NextGen VS. CurrentGen CPU:Cinebench R10

[cookerjc]

Episode 1 ：Phenom X4 with SuperPi

Episode 2：Phenom X4 with Crysis time demo

Episode 3：Phenom X4 with 3DMark06

As promised. here comes Phenom X4@Cinebench R10:

In this test, we find out that Multiprocessor Speedup of Phenom X4 is the highest one. It is because the architecture of Phenom X4 can split the workload to each core effficently. Comparing to QX9650's poor 3.56X, we think THAT'S THE POWER OF NATIVE QUAD CORE!

But sadly, each core's poor perfromance limit the total output. We can see this from Rendering(1CPU). Comparing to QX9650,the huge gap of 875 throw Phenom far away. And pls remember this is a OCed ES Phenom X4. a retail 9500 or 9600 can not OC to 3GHz that easily.

In OpenGL test we think Phenom X4 still done bad. Away from QX9650, Phenom X4 can even beat her sister FX62.Will this be another undulate like Core 2 Duo's ultra high score? judge by yourself.

[RPGWiZaRD]

Lol, that's some dissappointing results but are this one tested really the newest revision or some old ones? There's quite a difference between them for Phenom and AMD is putting a lot of work in order to try and improve it where there's possibility although hopes sure start to run out by now. That Phenom won't even match Core 2 clock for clock was expected to me since around 2 months now but that it would be this much behind is a bit unexpected. Ignoring these results I'd guess perhaps 25~30% clock for clock advantage over K8 which won't be enough for Penryn, a ~60% clock for clock performance increase over K8 would have been required for it to become a very successful launch so I knew AMD was gonna have a tough job to do.

Still I think retail samples will probably do a little better and on motherboards designed for this chip but I'm certain now Phenom won't be as good as AMD would want it to be. Matching clock for clock isn't even enough for AMD, it should be the leader in this scenario due to the clock scaling issues so yea Phenom will have to sold for a slightly less price compared to Intel's offering at same clock speed. That's not the very best scenario.

Although this test indicates it would have good quad core performance but too bad quad core 100% utilization won't be of any major importance until sometimes 2009-2010 perhaps and by then Phenom will be a forgotten tale already. Going non-native solution certainly has paid off for Intel this time around, due to being able to release a quad core a lot earlier as it's relatively easy to do the 2x dual core connected solution and since the native quad core performance scaling won't be of a significant matter for a while for the average user, it pays off both ways.

For AMD's sake I hope they've aimed a lot higher for the upcoming architecture, K8 has been milked enough already, only with a significant change in architecture can bring performance crown back to AMD. I'd like to see AMD aiming for minimum 90~100% performance boost over Phenom for the new architecture.

[informal]

Ok,i'll have to comment on this.
We have K10 with all the upgrades to the uarch. and we see it here underperforming the K8 per core basis???Wtf?

And you guys over at expreview didn't find this at least a little bit weird?

And one more question:is this "one by one" benchmarking with this "K10 system" (i put it in " " since it's more like K8 X2 @ 3Ghz) going to last 'till the 19th?You could better do a review instead doing teasing since the more you post about this Phenom,the more it seems it's slower/equal to a K8.I don't say you have a K8 and label it as a K10,i just say that you could loose a lot when the thing launches @19th and trounces K8(which it doesn't do in your tests).

[BeardyMan]

Ok,i'll have to comment on this.
We have K10 with all the upgrades to the uarch. and we see it here underperforming the K8 per core basis???Wtf?

And you guys over at expreview didn't find this at least a little bit weird?

And one more question:is this "one by one" benchmarking with this "K10 system" (i put it in " " since it's more like K8 X2 @ 3Ghz) going to last 'till the 19th?You could better do a review instead doing teasing since the more you post about this Phenom,the more it seems it's slower/equal to a K8.I don't say you have a K8 and label it as a K10,i just say that you could loose a lot when the thing launches @19th and trounces K8(which it doesn't do in your tests).

The funny part is that their K8 systems are running 375mhz on the mem to have teh exact settings as teh phenom.

is it me or wasn't 375 supposed to be 428mhz real on the Phenom?

I hardly believe any of these reviews

[Dimitriman]

Definately does not make sense that it would lose at single thread test to an X2 at the very same speed. Either this CPU is buggy or there is in fact no point in Phenom X4 at all.
Something to think about.

Nov. 19th is it? I hope it does good, I for one pre-ordered a Q9450 but that can be canceled in case Phenom proves itself better than those tests say.

