SiSoft Leaks AMD FX "Zambezi" Scores: Worse Than Intel Core i7

[Ancalagon44]

There's no point into claiming fakeness, if the benchmarks were valid. What for? So to make people wait another month and the go to Intel anyway? It's one thing getting the better product and a quite different thing being also angry to the rival company for making you wait. JF-AMD wouldn't want that, so he was telling the truth. The guys above (brazos example) already showed how they can't be true anyway.

As for the rv770, I meant its surprising performance. Yes it was not delayed.

Valid benchmarks proving BD is faster than SB - AMD's sales pipeline stalls and no one will buy current Athlons (*)
Valid benchmarks proving BD is slower than SB/P4/whatever else - people lose confidence in BD and wont buy it.

In either case, it causes a problem as far as BD is concerned. Hence why JF AMD would rather you believe that any benchmarks are either completely fake or not representative.

* I dont agree with AMD's logic on this because I dont think it will cause a sales pipeline stall, and I think the result would be a net gain. But that is their thinking behind not releasing any benchmarks ahead of time, look it up in other threads.

 

[sm625]

The only reason to buy anything AM3+ is bulldozer. If BD is junk then it makes llano look like the best (non-intel) option. If BD is really as slow as brazos, then the 8 core should be selling for $100. And they should be able to make money even at that price. It is impossible for a company like intel to run away with 60% profit margins. AMD can beat them on price/performance long before the profit margin is gone. But of course it is laughable to think they would bother releasing a new cpu if it is significantly slower than their previous design.

 

[BD231]

He said this he said that, sounds like HE needs to STFU.

 

[Black96ws6]

Say what you want about them possibly being fake, but that Google cache page hit is legitimate. Here is the URL so everyone can see for themselves:

http://webcache.googleusercontent.c...zer&cd=2&hl=de&ct=clnk&gl=de&client=firefox-a

Final Thoughts / Conclusions


Looking back at all the benchmarks of the new CPU, it is obvious that AMD is a generation behind Intel performance-wise. The "all new" design is better than the K8-variants and some results (e.g. Cryptography, Multi-Media) than the competition, but the other results are dissapointing. 256-bit performance (AVX/FMA) - what modern software will use - is "slow", most likely due to the shared FPU within a CU; the competition saw great gains from AVX, this is not the case here.
It will be challenging for AMD to compete in the "enthusiast market" with a CPU with (only) dual channel memory against tri-channel; maybe that's why Intel has not bothered to upgrade the 3-year old X58, but when it does (X79) the gap will widen even more. For AMD fans the upgrade is definitely worth it, especially if they already use a socket AM3+ motherboard; for those that have a socket AM3 boards you'll need to check whether a BIOS update is enough to ensure compatibility. In any case, the CPU is much faster than the mainstream "Llano" (see AMD Desktop "Llano" APU (CPU+GPU)).

 

[RussianSensation]

Black96ws6,

Right now, SB is 2 generations ahead of Phenom II (since Phenom II is a Core 2 Quad 45nm competitor). So if BD moves within 1 generation (around Core i7 Nehalem in performance), that seems pretty reasonable. To get to 2600k in performance would imply closing a 2 generation gap, which is more on the optimistic side (and more and more indications are pointing that this isn't happening). Obviously, we won't know for sure until the final release date. But even if BD = SB, SB has been out for 9 months now. So in the overall context of execution, AMD's management leaves much to be desired.

From an enthusiast perspective, an overclocked Bulldozer FX-8150 has to beat an overclocked Core i7 920. This is my personal benchmark level for Bulldozer because i7 920 has been available for $284 since November 16, 2008. If BD in its overclocked state cannot beat an overclocked 920 CPU, then I will view it as a failure.

 

[chihlidog]

Black96ws6,

Right now, SB is 2 generations ahead of Phenom II (since Phenom II is a Core 2 Quad 45nm competitor). So if BD moves within 1 generation (around Core i7 Nehalem in performance), that seems pretty reasonable. To get to 2600k in performance would imply closing a 2 generation gap, which is more on the optimistic side (and more and more indications are pointing that this isn't happening). Obviously, we won't know for sure until the final release date. But even if BD = SB, SB has been out for 9 months now. So in the overall context of execution, AMD's management leaves much to be desired.

From an enthusiast perspective, an overclocked Bulldozer FX-8150 has to beat an overclocked Core i7 920. This is my personal benchmark level for Bulldozer because i7 920 has been available for $284 since November 16, 2008. If BD in its overclocked state cannot beat an overclocked 920 CPU, then I will view it as a failure.

