Barcelona/Phenom single-thread SPEC performance numbers arrive

[terrace215]

It's not looking very good.

Intel has almost a 50% advantage CLOCK/CLOCK, at ~2GHz, in both int and fp.

Bodes poorly for desktop performance:

http://realworldtech.com/forums/index.cfm?action=detail&id=83478&threadid=83478&roomid=2

Name: David Kanter (dkanter@realworldtech.com) 10/7/07

Ed Brinker, an alert reader of RWT just pointed out that AMD has finally got their benchmarks together for Barcelona. They have now posted SPECint/FP results, although we are still missing SPECjbb2005 and TPC-C. Here's the deal:

SPECint2006 (base/peak)
AMD 1.9GHz - 9.97/11.3 - Note 2 socket system
Intel 3GHz - 18.9/20.8 - Note 2 socket system
Intel 2GHz - 14.2/15.6 - Note 2 socket system

There are higher SPECint2006 scores for Intel, but they use some sort of OpenMP flags which makes me think they might have been doing something 'odd' and unreflective of the benchmark.

SPECfp2006 (base/peak)
AMD 1.9GHz - 10.7/11.2 - Note 2 socket system
Intel 3GHz - 18.4/21.4 - Note 2 socket system
Intel 2GHz - 14.5/16.9 - Note 2 socket system

Looking at these scores, it's easy to understand why AMD didn't want to show single threaded performance at the Barcelona launch as the numbers are fairly disappointing. The good news is that single threaded performance is less likely to be a key purchasing concern for servers, so this isn't necessarily a substantial problem for AMD.

However, it clearly does mean that AMD will need to increase the frequency substantially to keep competitive with Intel on the desktop, where single threaded performance is much more important. Given that they still have a fair amount of time before they launch on the desktop, I certainly hope AMD can get up to 2.8GHz in a dual core configuration.

I see dual core as being much more important for desktop, as I don't think QC is very prevalent or particularly useful at the moment. I also think that if the problems are related to thermal headroom, that AMD should be able to pull off 2.8GHz or more.

DK

[Zytek_Fan]

Single thread isn't too important anymore, so I wouldn't be that worried, although the performance is worrying.

[JumpingJack]

Single thread isn't too important anymore, so I wouldn't be that worried, although the performance is worrying.

You may want to rethink this.... the data is pretty clear that in CPU bound workloads, where bottlenecks in the bussing are no issue, AMD is coming up short clock for clock. In light of this data, it is unlikely AMD will be ahead on desktop in most anything... even the multithreaded apps for desktop have working sets that are not stressing the bus, not even a 1067 MHz FSB.

This is both good and bad news. The good news is that AMD will need to continue competing on price and this will afford AMD users, and those with AM2 boards, a cheaper CPU that performs better than a K8, and it will bump up AMDs ASPs a little. Tricore now becomes more important for AMD overall.

The bad news is that Intel will not really need to clock up the Penryn lineup to keep a performance lead, hence they will probably keep lower clocked top bins and waterfall the prices from there to bouy the lower end and keep ASPs up, so Intel users will likely be paying more for the Intel CPUs than they otherwise would need to since AMD really isn't going to challenge the top end.

Jack

[awdrifter]

Hopefully this will mean there's some cheap quad-cores soon. If K10 can only match C2Q clock for clock, then we might get a 2ghz quad fore less than $200.

[Motiv]

Frankly, the results are awful.

[DoubleZero]

It's not looking very good.

Intel has almost a 50% advantage CLOCK/CLOCK, at ~2GHz, in both int and fp.
First they nowhere in spec to be found, those are the same numbers that was in the IBM slide... so nothing new here.
Second it's a 1.9Ghz not a 2ghz part and it's nowhere near 50%.

And it's not really fair and smart comparing the best result out of dozens with the only one that it's known and that it isn't even at Spec yet.

