AMD "Zembezi

[illfindu]

Hey im looking towards the future and I'm eyeing the AMD 8 core bulldozer CPU'S coming down the line. I'v seen some source say there going to use a AM3+ Socket and im wondering if that means youll be able to toss one in a current AM3+ compatible board?
Will my current http://www.newegg.com/Product/Product.aspx?Item=N82E16813130297&nm_mc=OTC-Froogle&cm_mmc=OTC-Froogle-_-Motherboards+-+AMD-_-MSI-_-13130297 msi 870A Fuzion work i noticed just now that its a AM3 not a AM3+ I'm guessing that means ill need a new mother board cause the sockets are comparable?

 

[Haserath]

I thought AMD had already said the scaling was 90% per core if both cores were in use. Maybe I got that from somewhere else.

That 180% would be with 2 threads on the module, just to be clear.

Percentages do not add up.

90% + 90% = 90% not 180%

What you're effectively saying is that the scaling is over 100% and near 200%. In other words that performance of a single Bulldozer Module (containing two "cores") is more than that of a fully functional set of two complete cores.

Not sure if I am explaining myself properly.[/quotemsg]
I don't think I'm making this very clear either... :lol:

2 complete cores- 200%

Bulldozer module- 180%

1 complete core- 100%

Just for integer performance, this is what it should act like if AMD made each module into a core with 'Hyperthreading', unless I'm mistaken.

 

[dogman_1234]

Zambezi will be good. I feel this new tech AMD is bringing will change the CPU interface a little for the future.

I feel that implementing Modules, cores can work as a team, reducing the most needed power, time and resources to get a job done. Like two guys lifting metal across a waterway.

 

[kg2010]

Percentages do not add up.

90% + 90% = 90% not 180%

What you're effectively saying is that the scaling is over 100% and near 200%. In other words that performance of a single Bulldozer Module (containing two "cores") is more than that of a fully functional set of two complete cores.

Not sure if I am explaining myself properly.

I don't think I'm making this very clear either... :lol:

--------------

2 complete cores- 200%

Bulldozer module- 180%

1 complete core- 100%

Just for integer performance, this is what it should act like if AMD made each module into a core with 'Hyperthreading', unless I'm mistaken.

So basically 1 module is slower than 2 cores? Which cores are these being compared to?

AMD's current cores, or Intel's cores?

But since BD will have 8 cores...

arrgghhhh, can we have some benchmarks already?

Here are 6 AMD cores vs. 4 Intel Cores:
http://www.anandtech.com/bench/Product/288?vs=146

 

[ElMoIsEviL]

Percentages do not add up.

90% + 90% = 90% not 180%

What you're effectively saying is that the scaling is over 100% and near 200%. In other words that performance of a single Bulldozer Module (containing two "cores") is more than that of a fully functional set of two complete cores.

Not sure if I am explaining myself properly.

I don't think I'm making this very clear either... :lol:

2 complete cores- 200%

Bulldozer module- 180%

1 complete core- 100%

Just for integer performance, this is what it should act like if AMD made each module into a core with 'Hyperthreading', unless I'm mistaken.

It is still out of 100%.

I get what you're saying but it confuses the math. What you're saying is that a Bulldozer Module performs at 90% that of two full cores. Two bulldozer modules perform at 90% of four full cores. (or 90/100).

S'all good... I get what you meant now but I still don't see how it could perform at 90% across the board (especially not under FP heavy loads) as the two Cores share the same Floating Point units. I'm thinking that the 90% is indicative of Integer-heavy and FP-light workloads.

 

[Haserath]

What was this thread about again? [:randomizer]

.
.
.
.
.

Oh yea, Zambezi.

Yea, yea, I tried to get this back onto topic, but I think I threw it further off course with my comment. I guess we're onto a tangent about fanboys for now.

BTW, I'm biased towards AMD(I don't know why, it must be me wanting the little guy to make things even again), but I try to keep an even playing field in my train of thought.

Here's hoping Zambezi is better than expected.

 

[Haserath]

So basically 1 module is slower than 2 cores? Which cores are these being compared to?

AMD's current cores, or Intel's cores?

But since BD will have 8 cores...

arrgghhhh, can we have some benchmarks already?

I guess you would have to compare them to what would be complete Bulldozer cores, so there really isn't anything to compare them to.