[RPGWiZaRD]

Definately does not make sense that it would lose at single thread test to an X2 at the very same speed. Either this CPU is buggy or there is in fact no point in Phenom X4 at all.
Something to think about.

Nov. 19th is it? I hope it does good, I for one pre-ordered a Q9450 but that can be canceled in case Phenom proves itself better than those tests say.

Yea you're not the only one thinking there's something weird about this test setup, like early adaption issues between mobo and CPU or some mysterious bug, I would take the expreview's results on Phenom with a quite huge grain of salt.

But that doesn't change the opinion I have about Phenom not quite delivering a great competition to its rival, but I sure think it will do better than the results we've seen from this site.

[kl0012]

Ok,i'll have to comment on this.
We have K10 with all the upgrades to the uarch. and we see it here underperforming the K8 per core basis???Wtf?

What about some uarch downgrades? After all, no one anticipated that Brisbane would be slower then Windsor.

[informal]

What about some uarch downgrades? After all, no one anticipated that Brisbane would be slower then Windsor.

Not a chance.Some instr. did get higher latency going from K8 to K10,but there are other means for compensating this(and AMD done everything in order to mask it).
Brisbane G2 Vs Windsor-512 F3 is 1% from each other in tests(yes L2 is faster on G2 as well as IMC is tweaked a bit)

[kl0012]

Not a chance.Some instr. did get higher latency going from K8 to K10,but there are other means for compensating this(and AMD done everything in order to mask it).
Brisbane G2 Vs Windsor-512 F3 is 1% from each other in tests(yes L2 is faster on G2 as well as IMC is tweaked a bit)

Still the uarch difference between K8 and K10 is not that big. There are no additional pipelines & execution units in K10. All uarch enhancement (like additional prefetchers e.t.c) may not play significant role in this particular benchmark.

[informal]

Still the uarch difference between K8 and K10 is not that big. There are no additional pipelines & execution units in K10. All uarch enhancement (like additional prefetchers e.t.c) may not play significant role in this particular benchmark.

You are wrong.The instr. fetch and cache bandwidth(L1&L2) alone are doubled,the prefetch is a lot improved going from K8 to K10(split into 2 levels,one is in the very IMC),prefetch directly into L1D ,SSE resources doubled per core,much improved IMC,out-of-order load execution,large shared L3(with higher associativity,but higher latency tho) etc.
The cores stayed 3 issue,but this is not much of a big deal when there are other ways to improve per core performance than widening the core to handle 4 instr./cycle

[mstp2009]

You are wrong.The instr. fetch and cache bandwidth(L1&L2) alone are doubled,the prefetch is a lot improved going from K8 to K10(split into 2 levels,one is in the very IMC),prefetch directly into L1D ,SSE resources doubled per core,much improved IMC,out-of-order load execution,large shared L3(with higher associativity,but higher latency tho) etc.
The cores stayed 3 issue,but this is not much of a big deal when there are other ways to improve per core performance than widening the core to handle 4 instr./cycle

Sorry. But I have to disagree w/ you on this one.

AMD worked HARD on the prefetchers and the SSE instruction width of K10, but they really should have widened the overall architecture. Even with all they did, the theoretical IPC for K10 is not quite up to what C2D can do on a per Core basis. For both Integer and SSE code. The only saving grace is the FPU code throughput for K10.

[informal]

Sorry. But I have to disagree w/ you on this one.

AMD worked HARD on the prefetchers and the SSE instruction width of K10, but they really should have widened the overall architecture. Even with all they did, the theoretical IPC for K10 is not quite up to what C2D can do on a per Core basis. For both Integer and SSE code. The only saving grace is the FPU code throughput for K10.

I would have to disagree with you too .The "width" of the K10 is not a problem.let me quote D.Kanter:

Looking at a comparison of the three microarchitectures, the K8, Core 2 and Barcelona below, shows some performance hints. For many of the most important features, AMD and Intel should be matched microarchitecturally because AMD has incorporated quite a few of the techniques that Intel used to boost the per clock efficiency of the Core 2. While it does appear that the Core 2 is 33% wider than Barcelona, in reality, neither processor comes close to peak capabilities on real code, so the performance will be much closer than the block diagrams imply. Barcelona's 3-wide issue, execute and retire capabilities are not a performance problem.

http://www.realworldtech.com/page.cf...1607033728&p=9

[mstp2009]

I would have to disagree with you too .The "width" of the K10 is not a problem.let me quote D.Kanter:


http://www.realworldtech.com/page.cf...1607033728&p=9

Just like you and I, that is an opinion article. Not any statement of fact (as measured by a suite of benchmarks and such).

We can agree to disagree, but I am of the strong opinion that AMD should have added at least one more decoder to their INT units and SSE/FPU units.


What they did with K10 is like strapping a pair of turbochargers on a 2.0L 4-cylinder engine when the competition has a supercharged 6.0L V8.

Simply put, there is no replacement for displacement.

/car analogy off

[nemrod]

It's funny all the people claiming it's a fake...
I don't see so many people on kyosen thread...
http://www.xtremesystems.org/forums/...d.php?t=163372

Score of Barcelona Opteron 2350(B1) at 2.4GHz 1 cpu: 1984

As cinebench scale very well with frequency score at 3GHz: 1984*3/2.4= 2480
here 2442. (less than 2&#37
Do you really think that phenom will be far better than barcelona?


I would be probably more polite to ask some information if you think something is wrong that saying such thing:

Ok,i'll have to comment on this.
We have K10 with all the upgrades to the uarch. and we see it here underperforming the K8 per core basis???Wtf?

And you guys over at expreview didn't find this at least a little bit weird?

And one more question:is this "one by one" benchmarking with this "K10 system" (i put it in " " since it's more like K8 X2 @ 3Ghz) going to last 'till the 19th?You could better do a review instead doing teasing since the more you post about this Phenom,the more it seems it's slower/equal to a K8.I don't say you have a K8 and label it as a K10,i just say that you could loose a lot when the thing launches @19th and trounces K8(which it doesn't do in your tests).

I 'm happy that some people who have samples share informations...

[informal]

I think that AMD made a wide range of simulations before they went with 3 issue design instead of 4 issue.There is a matter of perf. gain Vs the price to implement Vs power consumption,all playing a role in their decision.
My opinion is that they found out that 4 issue was not going to bring more than enough to justify its implementation.
This is probably left for Bulldozer/next generation that comes after the die shrink of K10.Bulldozer will be against the also 4- issue nehalem uarch(while we are at it,why do you think intel stayed with 4 issue cores in Nehalem uarch.?->the same reason AMD did the same with K10)

We have another ES scoring higher than expreview's ES per clock basis:
http://forums.vr-zone.com/showthread.php?t=201874



So we have 2Ghz :1757
3 Ghz would bring the 1.5x better results(assuming linear scaling which is proven to be correct in CB10's case)
3 Ghz K10 (vr zone):2635

expreview got 2440 for their K10 @3Ghz..

[savantu]

Ok,i'll have to comment on this.
We have K10 with all the upgrades to the uarch. and we see it here underperforming the K8 per core basis???Wtf?

And you guys over at expreview didn't find this at least a little bit weird?

And one more question:is this "one by one" benchmarking with this "K10 system" (i put it in " " since it's more like K8 X2 @ 3Ghz) going to last 'till the 19th?You could better do a review instead doing teasing since the more you post about this Phenom,the more it seems it's slower/equal to a K8.I don't say you have a K8 and label it as a K10,i just say that you could loose a lot when the thing launches @19th and trounces K8(which it doesn't do in your tests).

Oh yes , conspiracies.

Why shouldn't K10 be slower than K8 on a per core basis in some tests when :

-memory per core halved
-you have a small L3 cache with high latency , cache thrashing comes to mind
-we don't know the trade offs involved in the transition from K8->K10 ( the lack of SSE through output compared to Core for example )

I think you're smart enough to realize that not everything got better or improved significantly , somewhere cuts had to be made.

You mention cache BW , improved prefetchers , OOE , improved IMC.What if the K8 design was really balanced and didn't suffer in any meaningful way from the above ?
You could very well have the sum of all the above modifications to be 20% better perf than K8 at the same clock being dimished by the slower L3 (K8 needed badly low latency ) to 10-15% and in some tests to be negative ?

[mstp2009]

I think that AMD made a wide range of simulations before they went with 3 issue design instead of 4 issue.There is a matter of perf. gain Vs the price to implement Vs power consumption,all playing a role in their decision.
My opinion is that they found out that 4 issue was not going to bring more than enough to justify it's implementation.
This is probably left for Bulldozer/next generation that comes after the die shrink of K10.Bulldozer will be against the also 4- issue nehalem uarch(while we are at it,why do you think intel stayed with 4 issue cores in Nehalem uarch.?->the same reason AMD did the same with K10)

I think, and this is a bit of devil's advocate, that K10 was designed and finalized before they saw what C2D could do. At the time of that design Netburst was horribly inefficient and AMD probably didn't think they needed to increase the core to 4+ issue to still kick Intel's arse.

The 40% IPC increase from Netburst to C2D was something that I think totally took them by surprise. And by the time sold numbers were out, K10 was in the final stages of the design being finalized for tape out (i.e. it would cost a lot of money and more importantly TIME to go back and rework the core).

AMD probably figured they could close the gap w/ K10 and remain competitive until Shanghai or Bulldozer are released.


And do we have final specifications for Nehalem? I haven't heard that the version with the IMC has taped out (but I have been absent a lot lately).

[nemrod]

We have another ES scoring higher than expreview's ES per clock basis:
http://forums.vr-zone.com/showthread.php?t=201874

But the cpu speed up is worst In that case scaling lead to multi-cpu rendering of 9566. Equal to those present here.
All these score are similar, this is ES and the only thing to say is thanks to share results, asking some info to understand but not assaulting(?) people...

[Fatal Error]

I hardly believe any of these reviews

You are so obsessed that you wouldn't believe Hector de J. Ruiz himself if he comes and say to you "Dude we lost"

[informal]

But the cpu speed up is worst In that case scaling lead to multi-cpu rendering of 9566. Equal to those present here.
All these score are similar, this is ES and the only thing to say is thanks to share results, asking some info to understand but not assaulting(?) people...

The speed up is worse since :
1)the L3 has a somewhat larger latency since the IMC domain is probably ran at lower than 1.8Ghz frequency(1.6Ghz) when the CPU speed is 2Ghz.The L3 and IMC frequency play a huge/crucial role in multi core performance of K10.
2)test was ran in 32bit mode(Kyosen's tests clearly show that K10 has somewhat better multicore scaling in x86-64 bit mode)

[tictac]

The funny part is that their K8 systems are running 375mhz on the mem to have teh exact settings as teh phenom.

is it me or wasn't 375 supposed to be 428mhz real on the Phenom?

I hardly believe any of these reviews

The test ran on 200MHz HTT

if the NB Speed : 1600MHz
Memory setting at : DDR 800
Actual memory speed : 400MHz

if the NB Speed : 1800MHz
Memory setting at : DDR 800
Actual memory speed : 360MHz

[gallag]

Um... somethings just wrong here. I am quite sure amd did enough simulations and so on... if K10 has the same IPC as K8, then what's the point of all the tech improvements, when they could have sticked 2 brisbanes on a single chip and connect them via one HT link, it would still scale much better than intel's quads, and they could have offered quad core chips months before intel did

somethings rather fishy, or all amd engineers are idiots (highly unlikely IMO)

its just the usual happening here, Amd fanboys hyping something so much that when it is released and is pretty good it is still a disappointment.

AMD doubled the cores and stayed within the same power envelope as a x2, When they originally taped this thing out it would have been competing against netburst so the engineers would have been happy with k10, then conroe came along.

FANBOYS: STOP THE HYPE

[Shintai]

Ok,i'll have to comment on this.
We have K10 with all the upgrades to the uarch. and we see it here underperforming the K8 per core basis???Wtf?

And you guys over at expreview didn't find this at least a little bit weird?

And one more question:is this "one by one" benchmarking with this "K10 system" (i put it in " " since it's more like K8 X2 @ 3Ghz) going to last 'till the 19th?You could better do a review instead doing teasing since the more you post about this Phenom,the more it seems it's slower/equal to a K8.I don't say you have a K8 and label it as a K10,i just say that you could loose a lot when the thing launches @19th and trounces K8(which it doesn't do in your tests).

Uhm..there is a few basics you seem to lack understanding.

1. 90nm K8 vs 65nm K8. There is quite a difference on those alone, due to cache latency.

2. A L3 is not always good, far from. It depends on the workload type. And with a so hard penalized L3 in terms of latency (43 cycles). The rest has to suffer too.

Welcome to the real world. Atleast you get better per core scaling....

[Epsilon84]

Geez, at this rate informal is going to have a coronary before Phenom is even launched!

Calm down dude, just pre-release launches, take em with a grain of salt.

I'd hate to see this place when we get a full blown review and the results aren't to certain people's liking...

[PhilDoc]

I'm sure it would do better with improved memory timmings, but doubtful it would catch the Intels. I know we're still only looking at a few benchmarks, but its not looking good. Sorry to say as I was hoping for AMD to pull this one off, but it may be time to start singing the song Don Meredith use to sing at the end of Monday Night Football Games.