I dont expect BD to match SB. I'm hoping it matches first gen i5/i7 though. My i5-750 is smoking fast, and if I had the equivalent of EIGHT of my current cores, that would indeed be a worthy upgrade in my mind. Sure, not much uses them right now, but before long more and more applications will be making use of more cores. So I think in a sense I pretty much agree with you. If it matches intel's last gen CPUs, and has more cores available, it will be pretty attractive.

 

[BlueBlazer]

QC Bulldozer actually has 4 cores,as opposed to dual core Brazos. Who cares if the front end is shared?You still have 4 actual integer cores in the device. And no,there is not a way each integer Bulldozer core with all the improvements over F10h can end up being slower than each Brazos core.

Let me re-iterate again, these Bulldozer "cores" are not full blown cores like the ones in Brazos. In the original CMT patent, these Bulldozer "cores" are refered to as "clusters" (not cores!). The shared resources can be (IMHO) "choke points". Example, there's only one FP scheduler to serve both cores. Then there's the single front end Fetch and Decode as well. If each section have its own clock speed domain.... Then if one "core" is clocked slower then what clock speed would the front end run, and would it need to wait for the slower "core" to catch up (in async or sync manner)? Will this stall or drag down the other (higher clocked) "core"? Its a pretty complex there and lots need to be properly figured out when designing (which is why it can take years).

Not only you have much beefier units,you have L2 cache that's on another level compared to L2 of Bobcat(4x the size and double the frequency),you have prefetching and branch prediction that is on another level,unified scheduler for 4 execution units etc.

The L2 cache on Bulldozer is also shared between both "cores". There will at least some performance penalty there as well. You can see this example from Core 2 Duo which has L2 cache shared between two real cores.

For this thing to end up slower than Bobcat at the same clock,the very uarchitecure has to be fundamentally broken in some way.Which won't happen. I won't even mention FP resources since it's not even funny how big the difference is.

The fallacy of that line of thinking (that it should not happen) is similar to Intel's first Pentium 4 (Willamette) should have been overall faster (in almost every application) than Pentium III (which unfortunately is not the case). Need to wait and see how this Bulldozer CPU "matures" (new steppings/revisions and enhancements). And yes, possible delays as already witnessed. :hmm:

 

[inf64]

I won't address particular points,but you are wrong. Each bulldozer core is a real core in every sense of the word. It is as real as Bobcat core is. Sharing is design choice. All those possible "issues" you are listing were resolved long before AMD even scratched BD version 0 (45nm) and went for new one on 32nm. Sharing of L2 is beneficial(believe it or not ). Lastly,Bobcat,on paper is in every sense a half of the half of bulldozer module(or simply has roughly half the exec. resources per core versus Bulldozer core- not module).

 

[AtenRa]

Let me re-iterate again, these Bulldozer "cores" are not full blown cores like the ones in Brazos. In the original CMT patent, these Bulldozer "cores" are refered to as "clusters" (not cores!)

Since Bulldozer is AMD’s design they can call it whatever they want.
In the beginning, FP execution units where outside of the Core as you would like to call it, Core at that time was the Integer Execution Unit and FP was called a mathematical co-processor.

So one module has two cores, that’s how AMD call them and that’s how they actually work, so no point arguing that.

The shared resources can be (IMHO) "choke points". Example, there's only one FP scheduler to serve both cores.

Intel’s Core has one Scheduler for both Integer and FP, you better research more

 

[BlueBlazer]

I won't address particular points,but you are wrong. Each bulldozer core is a real core in every sense of the word. It is as real as Bobcat core is.

I still have to disagree with that. Also, AMD calling them "cores" IMHO a marketing strategy.

Sharing is design choice. All those possible "issues" you are listing were resolved long before AMD even scratched BD version 0 (45nm) and went for new one on 32nm.

How do you know they are fully resolved? CMT existed before "Turbo" feature was introduced. Then there's also the power management (each section have their own clock speed domain, as well as power gating). Have you thought about the effects of introducing these features into the design? Some very early engineering sample (B0 stepping) benchmarks leaked shows lower performance (in single threaded SuperPi) at higher overclocked speeds, which IMHO could be either be power management (C states), clock domain issues (speculations such as FP scheduler clock speed, slower "core" held back the active faster "core", etc), or simply "bugs". :hmm:

Sharing of L2 is beneficial(believe it or not ).

Like I've said earlier, sharing L2 cache does reduce performance a bit, example >> Cinebench benchmarks (compare single CPU with multi-CPU). You can easily see where AMD's original 90% per "core" performance derived from.

Lastly,Bobcat,on paper is in every sense a half of the half of bulldozer module(or simply has roughly half the exec. resources per core versus Bulldozer core- not module).

That is why I've mentioned earlier that if you treat these Bulldozer "quad core" as dual core, then those disrepencies can be eradicated (Bulldozer should be faster than Bobcat "module"-to-core).

 

[sm625]

I dont get the PIII to P4 comparisons. P4 was built for speed, and back in those days they had no reason to think the design wouldnt scale all the way to 6, 7 8 GHz. AMD knows damn well they aint gonna hit 6GHz, and they knew that when BD first got on the drawing board, so why design for speed?

 

[inf64]

That is why I've mentioned earlier that if you treat these Bulldozer "quad core" as dual core, then those disrepencies can be eradicated (Bulldozer should be faster than Bobcat "module"-to-core).

I don't know if you udnerstood me correctly: each Bobcat x86 core has roughly half the execution resources of each Bulldozer core(or what you call cluster). So a dual core bobcat has roughly 2x less execution resources than a BD module. So there is no sense that one BD core(or what you would refer to as cluster) can perform worse than one Bobcat core.

edit: as for Cinebench speedups,have you ever heard of multicore scaling? Nothing scales perfectly,so single to dual core (that has no SMT-like technique onboard)will never see perfect scaling . What you get usually is 1.85-1.95 range.

 

[cantholdanymore]

Let me re-iterate again, these Bulldozer "cores" are not full blown cores like the ones in Brazos.

Not again; dude where were you about a thousand post back?

 

[BlueBlazer]

I don't know if you udnerstood me correctly: each Bobcat x86 core has roughly half the execution resources of each Bulldozer core(or what you call cluster). So a dual core bobcat has roughly 2x less execution resources than a BD module. So there is no sense that one BD core(or what you would refer to as cluster) can perform worse than one Bobcat core.

Just did some calculations (adjusting for clock speed differences) and it does seem each Bulldozer module is slower (than a Bobcat core)! Not sure where's the issue here. Either its due to its pipeline design for high frequency (ala Netburst), or some hardware bugs.

edit: as for Cinebench speedups,have you ever heard of multicore scaling? Nothing scales perfectly,so single to dual core (that has no SMT-like technique onboard)will never see perfect scaling . What you get usually is 1.85-1.95 range.

Perfect scaling would be 2.00 for two cores (theoretically), and in the real world it would be less than that (due to hardware factors such as cache coherency and latency, memory read/write speed, bus, etc). But that's not what we see in the Cinebench results, which looks more like 1.83 (or if adjust for core speedup, around 91.7&#37. IMHO the FSB and memory has little to do here, rather its the cache (and core-to-core intercommunications).

Not again; dude where were you about a thousand post back?

You can refer to this post. I believe chew* has handled Bulldozer before, thus he knows how it performs like.

Edit: Latest chew* comments.....

Wow someone with common sense...........

Like i said before I discussed this face to face with engineers at AMD, BD is a native 4 core 8 thread part.

and then the lynch mob got mad..........

 

[gmaster456]

I don't buy it one bit.

 

[munchkincakes]

There's no real use in looking at benchmarks of things before they come out. As of right now though, Bulldozer is not looking very good. There's a revision on the way in Q2 2012, the current version is repeatedly delayed month after month, and apparently things aren't looking too great on the server side. I'm hoping that the improvements can be made soon so I can upgrade my 2nd machine, but I have serious doubts.

(I run two machines in my house, one using Intel while the other uses AMD, one using Nvidia while the other also goes AMD/ATI. It keeps me up to date on things.)

 

[Gigantopithecus]

Black96ws6,

Right now, SB is 2 generations ahead of Phenom II (since Phenom II is a Core 2 Quad 45nm competitor). So if BD moves within 1 generation (around Core i7 Nehalem in performance), that seems pretty reasonable. To get to 2600k in performance would imply closing a 2 generation gap, which is more on the optimistic side (and more and more indications are pointing that this isn't happening). Obviously, we won't know for sure until the final release date. But even if BD = SB, SB has been out for 9 months now. So in the overall context of execution, AMD's management leaves much to be desired.

From an enthusiast perspective, an overclocked Bulldozer FX-8150 has to beat an overclocked Core i7 920. This is my personal benchmark level for Bulldozer because i7 920 has been available for $284 since November 16, 2008. If BD in its overclocked state cannot beat an overclocked 920 CPU, then I will view it as a failure.