Just to prove my point, first Xeon E5335 on spec:
SPECint2006 12.3 12.8
SPECfp2006 10.9 11.2

A lot different from what you have now, so sorry to burst your bubble but those barcelona numbers will go up as systems are submitted.

And yes Intel will be faster at desktop clok-for-clock, maybe AMD can sneak up at 3.0ghz and faster but that's it, and again nothing new.

[Shintai]

Actually I think people dont understand the rates vs non rates. Its not really multithread vs singlethread. But high troughput vs low troughput.

So if you do something like Mental ray rendered both Clovertowns and Barcelonas scales 100%. Its all about how much memory bandwidth they need to substain the peak.

[Hornet331]

First they nowhere in spec to be found, those are the same numbers that was in the IBM slide... .


Name: David Kanter (dkanter@realworldtech.com) 10/8/07

Mike (overclockerfan@nospamATgmail.com) on 10/8/07 wrote:
---------------------------
>Do you have the direct links to the SPEC.ORG pubs so we can >see the configs used?
SPECint2006:
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01948.html

SPECfp2006:
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01949.html

Strangely enough, I just noticed that for peak, they use Pathscale's compilers on FP, but for base they use PGI. On SPECint, they use PGI for most everything except 3-4 benchmarks on peak. Very strange indeed.

DK

Enjoy. :)

[Shintai]

For desktop, the single and dual threaded performance is most important.
Anyway this is not giving any image about K10 performance on a single socket.

Notice the poor scaling of the dual Intel 3GHz vs dual Intel 2GHz. With a 50% higher clock, the Intel platform achieves 33% and 27% for INT/FPU respectively. Obviously the whole platform is bottlenecked by the slow parallel buses and the indirect CPU to CPU communication. The scaling of the performance on the single CPU configuration using FSB1333 is almost linear for these frequencies(3GHz and below).

Not really, the main issue is that the main memory do not scale with it. Barcelona would scale a little bit better but still suffer from the same.
And remember the dual sockets Xeons are dual 1333FSB, not single.

[knightwolf654]

i would really change the title, those are barcelona scores not phenom, barc has several things turned of so it would be a direct drop into older motherboards.

[Shintai]

i would really change the title, those are barcelona scores not phenom, barc has several things turned of so it would be a direct drop into older motherboards.

Eh? Phenom is the same thing. It can be dropped into older motherboards too.

Only difference is Phenom can use faster non ECC memory. Just like desktop Core 2s can plus no FB-DIMM.

[knightwolf654]

ok then its still not using a 1207+ mobo their for HT3 is turned off.

so these are not the final numbers.

[kl0012]

ok then its still not using a 1207+ mobo their for HT3 is turned off.

so these are not the final numbers.

Doubt if HT1 instead of HT3 can affect single threaded performance.

[DoubleZero]

SPECint2006:
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01948.html

SPECfp2006:
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01949.html


I searched for 2349 :shakes:
But again.. it's the same numbers and same system that was in the IBM slide.

[Syn.]

There are higher SPECint2006 scores for Intel, but they use some sort of OpenMP flags which makes me think they might have been doing something 'odd' and unreflective of the benchmark.


stopped careing when i read that bit.

[BrowncoatGR]

Well with the current bios problems that seem to exist i think people should take this with a pinch of salt (or two...);)

[BrowncoatGR]

Here's some quick K8 results @2GHz specint base=9.77 rate=10.8
specfp base=10.4 rate=10.9
This is way too close. Either there is something wrong with the K10 IBM system or K10 is going to be a huge disappointment for everybody except the HPC crowd. I would wait for at least a couple more results to get published. Just look at the problems stephen has been having with the various Bioses.
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070828-01902.html
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070828-01900.html

[Shintai]

Well with the current bios problems that seem to exist i think people should take this with a pinch of salt (or two...);)

How long are we gonna hear on that joke? What in a year when the performance is the exact same. Then whats the next excuse?

Its a retail final and shipping product and been for some time now.

[Shintai]

Here's some quick K8 results @2GHz specint base=9.77 rate=10.8
specfp base=10.4 rate=10.9
This is way too close. Either there is something wrong with the K10 IBM system or K10 is going to be a huge disappointment for everybody except the HPC crowd. I would wait for at least a couple more results to get published. Just look at the problems stephen has been having with the various Bioses.
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070828-01902.html
http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070828-01900.html

Or maybe K10 was hyped to something its not. K10 is a tweaked K8. Then mixed in a quad design.

[[XC] hipno650]

it's really not looking that good for AMD. so far what we have seen is

-clock perclock AMD loses
-clock speed AMD loses

and the worst part that AMD has better multi cpu platforms than Intel so on a single socket we may see Intel take more of a lead. i just hope they can come up with something or the only hope for them will be dirt cheap prices (which means more quads for my farm:D ) but also that they are getting in a financial bind.

[redpriest]

oh em gee. The Intel scores are using autopar, so they're not directly comparable.

[GoThr3k]

kinde strange AMD loses in FP????
even K8 beats conroe in that part :s

[redpriest]

To wit, let me include the intel result here and explain to you why their specFp score seems off the chart.

http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01963.gif

http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01963.html

cactusADM look funny anyone?

Here by comparison is a "sane" cactusADM score without autopar:

http://www.spec.org/cpu2006/results/res2007q1/cpu2006-20070213-00433.gif

which comes from IBM's intel submission here:
http://www.spec.org/cpu2006/results/res2007q1/cpu2006-20070213-00433.html

[Kunaak]

all I care about is how well they OC, and how they do in games.

till I see them in action in games, the rest doesnt mean a whole lot to me.

[PhilDoc]

all I care about is how well they OC, and how they do in games.

till I see them in action in games, the rest doesnt mean a whole lot to me.


With you all the way. The only thing I'd add was the cold bug.

[Periander6]

To wit, let me include the intel result here and explain to you why their specFp score seems off the chart.


Here are some Intel Specfp results without Auto parallelization.

1.86GHz: 10.9/11.1

http://www.spec.org/cpu2006/results/res2007q2/cpu2006-20070514-01058.html

2.0GHz 11.9/12.1

http://www.spec.org/cpu2006/results/res2007q2/cpu2006-20070427-00929.html

2.66GHz 14.3/14.5

http://www.spec.org/cpu2006/results/res2007q2/cpu2006-20070514-01072.html

3.0 GHz 16.5/16.9

http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070820-01861.html

Compare to AMD

1.9GHz 10.7/11.2

http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01949.html

Close, but even here Intel seems to have a slight edge clock for clock. For Specint it's a huge clock for clock margin for Intel.

[Zytek_Fan]

AMD fanpois still have hopes:
- the newer BIOS will improve the performance dramatically....
- the low latency DDR2-1066 will improve the performance even further...
- at higher frequency K10 scales better....

K10 does scale really well.

[Hornet331]

K10 does scale really well.

with more cores, yes... with frequency, needs to be determined.

[Shintai]

kinde strange AMD loses in FP????
even K8 beats conroe in that part :s

I think you mix specFP and specFP_rates.

[KTE]

Here's some quick K8 results @2GHz specint base=9.77 rate=10.8 specfp base=10.4 rate=10.9

SPECint2006 (base/peak) -- SPECfp2006 (base/peak)
AMD 1.9GHz - 9.97/11.3 -- AMD 1.9GHz - 10.7/11.2

So basically, 8 cores from 4 cores, keeping the frequency the same didn't show any noticeable improvement? Even if Barcelona is that bad of a performer (relative to K8), new arch and upgrade, what many would like us to believe, that is below ridiculous performance change being shown, and simply lunatic if compared with AMD press releases. I must admit, a month after Barcelona release, I'm fed up with BIOS problems and so on. Meanwhile AMD denies it all emphatically saying platform support is superb. Same as what happened with R600 from them. :shakes:

Here are some Intel Specfp results without Auto parallelization.

