Aceshardware

AtWork · Joined: 31 Jul 2007 Posts: 116

Manufacturing date is from mid April -

http://www.xtremesystems.org/forums/showthread.php?t=194436

Here's one comment from the posts:
It utterly destroyed K8 in SPi .Around ~17.5% faster than K8.Deneb brought us a lot more than Agena did in this particular test.
So all they needed was 3.48Ghz to break a WR that needed almost 4Ghz 65nm K10 chip to do it .
Being an ES still means this is not quite "tweaked" from BIOS perspective.Par this thing up with upcoming SB750 motherboard and 2 4870X2 card,what you get is all-amd-uber-system .K10.5 is starting to look a lot like RV770

JumpingJack · Joined: 05 Oct 2007 Posts: 112

The more impressive feat is the clock speed, highest OC yet on an AMD quad (well, a few exceptions of LC OCs)....

For comparision, here is a 2.5 GHz clocked yorky

Groo · Joined: 22 Jul 2007 Posts: 127

AtWork wrote:

Manufacturing date is from mid April -

http://www.xtremesystems.org/forums/showthread.php?t=194436

Here's one comment from the posts:
It utterly destroyed K8 in SPi .Around ~17.5% faster than K8.Deneb brought us a lot more than Agena did in this particular test.
So all they needed was 3.48Ghz to break a WR that needed almost 4Ghz 65nm K10 chip to do it .
Being an ES still means this is not quite "tweaked" from BIOS perspective.Par this thing up with upcoming SB750 motherboard and 2 4870X2 card,what you get is all-amd-uber-system .K10.5 is starting to look a lot like RV770

Intel has nothing to fear, but it will narrow the gap. I hear they might pull 45nm in a month or so, but still, they will be in second place until Bulldozer.

-Charlie

JoeP · Joined: 28 Jul 2007 Posts: 52

bulldozer will be that good?

Groo · Joined: 22 Jul 2007 Posts: 127

JoeP wrote:

bulldozer will be that good?

It is radically different, and IF it works, it MAY be on par with Intel. Until then there is no hope, that is what I meant.

Platform vs platform it is more even, but CPU vs CPU, Intel owns.

-Charlie

someoneelse · Joined: 13 Sep 2007 Posts: 16

45nm clocks very well for AMD. The voltages you see in a lot of the screenshots may be wrong because cpu-z detects them incorrectly.

lux_interior · Joined: 26 Jul 2007 Posts: 235

Well I hope 45nm will at least bring down the consumption of K10 CPU's, so that one day I have an interesting upgrade path from my 65nm 3600+ :-)
(not that I need it right now anyway...)

Paul DeMone · Joined: 29 Aug 2007 Posts: 376 Location: Great white north

someoneelse wrote:

45nm clocks very well for AMD.

Yeah I am sure AMD's 45 nm process is loads faster than its 65 nm
process given they stuck to a polysilicon gate stack which gave them
little room to reduce either effective gate length or dielectric thickness
without sending leakage current through the roof.

IMO *if* AMD gains a bin or two at 45 nm it will be because it is their
second kick at the can to lay out and speed path tune this turkey, not
because their 45 nm process is noticably faster than their 65 nm.

LiamC · Joined: 23 Jul 2007 Posts: 70

Groo wrote:

JoeP wrote:

bulldozer will be that good?

It is radically different, and IF it works, it MAY be on par with Intel. Until then there is no hope, that is what I meant.

Platform vs platform it is more even, but CPU vs CPU, Intel owns.

-Charlie

Can you elaborate? You may want to add IF it ships on time.

First I heard, Bulldozer was scheduled for late 2008/early 2009. Then it was pushed back to late 2009/2010. Current rumours have it late 2010. Given where Intel will be with Nehalem, add two years to Nehalems performance and Bulldozer will have to give 50% speedup over Penryns performance, or more, and keep up in clock. Awesome if it comes to pass, but can you see it happening?

Opteron · Joined: 16 Mar 2008 Posts: 44

LiamC wrote:

Can you elaborate? You may want to add IF it ships on time.
(..)Current rumours have it late 2010.

If I remember correctly, the latest AMD info on this is samples in H2/10 and chips on sale sometime in 2011 @32nm.
No rumors, official statement of an AMD guy during the last earning conference. But I wonder myself if the credibility is better than that of a rumor ^^

See the positive side, maybe AMD already adopts AVX/AVX 2.0 then.

