Official AMD Barcelona Thread

Printable View

Show 100 post(s) from this thread on one page

07-06-2007, 07:54 PM
informal

Slow(er) and inferior?K10? Are you joking?
07-06-2007, 08:16 PM
Mav451

Quote:

Originally Posted by mstp2009

I've actually READ the K10 NDA (probably unlike you). The breach of NDA is a $10 million fine for an individual and up to a $1 billion for a corporation.

Well at least they went for a nice round number.
07-07-2007, 12:44 AM
generics_user

Quote:

Originally Posted by Periander6

You aren't bursting any bubbles.

You want to pretend that C2 doesn't have a 20% IPC advantage over K8, suit youself. The rest of the world gets results like the ones Xbit, Techreport, Anand and every other competent testing outfit gets, 20% or so, if anything that is being generous to the K8.

Sun is a seller of complete systems, and are still struggling. If they tried to make it just as maker of Sparc chips, they would be out of business in weeks. Absurd to compare them to AMD.

Yes, people can browse the web and read e-mail with inferior CPUs. The problem is that AMD has to make enough of a profit to make payments on their 5+ billions in debt, fund chip development, equip their fabs ect.. Slower, inferior chips don't fetch much of a price. Not nearly enough to keep AMD from sinking.

i'm with nn_step on this one, yes C2D got an 20% IPC advantage but they still WIN in SOME apps (I DO NOT TALK ABOUT THE MAJORITY APPS, but in some apps amd can keep up with C2D, and in 64bit the disadvantage is less significant ;) )

for gaming: i doubt that you see much of a difference when using mid range graphics cards in systems for up to 1000$ ;) , but in the higher and amd got no chance against intel :yepp:
07-07-2007, 04:43 AM
red

Quote:

Originally Posted by mstp2009

I've actually READ the K10 NDA (probably unlike you). The breach of NDA is a $10 million fine for an individual and up to a $1 billion for a corporation.

Um, I didn't sign anything and I just found it off Google so I'm fine right :D
07-07-2007, 04:51 AM
red

Quote:

Originally Posted by nn_step

He did kinda skip Slide 42, which effectively means double Virtualization performance. Which would relatively put a single 1Ghz Barcelona on even performance with a Quad core Kentsfield running at 8Ghz (virtualization applications and Operating systems only)

No one bothered with this one? For some reason :rolleyes: , I'm inclined to think that these numbers are poo, but if K10 is double performance, a 1 core 1GHz K10 matches a 4 core 8GHz Kent, by deduction, a 1 core 2GHz K8 matches a 4 core 8GHz Kent, therefore, increasing AMD's advantage from 16x to 32x. If a current 16x advantage isn't compelling enough to save AMD, what makes you think a 32x one will :rofl:
07-07-2007, 06:41 AM
savantu

Quote:

Originally Posted by mstp2009

BS. On average, C2D is 20% faster than K8 in 32-bit code and 14% faster in 64-bit code.

And the corollary is also true. In some things C2D is far more than 20% faster than K8. I know that in Folding@Home it is about 70-80% faster because of the SSE optimizations to that program.

Based on what ?
07-07-2007, 07:39 AM
mstp2009

Quote:

Originally Posted by savantu

Based on what ?

Have you been under a rock for the last 15 months?

http://www.xbitlabs.com/articles/cpu...w_9.html#sect0
http://www.xbitlabs.com/articles/cpu...review_11.html
http://setiathome.berkeley.edu/forum...d.php?id=39379
http://www.tomshardware.com/2006/07/...out_athlon_64/
http://www.extremetech.com/article2/...2014685,00.asp
http://blogs.msdn.com/rickbrew/archi...13/664890.aspx
http://www.techreport.com/reviews/20...s/index.x?pg=1

EDIT - 64-bit analysis here:
http://www.xbitlabs.com/articles/cpu...duo-64bit.html

Google 101: Core 2 Duo vs. AMD Athlon X2
07-07-2007, 07:52 AM
savantu

Quote:

Originally Posted by mstp2009

Have you been under a rock for the last 15 months?

EDIT - 64-bit analysis here:
http://www.xbitlabs.com/articles/cpu...duo-64bit.html

Google 101: Core 2 Duo vs. AMD Athlon X2

The 64bit part interested me.I knew you would point out the Xbit test.Then why did I bother to ask ?