It's worth the slight performance loss to gain power and die savings. Too bad AMD didn't want to go with a complete performance monster and say each module was a core, that would absolutely wipe the floor with Intel in FPU and highly threaded integer tasks.

I don't know how much the die area costs, though, because a chip like that would almost double the size of the entire chip. Making it more power hungry(thus lower clocks) and cost almost double for each die. Not to mention that might require more than dual channel memory for something like that.

 

[kg2010]

I guess you would have to compare them to what would be complete Bulldozer cores, so there really isn't anything to compare them to.

It's worth the slight performance loss to gain power and die savings. Too bad AMD didn't want to go with a complete performance monster and say each module was a core, that would absolutely wipe the floor with Intel in FPU and highly threaded integer tasks.

I don't know how much the die area costs, though, because a chip like that would almost double the size of the entire chip. Making it more power hungry(thus lower clocks) and cost almost double for each die. Not to mention that might require more than dual channel memory for something like that.

One step at a time, one step at a time. 😉

They do have 16 core server processors, so just turn those chips into Enthusiast chips and AMD is golden, and oh yeah, make sure it can do 5 GHZ on air, AND price them @ $329. :lol:

I will jump ship. 😉

JF-AMD said this before:

Want a processor 2X faster than Intel? We could do that. You couldn't afford it.

Want a processor with 5GHz and 16 cores? Probably possible. But how would you cool it?

 

[dogman_1234]

arrgghhhh, can we have some benchmarks already?

They should be out soon. I think by the time CeBit is out, Anandtech, tom's Hardware, and others will release the benches.

 

[dogman_1234]

One step at a time, one step at a time. 😉

They do have 16 core server processors, so just turn those chips into Enthusiast chips and AMD is golden, and oh yeah, price them @ $329. :lol:

I will jump ship. 😉

AMD Netburst anyone?

 

[Haserath]

One step at a time, one step at a time. 😉

They do have 16 core server processors, so just turn those chips into Enthusiast chips and AMD is golden, and oh yeah, make sure it can do 5 GHZ on air, AND price them @ $329. :lol:

I will jump ship. 😉

JF-AMD said this before:

Want a processor 2X faster than Intel? We could do that. You couldn't afford it.

Want a processor with 5GHz and 16 cores? Probably possible. But how would you cool it?

Yea, I guess yields would also be an issue there, because doubling the size also more than doubles the chance of something being wrong on the chip.

That first one seems to be a little off, because if they doubled their IPC per core with a small die increase they would have this in the bag. I'm not sure how they would really do that, unless they made major improvements to their branch predictor, load/store, or added something like Intel's micro-ops cache to help speed things up.

I think I have had way too much caffeine. My mind just keeps racing with too much of this...I'm off to bed before I hurt myself.

 

[enzo matrix]

Percentages do not add up.

90% + 90% = 90% not 180%

What you're effectively saying is that the scaling is over 100% and near 200%. In other words that performance of a single Bulldozer Module (containing two "cores") is more than that of a fully functional set of two complete cores.

Not sure if I am explaining myself properly.

90/100= 90%
180/200= 90%

So yes, they are the same.

What he is effectively saying is that on a single thread you get 100% throughput if there is not another thread on the core's partner but if there is, then the throughput is 90% for one core. If the baseline for one core is 100%, then the performance of two threads with each on a separate module will be 200%. 180% for the same module.

 

[fazers_on_stun]

Of course there is. Still its the same idea. Intel did with with the FSB. Still it worked for both for the most part.

Agree 100%. Really the FSB solution only started slowing down when you got to quad socket or higher. For desktop it was more than adequate, as witnessed by the benchmarks.

I myself am a MT Dew fanboy. Love it.

Ahh, but are you a Mountain Dew Red or Nuclear Green fanboy?? 😛

 

[dogman_1234]

He's a green! 😛

What about the FSB? Why is it in Quad socket that it gets slower?

 

[fazers_on_stun]

He's a green! 😛

What about the FSB? Why is it in Quad socket that it gets slower?

Because the FSB had lower bandwidth and it got saturated once you went over 2 processors. In fact, I believe Intel went with 2 FSB chipsets in their quad-socket servers because of that issue, and of course then there was one more step involved in communicating between the chipsets..