I think this is a very reasonable assessment of the situation, and I agree 100% with your perspective as a fellow enthusiast.

I also think it's astonishing that a company with a market cap of less than $5B can stay within one or two generations of a company with a market cap of over $100B.

 

[Red Storm]

I also think it's astonishing that a company with a market cap of less than $5B can stay within one or two generations of a company with a market cap of over $100B.

Agreed. I don't think people realize just how much of an advantage Intel has always had.

 

[Dresdenboy]

Say what you want about them possibly being fake, but that Google cache page hit is legitimate. Here is the URL so everyone can see for themselves:

Nobody said that this Sisoft article is fake. It just looks like the results are flawed. Or why has it been removed that quickly?

Sent from my GT-I9000 using Tapatalk

 

[exar333]

The only reason to buy anything AM3+ is bulldozer. If BD is junk then it makes llano look like the best (non-intel) option. If BD is really as slow as brazos, then the 8 core should be selling for $100. And they should be able to make money even at that price. It is impossible for a company like intel to run away with 60% profit margins. AMD can beat them on price/performance long before the profit margin is gone. But of course it is laughable to think they would bother releasing a new cpu if it is significantly slower than their previous design.

This got me thinking. AMD is hurting themselves, in my opinion, by not releasing benchmarks for BD. Because of the uncertainty of BD performance, people do not know what AMD path is better (AM3+(BD) or Llano). No one wants to choose the wrong path, right? So they buy Intel.

Be it good, or bad news about BD, let people decide. If good, AM3+ is a great buy even with a delay, because they can drop it in later. If bad, grab a Llano and wait for revisions there.

No news just begs you to go choose a 2500k and call it a day.

Just my $0.02...

 

[Dresdenboy]

Agreed. I don't think people realize just how much of an advantage Intel has always had.

This is a result of only looking at SuperPi

Sent from my GT-I9000 using Tapatalk

 

[soccerballtux]

benchmarks ho hum. SiSoftware has always favored? Intel. Wonder if Intel gave them a free Intel compiler.

 

[blackened23]

Anand posted on twitter to not regard these benchmarks floating around with any kind of accuracy.

My guess is that he's in the loop on these things, so if he's defending BD (even though he hasn't released numbers) that seems like a good sign. I'm neither an intel or amd fanboy, but I hope BD turns out good.

As far as the benchmarks in this thread. Ridiculous. There's no way BD would be that awful, sounds like whoever posted those benhcmarks fabricated them or used an early engineering sample.

 

[RussianSensation]

I think this is a very reasonable assessment of the situation, and I agree 100% with your perspective as a fellow enthusiast.

I also think it's astonishing that a company with a market cap of less than $5B can stay within one or two generations of a company with a market cap of over $100B.

I also find it amazing that after Athlon XP+, Athlon 64, Athlon X2/FX, AMD twirled thumbs and barely improved IPC over A64 in Phenom I/II. Basically, they underestimated Intel by pretty much rolling over Athlon 64's architecture into Phenom II. When AMD took risks and was focused on innovation, it was not far behind Intel (and sometimes ahead). The intellectual capital (of engineers) was there.

I can only blame poor management of Ruiz and an insane premium that AMD paid for ATI in the $5.4B acquisition......Remember that they purchased ATI using a leveraged buyout, and the interest on the debt killed any positive cash flow AMD had to reinvest into R&D. So essentially, the buyout of ATI took the wind out of AMD's sails, and 5 years later it seems it still has not recovered. The fact that GPU profits/revenues started to shrink in recent years as users shifted to gaming on consoles/tablets/smartphones isn't helping AMD's graphics division either.

 

[BlueBlazer]

Anand posted on twitter to not regard these benchmarks floating around with any kind of accuracy.

My guess is that he's in the loop on these things, so if he's defending BD (even though he hasn't released numbers) that seems like a good sign. I'm neither an intel or amd fanboy, but I hope BD turns out good.

If you had followed posts in the other threads like mine here and inf64's, then you would know what's Anand is actually saying. And for the (latest) current B2F stepping, the performance is not up to par.

As far as the benchmarks in this thread. Ridiculous. There's no way BD would be that awful, sounds like whoever posted those benhcmarks fabricated them or used an early engineering sample.

These are not fabricated benchmarks, and also some of those posted here recently are latest Engineering Sample B2 steppings (possibly "B2F" stepping). The final shipping one is going to be a newer B2G stepping, though we may not know how much improved its performance will be (you can try to speculate). The message from Anand is that the current leaks "may not be representative" of the final products. :hmm: