my evidence.

Using Intel analysis tools, running business applications, I see FSB utilization in the 15% percent area on a dual core HP system. Running the same on a quad bumps the utilization to 18%-20%. Business applications run out of cpu long before then run out of FSB bandwidth.



By Viditor on 2/1/2007 1:38:14 PM , Rating: 2
quote:
Yeah, prove a negative, that will work

Sigh...it's not proving a negative, it's demonstrating that data throughput on a FSB modeled system is equivalent to an HT modeled system.

I remind you that you are the one who said:

"I'ts been shown again and again that there is no FSB bottleneck"

All I'm asking for is for you to show that...seems reasonable to me.


By saratoga on 2/2/2007 6:24:55 PM , Rating: 2
quote:
Please point me to this scientific evidence you speak about.


What kind of logic is this? You said there was evidence, so you provide it. Don't say it exists and then expect other people to find it for you.

quote:
Even though I'm not the one making the statements (you are)


You mean "making claims". This whole post of yours is a statement.

Anyway, remember when you posted this:

"I'ts been shown again and again that there is no FSB bottleneck. "

So you did make a claim. Now back it up or retract it.

quote:
Using Intel analysis tools, running business applications, I see FSB utilization in the 15% percent area on a dual core HP system. Running the same on a quad bumps the utilization to 18%-20%. Business applications run out of cpu long before then run out of FSB bandwidth.


All that proves is that your specific business app isn't constrained. No one even doubted that. Rather, your claim that the Core 2 was not limited on all workloads is in doubt. Unless you've some reason to think its relevant to the rest of the world, theres no sense in even mentioning it.


"People Don't Respect Confidentiality in This Industry" -- Sony Computer Entertainment of America President and CEO Jack Tretton










botimage
Copyright 2016 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki