Printable View
http://blogs.zdnet.com/Ou/?p=735
George Ou has happened upon some slides that are (were?) to be presented on September 10th.
For 4P, 2GHz K10 scores 160, losing to 8MB 1.86GHz Tigerton score of 166 though it bests the crippled 4MB 2.13GHz Tigerton.
For 2P, 2GHz K10 loses to 2GHz Clover by 4.2 points (88 vs 92.2)
What a disaster. Once the Seaburg chipset launches in Q4, Intel's 2P specINT score will also beat Barcelona clock/clock.
Tigerton runs on Clarksboro, which already has the updated snoop filter designed for quad core.
Seaburg will add this feature (vs. current Blackford).
reply to linked Ou blog:
Some interesting points.Quote:
Hi George,
The 4-processor SPECint_rate2006 slides you posted are actually GOOD news for AMD !
They show the Opteron 8350 platform ($1004/proc) being faster, or as fast as, the Xeon L7345 and E7320 platforms (respectively $2301 and $1177/proc) ! Then, the Xeon X7340 score is indeed higher than the Opteron 8350, but this Xeon costs twice as much: $1980. (It would have been interesting to see the score for the intermediate Xeon X7330 by the way.)
Everybody is so focused on absolute performance, but, in the end, all what matters is how much bang you get for your buck. So unless you stricly buy the highest-end processors ($2000+ proc representing the top 20% of the 4P market), these slides prove that Barcelona is faster & cheaper for everything else (the remaining 80% of the 4P market) !
I don't even pay attention to Ou, because all he seems to do is bash AMD.
Yep,Gearge Ou is like an intel-Sharikou version,only at the larger scale(since he writes for the "famous" ZDnet)
QFTQuote:
Originally Posted by informal
Yeah , too bad AMD changed their policy wrt to benchmark scores based on his input ; OTOH who gives a ***ss about Sharikou ?Quote:
Originally Posted by informal
This , if anything , shows your level of fanboyism , even when AMD gives credit you still beat the dead horse...:horse:
Yeah whatever,i couldn't expect more from you,the same story all over again.
If your this pi ssy now, it's going to be entertaining to see how you act come Monday.Quote:
Originally Posted by informal
Can we keep flame baiting out of this one? :rolleyes:
You do realize that these scores came directly from AMD, right?Quote:
Originally Posted by Zytek_Fan
well, didnt you know everything that has no amd in the link = bad ? :rofl:Quote:
Originally Posted by Proesterchen
assuming george ou's numbers are accurate:Quote:
Originally Posted by hawkeyefan
Intel can change the pricing sooner than later, but AMD NEED to recover R&D and their huge losses, Barcelona die is huge (it equals the clovertown die, but don't forget that Intels 65nm is more mature and that they just take two conroes and put them on one die - increasing yields. and penryn is launching very soon), they will have no chance when tigerton's 45nm successor arrives (when is this going to happen anyway? I don't think it's as early as Q1'08) or when Intel decides to lower their margins a little to hurt AMD. Besides we need to know performance/watt, oh and the E7330 (2,4ghz, 6mb cache) at $1391 should be a worthier competitor against the 2ghz Barcelona on price/performance.
The int numbers are really discouraging..
I thought HPC is a niche market?
How are they going to charge a price premium for HPC? Doesn't the HPC crowd use off the shelf opterons with some exotic chipsets and interconnects?
Then they can't even charge a premium for the SUPERIOR floatingpoint+bandwidth performance, because then no one will buy the opterons for normal servers (cause they are really weak on int) and if they price them according to their int performance, well...
btw
µopt was prolly right on the int performance beeing closer/worse than conroe. gj :clap:
Not only does AMD's 4P CPU has better price/performance ratio, it sits in the same mobo that customer has! For Xeon series 7300 you need new mobo. Also for Barcelona you need RDRAM and for Xeon 7300 series you need much more expensive FB-DIMM's.Quote:
Originally Posted by hawkeyefan
All in all 4P Barcelona is far more superior product.