Because the conclusion is flawed IMO.

When you average results , you need to eliminate anomalies , like the one provided by ScienceMark , because they skew your results.

ScienceMark is a test where K8 does abnormaly well ; in the same way SuperPI is a test where C2D does abnormaly well.Including them has a profound effect on the average.

Once you eliminate ScienceMark , the results are damn close.The graphic at the end shows this best.
07-07-2007, 08:00 AM
informal

Savantu,then you need to check the techreport x64 testing,where 6000+ gets really close on many occasions to e6700.
07-07-2007, 08:29 AM
madcho

The big advantage of the C2D is the SSE. K10 have better SSE ( full 128bits, C2D not all ).

if you disable the SSE on a C2D and a K8, K8 would be faster.

64bits mode use SSE, AMD don't loos too much here with the lot of GPRs in the K8. K10 will perform more fast in 64bits mode, with 8 more GPRs and double SSE.

K10 is a bit more fast in integer, but not a big improvement. about 15%.

Phenom will be a bit faster than barcelona clock for clock 8-10% i think, with 1066 non registered memory ( barcelona 800 registered ), and HT3 could be usefull for 4 cores.
07-07-2007, 08:46 AM
mstp2009

Quote:

Originally Posted by madcho

The big advantage of the C2D is the SSE. K10 have better SSE ( full 128bits, C2D not all ).

Wrong. C2D is fully 128bit SSE (on all 3 SSE decoders).

Quote:

Originally Posted by madcho

if you disable the SSE on a C2D and a K8, K8 would be faster.

Wrong. The integer unit in C2D is MUCH better than that in K8.

Quote:

Originally Posted by madcho

Phenom will be a bit faster than barcelona clock for clock 8-10% i think, with 1066 non registered memory ( barcelona 800 registered ), and HT3 could be usefull for 4 cores.

Do you have a clue what you are talking about? Phenom IS BARCELONA (same core, just different sockets - desktop vs. server).
07-07-2007, 09:18 AM
BrowncoatGR

Quote:

Originally Posted by mstp2009

Wrong. C2D is fully 128bit SSE (on all 3 SSE decoders).

Wrong. The integer unit in C2D is MUCH better than that in K8.

Do you have a clue what you are talking about? Phenom IS BARCELONA (same core, just different sockets - desktop vs. server).

While i dont think a 10% IPC for Phenom over Barcelona is reasonable these are not the same processors for different sockets. Barcelona is a HT2 cpu using up to DDR2 5300 registered memory whereas Phenom is a HT3 cpu using up to DDR2 8500, which gives Phenoma huge advantage in Memory bandwidth. Whether Phenom can use the increased memory bandwidth to get improved performance remains anyone's guess right now but id say it will at least give it a slight advantage clock for clock.
07-07-2007, 09:35 AM
mstp2009

Quote:

Originally Posted by BrowncoatGR

While i dont think a 10% IPC for Phenom over Barcelona is reasonable these are not the same processors for different sockets. Barcelona is a HT2 cpu using up to DDR2 5300 registered memory whereas Phenom is a HT3 cpu using up to DDR2 8500, which gives Phenoma huge advantage in Memory bandwidth. Whether Phenom can use the increased memory bandwidth to get improved performance remains anyone's guess right now but id say it will at least give it a slight advantage clock for clock.

It's all a moot point for the desktop.

What can fully take up that memory bandwith NOW on a desktop app?

Nothing.

It isn't until you get to 4P systems in a server environment with HEAVY database transaction loads that you see any difference.

This is exactly why we see no performance gain going from Socket 939 (DDR1) to Socket AM2. The CPU core didn't change except for the memory controller. Bench identically clocked systems and you will see that all that extra memory bandwidth is going to waste.

It will be the same thing in Phenom w/ HT2 vs. HT3. This is pointless for desktop apps.
07-07-2007, 09:39 AM
Jacky

Quote:

Originally Posted by BrowncoatGR

While i dont think a 10% IPC for Phenom over Barcelona is reasonable these are not the same processors for different sockets. Barcelona is a HT2 cpu using up to DDR2 5300 registered memory whereas Phenom is a HT3 cpu using up to DDR2 8500, which gives Phenoma huge advantage in Memory bandwidth. Whether Phenom can use the increased memory bandwidth to get improved performance remains anyone's guess right now but id say it will at least give it a slight advantage clock for clock.

there won't be any surprise. it will be about equal to the advantage core 2 gets when the fsb is upped from 1066mhz->1333mhz.. 2-3%, maybe 5% (core 2 isn't really memory starved at all)

Quote:

Originally Posted by madcho

Phenom will be a bit faster than barcelona clock for clock 8-10% i think, with 1066 non registered memory ( barcelona 800 registered ), and HT3 could be usefull for 4 cores.

Quote:

Originally Posted by madcho

if you disable the SSE on a C2D and a K8, K8 would be faster.

please stop the FUD and write decent english
07-07-2007, 11:10 AM
BrowncoatGR

Jacky well id say that is both application dependent and really difficult to predict right now. K8 benefits quite a bit from PC6400 memory in some applications and my guess is so will K10 from both PC8500 memory and the increased HT speed. Although in some applications the advantage will be close to zero. I expect the biggest impact to be in games where consumer apps are concerned. BTW imo Core2 could use additional memory bandwidth, in particular quadcore parts, but increasing the fsb does not have that much of an impact because of the limitations of the FSB bus. Im also guessing Core2 prefetching and cache structure was also optimized with a 1066 FSB bus in mind. Penryn will probably change this though.

Madcho C2D clearly hasa faster integer unit than K8, there is really no discussion here. But K8 is multitudes faster at x87 fp code but that is not really relevant for desktop apps. What allows K8 to somewhat keep up is its memory bandwidth mostly. And in 64bit macro-ops fusion does not work.
07-07-2007, 11:25 AM
Zytek_Fan

K10 won't be slower or inferior to K8. It seems a lot like Core, so it should at least be on par. The extra bandwidth should help AMD a bit.
07-07-2007, 11:31 AM
BrowncoatGR

Quote:

Originally Posted by Zytek_Fan

K10 won't be slower or inferior to K8. It seems a lot like Core, so it should at least be on par. The extra bandwidth should help AMD a bit.

If that was at me, I was referring at the impact increased memory bandwidth could have for K10. I was not comparing K8 to K10
07-07-2007, 11:31 AM
accord99

Quote:

Originally Posted by informal

Savantu,then you need to check the techreport x64 testing,where 6000+ gets really close on many occasions to e6700.

And the many times where it loses to the E6600.

Quote:

Originally Posted by mstp2009

EDIT - 64-bit analysis here:
http://www.xbitlabs.com/articles/cpu...duo-64bit.html

Google 101: Core 2 Duo vs. AMD Athlon X2

Xbitlabs made a mistake in their simple average. Going to 64-bit, the X6800 ends up having the same % advantage as in 32-bit.

http://img85.imageshack.us/img85/32/...4bitrunol4.jpg
07-07-2007, 12:07 PM
Zytek_Fan

Quote:

Originally Posted by BrowncoatGR

If that was at me, I was referring at the impact increased memory bandwidth could have for K10. I was not comparing K8 to K10

No, not at you.
07-07-2007, 12:15 PM
nn_step

Since I just love throwing wrenches into the works
http://www.silentpcreview.com/article735-page1.html
07-07-2007, 12:52 PM
lapdog

Quote:

Originally Posted by nn_step

Since I just love throwing wrenches into the works
http://www.silentpcreview.com/article735-page1.html

What in tarnation does that have to do with Barcelona?
07-07-2007, 01:40 PM
madcho

Quote:

Originally Posted by mstp2009

Wrong. C2D is fully 128bit SSE (on all 3 SSE decoders).

not true. the units are 128bits but some functions are not available in 128bits

Quote:

Wrong. The integer unit in C2D is MUCH better than that in K8.

yes.

Quote:

Do you have a clue what you are talking about? Phenom IS BARCELONA (same core, just different sockets - desktop vs. server).

The new IMC of the K10 will allow to use more bandwith, not like the K8 on AM2. The K8 on AM2 is not able to use the full bandwith of the DDR2, and cache speed is too slow.

Registered memory is slow, and have bad timings, Barcelona will not able to show all. Phenom will be a bit faster. HT3 will improve a bit for 4 cores. I'have seen with 1ghz HT with dual core a bottleneck. ;)
07-07-2007, 03:30 PM
savantu

Quote:

Originally Posted by informal

Savantu,then you need to check the techreport x64 testing,where 6000+ gets really close on many occasions to e6700.

We were discussing how much each CPU gains from 32bit -> 64bit , not absolute performance.

Secondly , too bad Xbit didn't include how much P4 gained from 32bit to 64bit.I wouldn't be surprised too see it exceeds both C2D and K8 according to their tests.I'll calculate this tomorrow , it's 2 am now and I need to sleep.
07-07-2007, 03:34 PM
savantu

Quote:

Originally Posted by madcho

not true. the units are 128bits but some functions are not available in 128bits

Really ? Which ones aren't available for C2D but are for K10 ?

Quote:

The new IMC of the K10 will allow to use more bandwith, not like the K8 on AM2. The K8 on AM2 is not able to use the full bandwith of the DDR2, and cache speed is too slow.

Well , K10 introduces another cache in the equation , an L3 with unknown latencies thus far.

Quote:

Registered memory is slow, and have bad timings, Barcelona will not able to show all. Phenom will be a bit faster. HT3 will improve a bit for 4 cores. I'have seen with 1ghz HT with dual core a bottleneck. ;)

In what tests did you notice a performance penalty due to the 1GHz HT link ?
07-07-2007, 05:26 PM
informal

Quote:

Originally Posted by mstp2009

Oh yeah, and your posts just overflow with intelligent discourse. :rofl:

Quick search on your posts reveals,well nothing interesting(except extreme intel bias).But that's nothing new these days.

On topic:

Article from tgdaily sums it up nicely:
http://www.tgdaily.com/content/view/32802/135/

Quote:

AMD feels “confident” about Barcelona performance estimates PDF Print E-mail
Hardware
By Wolfgang Gruener
Friday, July 06, 2007 13:04
Recommend article:
Slashdot
Digg
Delicious
Technorati
YahooMyWeb
Chicago (IL) – Two Barcelona benchmark charts published on AMD’s website caused quite a stir yesterday: Intel wasn’t happy about outdated performance numbers causing AMD to remove the charts.

In Pictures: AMD Quad-Core "Barcelona" technical details ...
AMD today in fact acknowledged that its performance charts were outdated and that new Intel numbers were not included in the comparison. Spokesman Phil Hughes conceded in a conversation with TG Daily that AMD “failed to update the numbers” and has taken down the charts as a result.

In the meantime, Intel had stressed that Xeon performance numbers, which far exceed the numbers used by AMD, are available. And according to Intel, the 2.66 GHz Xeon 5355 does not trail the estimated integer performance of a 2.6 GHz quad-core Opteron by about 20%, but by only about 2% (99.9 vs. 102 points in the Specint_rate2006 benchmark). The 3.0 GHz Xeon 5365, which is shipping in limited numbers at this time, is actually slightly faster than the 2.6 GHz Opteron (106 vs. 102 points) – which is expected to be AMD’s fastest quad-core to be available by the end of this year.

From that perspective, the future does not look favorable for AMD, given the fact that Intel’s 45 nm server chip will be available around the turn of the year and Intel already said that it does expect substantial performance gains in the double-digit range across the board. But the truth of the matter is that we are talking about unreleased products on both ends, with Intel holding the performance lead in products that you can buy today.

So, if AMD has taken down the benchmark numbers, does that mean that Barcelona will not be as fast as AMD claims? Hughes told us that AMD still expects the performance of Barcelona at 2.6 GHz to come in at what the estimate has shown – or “even a bit higher”. However, there is no doubt that the initially available Barcelonas (with up to 2.0 GHz) will not be able to match Intel’s current performance level.

Interestingly enough, plain speed isn’t really the whole story. Power consumption is a performance factor that has not received a whole lot attention. In the light of today’s server environments that try to decrease heat and use the available power as efficiently as possible, this may actually the discipline where the battle between Intel and AMD will be fought. At least according to AMD’s Barcelona website, the firm’s 68w/95w/125w chips will have an advantage over comparable Xeons (65w/80w), at least as long as 65w Xeons are pitched against 68w and 95w Opterons and the 80w Xeon against the 95w Opteron. And yes, the low-power Clovertown processors (50w) were left out of consideration in this comparison.

If industry rumors are right, then AMD will be releasing Barcelona in September and we soon will be able to verify what the processor is really capable of and what impact it may have in the market.

Show 100 post(s) from this thread on one page