AMD Zambezi news, info, fans !

[FlanK3r]

If will be FX better core to core (or clock to clock), must be FX61xx better performance than x6 1100t. Example efectivity of performance can be 6C (or 3c/6tsm) FX simillary as efectivity of performance 4c/8T (intel hyperthreading)

[Solus Corvus]

Right now I think it would be best to moderate your expectations and be patient. You guys are getting really twisted up trying to justify your expectations. I don't think that level of emotional pressure on an unknown future is healthy. If you are right it fills you with a sense of self–righteousness that reality legitimized your expectations. And if you are wrong it can be depressing or even turn you into the dreaded anti-fan. There are examples of the latter here in our very own forums. Instead practice patience and try to see reality as it is, free from our preconceptions.

I want bulldozer to succeed as much as any of you guys. Partly because of nostalgia and partly because it would be the better outcome for the computer market. Those with a long memory will remember that I speculated favorably about bulldozer since the beginning, against some of the worst Intel trolls. But my speculations, hopes, and expectations can't change the reality of making a chip. The reality is that when the theory of making a chip turns in to the actual implementation of physically creating that chip things can go wrong at many stages of that process.

Quite a while ago (just after we learned BD has 2 ALU) I made a rough estimate. I figured that if early 8 core BD could attain 4.25GHz base clock then AMD and Intel would be near parity for regular workloads and AMD would have a serious advantage in high thread count workloads. Details about the pipeline, cache hierarchy, memory subsystem seemed to bear this theory out - AMD is aiming for high clocks to make up for any IPC deficit compared to SB/IB.

But then rumors of the release clocks started rolling in. 3.6GHz, WTF? Of course my estimates could have been way off, but then why the long pipeline, etc for clocks not any faster than current products? Something doesn't smell right. Then there was the rumor of FX-8170 at 4.2, then 3.9. And the rumor about the 4.2GHz 4 core part. The reality is that to make a high frequency product everything has to go right, the architecture and the process. What do you do as a company if your foundry (for example) can't meet your original design clocks? You release what you can and try to fix the process or tread water until a new better one can be implemented. No amount of me or you wishing that it can be fixed with a respin will actually fix it if the problem lies elsewhere (if there is a problem at all). So here I am closer to release and I am letting go of my emotional attachment to all the speculations and hopes - speculation is for fun and I'm not going to let it wreck my emotional state if reality doesn't want to cooperate.

As for all this arguing over what defines a core, it's garbage. You guys are arguing over a fuzzy term that is more useful to the marketing team then the engineering team. What matters is how the processor as a whole performs with various workloads, not the number of cores. Cores is just a term to plaster in big letters on the box to sell to people who are too clueless to do actual research before impulse buying it. Just like GHz was in the P4 days.

[drfedja]

SiSoft Sandra 2011 does have AVX/FMA support, but we don't know anything about implementation of FMA.

AVX/FMA instruction set support for new CPUs
Processor Multi-Media, Processor Cryptography, Memory Bandwidth, Cache and Memory Benchmarks

http://www.sisoftware.net/?d=news&f=...elease&l=en&a=

[informal]

Well I basically agree with you Solus Corvus. I'm now quite pessimistic when it comes to Zambezi and desktop market. Interlagos can still do pretty good in server market,it has relatively good clock speed and number of threads per chip.

[xsecret]

As for all this arguing over what defines a core, it's garbage. You guys are arguing over a fuzzy term that is more useful to the marketing team then the engineering team. What matters is how the processor as a whole performs with various workloads, not the number of cores. Cores is just a term to plaster in big letters on the box to sell to people who are too clueless to do actual research before impulse buying it. Just like GHz was in the P4 days.

I disagree with this statement. No matter how much raw power a CPU can deliver, the power is nothing if software can't use it in real world application (and without too much work for devs). Threading is a mess and an high time-consuming task for developers, so single thread performance and core count have a major impact in real world performance.

This said I wonder what u guys will think if benchmarks published after the NDA are similar to actual ones ? Reviewer sucks ? AMD issue with a TLB or anything else like for first phenom ? crappy µarch ?

[Solus Corvus]

Well I basically agree with you Solus Corvus. I'm now quite pessimistic when it comes to Zambezi and desktop market. Interlagos can still do pretty good in server market,it has relatively good clock speed and number of threads per chip.

I don't think being pessimistic is good either. It's reasonable to have concerns, but there is plenty to look forward to as well. I think it's best to replace all those positive and negative emotions we are placing on this product with simple curiosity. Even if it isn't a fire breathing monster it is still a new architecture to play with. We have no idea what overclocking will be like, how it will perform with our specific applications, etc. Reality is more interesting than all of our emotional baggage combined anyway.

No matter how much raw power a CPU can deliver, the power is nothing if software can't use it in real world application (and without too much work for devs).

I'm sorry if I didn't communicate my point well, you stated almost exactly what I meant. I have said it for a while now: core counts don't matter. What matters is how it performs in the applications that you as use. That's what I meant by workload - real world apps, not some theoretical measure of power.

[xsecret]

The real world strange performances of BD seems to come from major changes in instructions throughput. Some of them are much (much) faster than Thuban, some other are much (much) slower. So depending on which instructions are used by the benchmark/software, you will get various results. Use FDIV and you'll see major gain, use FCOS and that will sux more than a Prescott. Something is wrong in µops decoding on many instructions.

[radaja]

Some depressing info here,but i will stay optimistic
this new µarch from AMD might have some bugs
but i can and will be improved upon,its better than
them just shrinking the old,they now have somewhere
to go and it will be good after it matures

[informal]

@ xsecret

Do you think this can be microcode's fault? Can AMD forcefully impair Bulldozer samples with additional overhead via firmware? I know they can turn off prefetch completely .

[Dimitriman]

I see it like this then, if AMD cannot release a product on a new architecture it intends to use for 5+ years to come, at least moderately faster than its previous top product (nevermind slower), then it will:

1- Say goodbye to the desktop market and become a server only provider
2- Hold on to a 1-5% share of the desktop market where price is the only element of choice such as in the majority of 3rd world countries and China
3- Go bankrupt from lack of funds and casflow to invest in developing a new replacement project/ be bought out by Nvidia or other

Don't tell me AMD does not need Bulldozer to survive because everyone knows very well that this company won't survive on the graphics department alone...

[xsecret]

I see it like this then, if AMD cannot release a product on a new architecture it intends to use for 5+ years to come, at least moderately faster than its previous top product (nevermind slower), then it will:

1- Say goodbye to the desktop market and become a server only provider
2- Hold on to a 1-5% share of the desktop market where price is the only element of choice such as in the majority of 3rd world countries and China
3- Go bankrupt from lack of funds and casflow to invest in developing a new replacement project/ be bought out by Nvidia or other

Don't tell me AMD does not need Bulldozer to survive because everyone knows very well that this company won't survive on the graphics department alone...

Bulldozer will be able to compete with Intel mainstream product in the 150-250$ range. That's a lot better than the actual product line. But I know many guys are expecting more for a "brand new" architecture.

[Oese]

The real world strange performances of BD seems to come from major changes in instructions throughput. Some of them are much (much) faster than Thuban, some other are much (much) slower. So depending on which instructions are used by the benchmark/software, you will get various results. Use FDIV and you'll see major gain, use FCOS and that will sux more than a Prescott. Something is wrong in µops decoding on many instructions.

makes sense.. damn this will give nightmare public relations for amd, even if the design would somewhat mature. Worst case will be if it depends on special compilers and therefore software/hardware market. Then it might only be good in server department with specialized solutions... And even then it will be hard..

Even if bulldozer would rock with specialized software, I doubt AMD would get the deserved support.. It may be 10 years before its time, and even if the future will proof it like with x64 and internal imc, if AMD sacrificed too much everyday-performance, it will be a serious problem for bulldozer success...

[Oese]

it should be 10-15% better then thuban, and reach higher clocks/suck less power, everything else would be a great fail..

[FlanK3r]

xsecret: I dont know now, where is it true. One man with FX B2 say here, single performance is better than Phenom II and u said "its whorse" ...:-/

[marten_larsson]

Maybe that's the problem. People expect bulldozer to be much faster than a chip that's considered very good (SB 2600K) when in reality that would be a huge step up for AMD in desktop.

[haylui]

xsecret: I dont know now, where is it true. One man with FX B2 say here, single performance is better than Phenom II and u said "its whorse" ...:-/

smoke screen is everywhere over BD on the internet..i can't take it no more!!!

[Dimitriman]

Bulldozer will be able to compete with Intel mainstream product in the 150-250$ range. That's a lot better than the actual product line. But I know many guys are expecting more for a "brand new" architecture.

I find it difficult to understand how AMD can possibly repair the damage caused by another Barcelona while Intel will bring a world of pain in just 6 months time with Ivy Bridge + S2011 which will make the performance gap between AMD and Intel even greater than it is today with Phenom II vs. i7.

AMD sunk terribly with the Barcelona fiasco, another fiasco would confine them to a new VIA in the desktop market.

[undone]

The real world strange performances of BD seems to come from major changes in instructions throughput. Some of them are much (much) faster than Thuban, some other are much (much) slower. So depending on which instructions are used by the benchmark/software, you will get various results. Use FDIV and you'll see major gain, use FCOS and that will sux more than a Prescott. Something is wrong in µops decoding on many instructions.

Better prefetcher, especially bulldozer's double prefetcher will have a lot of work to do while the longer pipeline cause higher instruction latency, to save some cycle that have been wasted. If prefetcher is down, those higher latency will cause performance disaster cuz you have to spend more cycle to fetch code from cache/memory.

[maltrabob]

...This said I wonder what u guys will think if benchmarks published after the NDA are similar to actual ones ? Reviewer sucks ? AMD issue with a TLB or anything else like for first phenom ? crappy µarch ?

In that case I will calmly try to find my Kindle I put somewhere after I bought it and finally start reading books.

Joking apart. You make impression you already know the final performance when I look at some of your posts:

"Anyway, I just put my hand on a final, retail CPU with retail box. About performances, nothing changed. So, if this CPU isn't with "shipping performance", the fix will come post launch. And that would sux a lot."

"I have the same performances with the B2 chip since weeks and more important, the benchmarks published under NDA by AMD itself are really close to that."

"Bulldozer will be able to compete with Intel mainstream product in the 150-250$ range."

On the other hand there is also a lot of confusion you are bringing to this discussion as well, for example you said that "The issue with previous stepping seems related to power and only power, not internal µarch." which can be seen as there wasn`t any problem with BD microarchitecture whatsoever. But then you said "This said, a big problem in BD µarch will comes from the Integer unit." which on the contrary is a claim that contradicts your previous opinion.

[xsecret]

xsecret: I dont know now, where is it true. One man with FX B2 say here, single performance is better than Phenom II and u said "its whorse" ...:-/

It could be better, it could be worst, depending on which benchmark you use.

[xsecret]

On the other hand there is also a lot of confusion you are bringing to this discussion as well, for example you said that "The issue with previous stepping seems related to power and only power, not internal µarch." which can be seen as there wasn`t any problem with BD microarchitecture whatsoever. But then you said "This said, a big problem in BD µarch will comes from the Integer unit." which on the contrary is a claim that contradicts your previous opinion.

Depend on what is "bad performance". If it's "not able to compete with Intel product for the same price", then all previous stepping than B2 are bad due to that bug. If it's "awesome roxxing performance able to destroy a Core i7 990X" (many guys expect that), well, even the current final CPU is "bad", but that's µarch related.

[PaganII]

That's all that really matters, the performance. If BD is not even faster than Thuban, then AMD is going to lose many life long customers to Intel, myself included. At this point in time, it's not looking to rosy for BD.

I don't think so. I stuck with 939s till Phenom II was released.

[rog]

The issue with previous stepping seems related to power and only power, not internal µarch.

Anyway, I just put my hand on a final, retail CPU with retail box. The box art is similar than the ones leaked some months ago, except they changed the colors and replaced the LGA CPU with a PGA one on the front. About performances, nothing changed. So, if this CPU isn't with "shipping performance", the fix will come post launch. And that would sux a lot.

Can you confirm B2G is the final stepping? I don't know if you aware of it but you should use a Bios with at least Agesa 0090 or 0091 code. If you have an old Bios with 0075 Agesa you are losing lots of performance in some applications.

[m411b]

smoke screen is everywhere over BD on the internet..i can't take it no more!!!

I gave up 2 months ago!

I can see this BD with it's 4 "Clusters" and 8 "Cores" turning into a P4 Hyper-threading situation, which today is terrible!

[radaja]

Looks like i will need a bigger PSU