Regarding the "niche" market... if you call auto, space, oil, mining, financial… niche (aka not worthy) markets, then OK, but targeting these market is the smartest thing that AMD could have done. Those guys have money, and they buy a lot ;)
And for the reference pricing of new Xeons, and new Opterons:
XEON
• X7350 @ 2.93GHz, 8MB L2 cache, 130W TDP at $2310
• E7340 @ 2.40GHz, 8MB L2 cache, 80W TDP at $1980
• E7330 @ 2.40GHz, 6MB L2 cache, 80W TDP at $1391
• E7320 @ 2.13GHz, 4MB L2 cache, 80W TDP at $1177
• E7310 @ 1.60GHz, 4MB L2 cache, 80W TDP at $856
OPTERON
• 8350 @ 2.0GHz, 95W TDP at $1004
• 8347 @ 1.9GHz, 95W TDP at $774
• 8346 HE @ 1.8 GHz, 68W TDP at $688
Pricing can change in a matter of days - performance on the other hand.. :rolleyes:
It's only the same if the customer has Socket F based systems. What about those who have Socket 940 systems? FB-DIMMs are essentially the same price as registered memory, while IBM's X4 based servers will use registered DDR2.Quote:
Originally Posted by Nedjo
But they haven't been buying much of AMD lately, given AMD's significant loss of market share in servers.Quote:
Regarding the "niche" market... if you call auto, space, oil, mining, financial… niche (aka not worthy) markets, then OK, but targeting these market is the smartest thing that AMD could have done. Those guys have money, and they buy a lot ;)
Performance dictates pricing.Quote:
And for the reference pricing of new Xeons, and new Opterons:
XEON
• X7350 @ 2.93GHz, 8MB L2 cache, 130W TDP at $2310
• E7340 @ 2.40GHz, 8MB L2 cache, 80W TDP at $1980
• E7330 @ 2.40GHz, 6MB L2 cache, 80W TDP at $1391
• E7320 @ 2.13GHz, 4MB L2 cache, 80W TDP at $1177
• E7310 @ 1.60GHz, 4MB L2 cache, 80W TDP at $856
OPTERON
• 8350 @ 2.0GHz, 95W TDP at $1004
• 8347 @ 1.9GHz, 95W TDP at $774
• 8346 HE @ 1.8 GHz, 68W TDP at $688
Quote:
Originally Posted by Nedjo
AMD's pricing only makes it very clear with which Intel parts they are competing ; that is 2GHz K10 is more or less equal to the E7320 in apps that customers care about.
If anything , that's worrying.You know why ? Because in that segment you're overlapping with DP , for that money you can buy a 5365 3GHz QC Xeon that would cream the K10.Few apps scale well to 16 cores , instead of 16 cores you're better off with 8 faster ones.
I'm pretty sure an 8 core 3GHz Xeon system will be cheaper and perform close enough to a 16 core 2Ghz K10 system in most commercial apps.
Quote:
Originally Posted by MR_SmartAss
Can we agree that product that cost 1004 USD wipes the floor with product that costs 1177 USD is superior?
Can we agree that product that does not require new expensive motherboard, and saves to server manufacturers big buck 'cos it uses mature and each day cheaper platform is superior?
Can we agree that product that uses far less expensive RDRAM in comparison with hot FB-DIMM is superior?
Can we agree that CPU that can independently clock its cores and logic inside of them is superior?
Can we agree that CPU that is not bottlenecked with antique FSB is superior?
;)
Since you probably backing this claim with insight to the performance of the apps that customers care about, could you be so kind and share those with us?Quote:
Originally Posted by savantu
But it doesn't, otherwise it wouldn't be priced at $1004.Quote:
Originally Posted by Nedjo
But it does require a new expensive motherboard. It's just that customers would have had to purchase it last year. But a lot didn't bother to go to Socket F and went with Intel's 2S quad-core solution instead which offered comparable performance to 4S Socket F at the time at less than half the power consumption.Quote:
Can we agree that product that does not require new expensive motherboard, and saves to server manufacturers big buck 'cos it uses mature and each day cheaper platform is superior?
They're roughly the same price. But IBM will offer registered DDR2 on its servers.Quote:
Can we agree that product that uses far less expensive RDRAM in comparison with hot FB-DIMM is superior?
No, since it's probably necessary for AMD to use every trick they can to match up with the spectacular performance/watt of the Core uarch.Quote:
Can we agree that CPU that can independently clock its cores and logic inside of them is superior?
Too bad it's bottlenecked by its lack of clockspeed, slower cache sub-system and just being slower in general.Quote:
Can we agree that CPU that is not bottlenecked with antique FSB is superior?
I'm talking TPC-C , SpecJBB , that kind of benchmarks.Quote:
Originally Posted by Nedjo
SpecFP_rate is nice , but nobody except HPC folks give a damn about it.