1.86GHz: 10.9/11.1
http://www.spec.org/cpu2006/results/res2007q2/cpu2006-20070514-01058.html

The one socket scores are little relevance here. This is the E5320 2S 1.86GHz score you should quote to show minimal difference going from one to two, even with multithreaded base optimization (Intel):

Published Oct 2007

SPECint2006 (http://www.spec.org/cpu2006/results/res2007q4/cpu2006-20070917-02080.html) (base/peak) -- SPECfp2006 (http://www.spec.org/cpu2006/results/res2007q4/cpu2006-20070918-02100.html) (base/peak)
Xeon 1.86GHz 12.7/14.0 -- Xeon 1.86GHz 12.0/12.4

Published Sept 2007

SPECint2006 (http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01948.html) (base/peak) -- SPECfp2006 (http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01949.html) (base/peak)
Opteron 1.9GHz 9.97/11.3 -- Opteron 1.9GHz 10.7/11.2

[BeardyMan]

AMD fanpois still have hopes:
- the newer BIOS will improve the performance dramatically....
- the low latency DDR2-1066 will improve the performance even further...
- at higher frequency K10 scales better....
well i don't care about server performance really..all the gibberish it doesn't make sense to me. Like kunaak stated game and oc possibilities.

We haven't seen any desktop performance yet with real desktop parts.

Shall i define myself mister Pois?:D

-Motherboard equiped with RD790 or any Nvidia chipset, from a brand like ASUS ord MSI or DFI
-DDR3 from a brand like OCZ,Corsair or Centon
-Retail Phenom cpu

Stuff we can actullay buy in the store from brands that has proven themself :)
So there is no need for hoping yet :)

[Spawne32]

all I care about is how well they OC, and how they do in games.

till I see them in action in games, the rest doesnt mean a whole lot to me.

:up: cant argue with that, got some serious denial in this thread goin on.

[PhilDoc]

SPECint2006 (base/peak) -- SPECfp2006 (base/peak)
AMD 1.9GHz - 9.97/11.3 -- AMD 1.9GHz - 10.7/11.2

So basically, 8 cores from 4 cores, keeping the frequency the same didn't show any noticeable improvement? Even if Barcelona is that bad of a performer (relative to K8), new arch and upgrade, what many would like us to believe, that is below ridiculous performance change being shown, and simply lunatic if compared with AMD press releases. I must admit, a month after Barcelona release, I'm fed up with BIOS problems and so on. Meanwhile AMD denies it all emphatically saying platform support is superb. Same as what happened with R600 from them. :shakes:



The one socket scores are little relevance here. This is the E5320 2S 1.86GHz score you should quote to show minimal difference going from one to two, even with multithreaded base optimization (Intel):

Published Oct 2007

SPECint2006 (http://www.spec.org/cpu2006/results/res2007q4/cpu2006-20070917-02080.html) (base/peak) -- SPECfp2006 (http://www.spec.org/cpu2006/results/res2007q4/cpu2006-20070918-02100.html) (base/peak)
Xeon 1.86GHz 12.7/14.0 -- Xeon 1.86GHz 12.0/12.4

Published Sept 2007

SPECint2006 (http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01948.html) (base/peak) -- SPECfp2006 (http://www.spec.org/cpu2006/results/res2007q3/cpu2006-20070903-01949.html) (base/peak)
Opteron 1.9GHz 9.97/11.3 -- Opteron 1.9GHz 10.7/11.2


I guess I'm a little lost, if this is a single thread test then it shouldn't matter how many cores you have. Ether way, all I want to see is how it does in games, Video and Photo editing. How well it overclocks and if it has a cold bug. Other than that, like Kunaak, I could care less. I also suspect that as time goes on single thread test, even on the desktop will be meaningless

[BrowncoatGR]

Well NUMA can actually hurt single thread performance