Quote:

It utterly destroyed K8 in SPi .Around ~17.5% faster than K8.Deneb brought us a lot more than Agena did in this particular test.
So all they needed was 3.48Ghz to break a WR that needed almost 4Ghz 65nm K10 chip to do it .
Being an ES still means this is not quite "tweaked" from BIOS perspective.Par this thing up with upcoming SB750 motherboard and 2 4870X2 card,what you get is all-amd-uber-system .K10.5 is starting to look a lot like RV770

It wont be a "RV770", this is just SuperPi 1M, which memory footprint is around ~8MB. No wonder that it is faster on the K10.5 as the chip has 4MB more L3. The program is single threaded, therefore it can consume most of the L3 -> 6MB. Then there are additional 512kB from the L2, all in all not that far away from the 8 MB.

cheers

Opteron

shukla · Joined: 27 Jun 2007 Posts: 19

Paul DeMone wrote:

someoneelse wrote:

45nm clocks very well for AMD.

IMO *if* AMD gains a bin or two at 45 nm it will be because it is their
second kick at the can to lay out and speed path tune this turkey, not
because their 45 nm process is noticably faster than their 65 nm.

From benches I understand that the Phenom performance has not happened, but I'm curious as to what in the design you think is the cause?

Is there a specific aspect of the CPU design that's holding it back, pipeline? prefetch? cache (size and latency)? Or is this design just lost from an architectural standpoint.

What I also don't understand is why AMD has had such a hard time ramping clock speed. It seems that performance has increased with die shrinks but not speed.

EduardoS · Joined: 22 Mar 2008 Posts: 75

shukla wrote:

From benches I understand that the Phenom performance has not happened, but I'm curious as to what in the design you think is the cause?

Is there a specific aspect of the CPU design that's holding it back, pipeline? prefetch? cache (size and latency)? Or is this design just lost from an architectural standpoint.

What I also don't understand is why AMD has had such a hard time ramping clock speed. It seems that performance has increased with die shrinks but not speed.

Let's see if it ends with a good and healthy discussion, IMHO what hurts Phenom desktops performance most is it's cache (L3 size/latency) and integer SSE (latency/throughput), I didn't profiled it to be sure, so just a guess.

Paul DeMone · Joined: 29 Aug 2007 Posts: 376 Location: Great white north

shukla wrote:

From benches I understand that the Phenom performance has not happened, but I'm curious as to what in the design you think is the cause?.

Three basic things here:

1) very hard compare. Face it, C2D is an out of the park home run
in terms of x86 microarchitectural goodness and excellence of its
65 and 45 nm implementations.

2) AMD got very complacent about five years ago and figured they
would hold onto scalar integer and client computing performance
lead by simple inertia given Intel was sticking with FSBs for a while
longer. IMO they focussed too much on trying to attack IPF in HPC
and spent too much effort adding cache and FPU support for 4 DP
FLOPs/cycle and far too little on modernizing everything else. Intel
then surprised AMD with 4 DP FLOPs/cycle in C2D *and* a huge
increase in integer/client performance as well as a huge jump in
computational power efficiency over P4. Now, the only advantage
AMD retains, IMCs, goes away in a few quarters.

3) Part of 2 was AMD wasted years puttering around with radical
design approaches to succeed K8 that went no where. The K8L
development was a total crash program pushed on a completely
unreasonable schedule. It is a wonder that more serious bugs
didn't reach OEMs and frequency bins weren't even lower. AMD
engineers had their own Bataan death march to save Hector and
Dirk's dumb a**es but their premature and sickly baby ended up
next to the systematically and relentlessly polished C2D for the
world to compare and contrast; a real tough break. AMD has been
in headless chicken mode ever since.

Hans de Vries · Joined: 07 Aug 2007 Posts: 74

The PS3 cell processor was ported from 65nm to 45nm resulting
in either a 40% improvement in power consumption or a 20%
improvement in frequency.

This is using IBM's 45nm SOI process so the scaling for AMD's
Deneb could yield similar results.

http://www.beyond3d.com/content/news/582
http://www.trustedreviews.com/images/article/inline/6744-ibmcell.jpg

Article: Migration of the Cell Broadband Engine to 45nm SOI
http://www.edatechforum.com/journal/june2008/migration.cfm
(registration required)

Regards, Hans

Paul DeMone · Joined: 29 Aug 2007 Posts: 376 Location: Great white north

Hans de Vries wrote:

The PS3 cell processor was ported from 65nm to 45nm resulting
in either a 40% improvement in power consumption or a 20%
improvement in frequency.

This is using IBM's 45nm SOI process so the scaling for AMD's
Deneb could yield similar results.

What is the relative die area and transistor count between the 65
and 45 nm Cell?

What is the relative die area and transistor count between the 65
and 45 nm K8L?

Do you think AMD and IBM use identical processes at each feature
size?