Oh yes it does:Quote:
Originally Posted by accord99
http://blogs.zdnet.com/Ou/images/4p-...amd-server.png
If we talk about 4P customers, we need to talk about HP, IBM, SUN, DELL, and they 4P AMD offering is SocketF, and they are very, very happy 'cos of backward compatibility.Quote:
But it does require a new expensive motherboard. It's just that customers would have had to purchase it last year. But a lot didn't bother to go to Socket F and went with Intel's 2S quad-core solution instead which offered comparable performance to 4S Socket F at the time at less than half the power consumption.
RDRAM (DDR-667): http://www.newegg.com/Product/Produc...82E16820134329Quote:
They're roughly the same price. But IBM will offer registered DDR2 on its servers.
FB-DIMM (DDR2-667): http://www.newegg.com/Product/Produc...82E16820134327
So, $50 price diference for one 4 GB kit. And since you need at least two of them that's 100 USD pricce diference. if you can call that "roughly same price" then its fine with me!
pff with all that spectacularity that you talk about, AMD still leads when it comes to perf/wat on server market: http://www.worlds-fastest.com/wfz991.htmlQuote:
No, since it's probably necessary for AMD to use every trick they can to match up with the spectacular performance/watt of the Core uarch.
if this is slower, then there is no point to this discusion:Quote:
Too bad it's bottlenecked by its lack of clockspeed, slower cache sub-system and just being slower in general.
http://blogs.zdnet.com/Ou/images/2p-...ptember-fp.png
Nedjo..for a 4way server...100$ is nothing. And your worldfastest link is heavily biased. Look on all their previous stuff.
Also i´m getting somewhat bored of these rates benches.
Rates=throughput, non rates=speed.
Is K8 faster than Core 2? I´m just curious because its faster in rates...
And backwards compability is a joke. I can also throw in harpertowns in my servers...so what? By the time there is a new CPU I would much rather get a new server. Because the I/O and such is old! And you would need new memorya swell to maximize the gain wouldn´t you? Not to talk about HT3 later on when they fix that.
Marketing FUD ftl.
Now try to post the non rates benchmarks..just to compare!
ffs it's not superior, ffs AMD would price it higher if it was superior and they COULD (they NEED higher margins ASAP), but it seems that they can't... (don't forget gaining marketshare is useless if you go bankrupt)
if this pricing won't change, it look's like their margins will take another dive... because Intel can change prices whenever they want.
btw is someone willing to answer my question? :)
"does the HPC crowd use off the shelf cpus with custom interconnects, chipsets, etc" - if yes, how is AMD going to differentiate their products. if they price them high (for the HPC people) no one is going to buy AMD for normal servers (which would be a HUGE loss), so do they have to price barcelona low (cause of its sucky int/general use performance..)?
Can someone explain what is so bad here to me. On a 2p system the AMD opteron is 4% slower in SPECint_rate.
While in the SPECfp_rate the AMD is 40%+ quicker. Surely this just shows it's a competative processor across the board.
Granted, the clockspeeds are low and until they speed up I will stick with the Xeons but it doesn't look too bad to me.
rates!Quote:
Originally Posted by Nedjo
:fact:Quote:
Originally Posted by Shintai
non-rate benchmarks should be available in a couple of days, then we'll talk performance
Not really. The rates are a very bad bench in itself. Because its mainly memory bandwidth dependent.Quote:
Originally Posted by Motiv
If rates was a sign of actual speed. K8 would beat Core 2.
sorry I'll explain a little better.Quote:
Originally Posted by Shintai
Surely the figures that are shown, does not put AMD in a bad light. I agree, the benchmark itself is of little importance but saying it makes AMD look bad (by the original Author) is no more than FUD?
Ou has updated September SPEC scores. AMD 2GHz K10 now loses by 4.2 points in int_rate.
page 20/21, I see around 7%+ for int_rate and 9.7% for fp_rate, with them supposedly just isolating fsb increase.Quote:
Originally Posted by terrace215
The problem is that you don't buy servers at a certain clock speedQuote:
On a 2p system the AMD opteron is 4% slower in SPECint_rate.
I don't understand what you mean by this.Quote:
Originally Posted by red
It means it doesn't matter if it's clock for clock equivalent if it's clocked 2GHz to 3GHz
But that would indicate I was after a 2p 3ghz system. Where I work, we have more than just speed to think of.Quote:
Originally Posted by red
If AMDs 2ghz cpu is costed equivalent to Intels 2ghz Xeon. If the rest of the system is all based the same, then both CPUs have an equal amount of chance of being bought.
Granted, if I was after pure speed then the 3ghz Xeons are the best bet but other considerations are taken into account before purchasing.
On a clock for Clock basis (and going of the price points), the benchmark does not put the AMD at a disadvantage or am I wrong?
"clock for clock", it looks fine, but it's a pity that 2Ghz quad can only command top price of $372 (it's currently $873 for 2P dual 3GHz k8).
If you look at old numbers, you'll see that AMD wasn't much behind in int_rate per core/clock anyway
http://www.spec.org/cpu2006/results/...604-01203.html
http://www.spec.org/cpu2006/results/...803-01638.html
http://www.spec.org/cpu2006/results/...821-01887.html
http://www.spec.org/cpu2006/results/...820-01831.html
http://www.spec.org/cpu2006/results/...412-00810.html
So AMD has gone from a bit behind in int_rate, to creepily close, bravo! Now if you consider that AMD scales better, on nonrate speccpu... Just comparing Ou's 2P numbers, 88/(6.4 or 7.2 which is 80-90% scaling), gets you 12.2-13.8 nonrate specint score. 92 for Intel/(5.6-6.4) which is 70-80% gets you 14.4 and 16.4.
btw, specint is a single bench suite. specintrate and sorts is running multiple instances of specint, voila
Perhaps someone else can jump in here.
"Clock for clock" it seems on par with intel. If that is the case and the prices are equal. Surely it doesn't make the CPU a bad buy?
Ofcourse not.Quote:
Originally Posted by Motiv
QFT :up:Quote:
Originally Posted by Nedjo
Using rates benches is like using DiVX encoding with a Conroe vs Penryn and claim average performance of Penryn is double of Conroe.
Imagine if we didnt have to be feed with marketing spins and lies. So many people that would be unemployed.
Not quite. Believe it or not, there are some people who like to know how a system performs when all cores are loaded. This is what the rate benchmarks test. Not everyone runs a single thread on a quad-core. ;)Quote:
Originally Posted by Shintai
Rates doesnt even reflect that either. Try do some rendering or the like. Rates is only useful for heavy bandwidth applications. And rates isn´t even multithreaded as such. Its more a bunch of singlethreaded applications running in its benching method to simulate alot of "clients".Quote:
Originally Posted by oldblue
So that somehow brings us back to 4x4 and "megatasking".
ftp://ftp.software.ibm.com/eserver/b...cpu_091007.pdf
2 socket, 1.9ghz k10s
SPECint2006 11.3
SPECint_rate2006 83.2
SPECint_rate_base2006 72.8
SPECfp2006 11.2
SPECfp_rate2006 73.0
SPECfp_rate_base2006 68.5
(1) Planned availability for the x3455 model using the AMD Opteron Model 2347 processor
(1.9GHz, 512KB L2 cache per core) is November 16, 2007.
------
ftp://ftp.software.ibm.com/eserver/b...yna_091007.pdf
The x3455 server achieved an elapsed time of 31,526 seconds, which is better than the 32,578
seconds of elapsed time achieved by a single-node 8-core blade system using the Quad-Core
Intel® Xeon® X5355 processor at 2.66GHz.
So according to the 2 PDFs,1.9Ghz DP Barcelona(2347) system is a little bit faster than 2.66 DP intel(X5355) quad system in this LS Dyna set of benchmarks.
nice.. K10 owned :up: :yepp:Quote:
Originally Posted by red
Interesting... the non-rate scores look like they will fall short of a 2.0 GHz clovertown....Quote:
Originally Posted by red
Take for example, a 5335 http://www.spec.org/cpu2006/results/...0301-00593.pdf
Barcey at 1.9 GHz SpecFP2006 -- 11.2
Clovertown at 2.0 GHz SpecFP2006 -- 13.6 (base)
It appears Barcy's key strength is BW driven.
qft, I mean come on people, does it not seem a bit funny that the only semi official leaked benches are rate benches that k8 already fears well in?Quote:
Originally Posted by Shintai
Those Specint2006 and Specfp2006 scores are very very weak, well behind what Clovertown offers on a clock for clock basis. This bodes quite badly for what K10 will offer in terms of desktop performance.Quote:
Originally Posted by red
So when we stop using bandwidth/rates benches. K10 suddenly drops behind...how funny...and how predictable.
no one in the industry who buys these CPUs gives a damn about how wellQuote:
So when we stop using bandwidth/rates benches. K10 suddenly drops behind...how funny...and how predictable.
a single core performs. System performance is what matters period.
your bashing of the _rate benchmarks and aproval of the nonrates is really
hard to understand and makes no sense. For your information the _rates
run the same nonrate benchmarks but on all the cores of the system, thus
giving us a good clue of how the whole system works on certain type of
applications. hard to pick your logic... seriously...
1 more day to go till our new rendering monster;)
Except the same cores will be used for the desktop market, and where the non-rate SPEC tests are more relevant than the rate tests.Quote:
Originally Posted by alayashu
I don't think it's a bashing of the rate benchmarks specifically, merely a desire for more benchmarks that paint a more complete picture of the performance of Barcelona, vs AMD's continual usage of only SPECfp_rate for months to hype the performance of Barcelona while being silent on the many other available server benchmarks.Quote:
your bashing of the _rate benchmarks and aproval of the nonrates is really
hard to understand and makes no sense.
Running multiple copies of single-threaded applications is not really how most server applications work.Quote:
For your information the _rates
run the same nonrate benchmarks but on all the cores of the system, thus
giving us a good clue of how the whole system works on certain type of
applications. hard to pick your logic... seriously...
This whole debate reminds of when the P4 first made its appearance, at that stage memory bandwidth was the buz word and how important it was for Quake 3. Strange how times have changed.
And you know what are the scores of Clowerton @1.9 GHz? Please share them with us!Quote:
Originally Posted by Periander6
The best published scores for a 1.86GHz/1066MHz FSB E5320 isQuote:
Originally Posted by Nedjo
SPECint2006 13.6
SPECfp2006 11.1
http://www.spec.org/cpu2006/results/...806-01697.html
http://www.spec.org/cpu2006/results/...514-01054.html
Intel Xeon processor 3040, 1.86GHz:
SPECfp®2006 = 12.5
SPECfp_base2006 = 12.3
http://www.spec.org/cpu2006/results/...625-01332.html
Dual core, though.
And when we stop using benches that fit in the cache C2D suddenly drops behind... it's also funny...and predictable:Quote:
Originally Posted by Shintai
http://blogs.zdnet.com/Ou/images/4p-...amd-server.png
Look how L7345 beats the crap out of the E7320 despite being 200mhz slower... if that isn't funnier then any result K10 can produce i don't know what is.
But i guess it's ok to use cache, but completely wrong to use bandwidth... go figure.
Since there is only 4MB DC Xeons and 8MB Quad Xeons..go figure it out. There is no 4MB quad Xeon. Its a wrong posting.Quote:
Originally Posted by DoubleZero
http://processorfinder.intel.com/det...px?sSpec=SLA69
But I would guess they use a single Quad there. And rest is dual.
again, (as i've said ad naseum), unless folks are using the production stepping vs. the B1 stepping tomorrow, you're going to see some pretty interesting results.
ask a lot of questions and try to understand what they're testing on. I've got no doubts that AMD wants the reviews to be positive, but given the huge populace of idiots in the reviewing world, someone's going to fubar it.
dave
That's a typo, E7320 is only 4mb.Quote:
Originally Posted by Shintai
http://www.hardwareschotte.de/hardwa...0+BX80565E7320
Don't you think the score is a bit high for a single Quadcore? :)
You mean run something other than SpecFP2006? Sorry... couldn't help it... this is going to be the bench people focus on.Quote:
Originally Posted by dave_graham
Clovertown 1.86 GHz best scoresQuote:
Originally Posted by Nedjo
http://www.spec.org/cpu2006/results/...0611-01262.pdf
http://www.spec.org/cpu2006/results/...0305-00613.pdf
SpecINT2006 peak 14.2, base 12.9
SpecFP2006 Peak 10.9, base 10.8
If you like, take the 2.0 clovertown results to
http://www.spec.org/cpu2006/results/...0806-01673.pdf
http://www.spec.org/cpu2006/results/...0301-00593.pdf
SpecINT2006 peak 15.5, base 14.0
SpecFP2006 Peak 14.0, base 13.6
Arh you are right. Ye could be so. Cache would have a large impact on the rates. Specially in 4 sockets.Quote:
Originally Posted by DoubleZero
Don't want to knock the results but why is the libquantum results so high? I didn't realise there was that kind of architecture improvement in clovertown?Quote:
Originally Posted by JumpingJack