 

[Reynod]

People have disagreements... people are human. Public disagreements are intellectual. You ought not seek to secure yourself away from reality. Reality is what it is.

If people are allowed to work out their disagreements... a problem is solved. When people are not allowed to work out a disagreement... animosity remains.

You cannot, by force, stop human beings from being human beings. Authority works counter to Freedom, Logic and the advancement of knowledge.

Don't like it... skip it.

When people create an environment where one has to be worried about what one says... nothing of value is ever said. That's called a police state.

But if you and Keith don't settle down I will remove both your posts entirely from the thread.

Ditto for anyone else wanting to exchange insults.

Focus on the topic and not on the person ...

 

[saint19]

@All posters:

Thread is about AMD "Zambezi" post about that or be banned and thread closed. Any other discussion or personal problem use PM.

 

[keithlm]

But if you and Keith don't settle down I will remove both your posts entirely from the thread.

Ditto for anyone else wanting to exchange insults.

Focus on the topic and not on the person ...

So we can now see that it is acceptable for Elmo to personally attack people and it is not allowable for someone to point out his shortcomings and defend against his unwelcome personal attacks. (You do realize you left several of his less than intelligent personal attacks and you also allowed him to yet again post even more crap. And I'm not allowed to say anything about his snarky personal comments because everything I post gets removed.)

Good to know exactly what level of standards this forum embraces.

 

[Hellboy]

But if you and Keith don't settle down I will remove both your posts entirely from the thread.

Ditto for anyone else wanting to exchange insults.

Focus on the topic and not on the person ...

go for it Reynod...


I came back after a year and now i cant be bothered anymore after 4 posts


still the same old crap - different colour thats all


infact im almost on the verge of giving up on pcs all together,,, its all down to the kpsm syndrome

knobs per square mile

 

[fazers_on_stun]

Personally I'm glad to see the mods step in finally. And despite the complaints about "standards", it is far better here than, say, over on AMDzone.

Back on topic, I'm thinking Zambezi's long pipeline might not be too bad if indeed they have improved their branch prediction as JF states. I sorta doubt it'll be as good as Intel's SB but far better than Netburst. And a long pipe should make for higher clocks. Intel also recycled a number of Netburst ideas with SB - I guess the technology has finally caught up to enable the design ideas.

And Zambezi's CMT may be the wave of the future as well - the areal cost is only something like 12% vs. SMT's 5%, and should provide better heavy thread performance. I wouldn't be surprised if Intel beefs up their SMT in Haswell.

IHMO, winning designs are not only about innovation, but also timing. So Intel's using MCM + FSB link for C2Q was right for the time - Nov. 2006 IIRC - because FSB provided more than enough bandwidth for desktop and even 2-socket server. In contrast, AMD's "native" quad core Barcelona turned out to be something of a disaster - a year late and way underperforming. Not that I'm implying that the HT links were the cause of the problems, but I'm sure the added complexity didn't help matters any.

It all comes down to execution and performance, but Intel hanging on to SMT may be something of a bad-timing mistake - the benchies should tell when BD is released.

 

[sonoran]

OK, first off jimmy, the cores are real cores. Please give me all of the reasons why you believe they are not.

And "because they are not" does not count. Intel shares a single scheduler between their integer core and their FPU. Does that mean they are not real cores? Older versions of the 7000 series had shared L2 cache. Does that mean they were not real cores?

When the world went from single core to dual core, everything in the processor went from being completely dedicated to shared. At that point manufacturers, including intel, have made decisions about what components to share and what to leave discrete. Bulldozer is no different.

There was a lot of discussion years back about whether Intel MCM quad cores were "real quad cores" or not. One could easily argue that this discussion was egged on by AMD itself with this little gem: AMD sends "Multi-Core Processing for Dummies" to IDF.

Given that 2 Bulldozer cores share FP resources, one has to expect that some people are going to say they're "not 2 real cores".

But your point is well made - what defines a "core" is ever changing, and what resources those cores should/should not share will always be a cost/benefit tradeoff.

* Not speaking for Intel Corporation *

 

[dogman_1234]

They are 'true' cores, just sharing cache and pipelines.

 

[jimmysmitty]

He's a green! 😛

What about the FSB? Why is it in Quad socket that it gets slower?

Basically the memory bandwidth sucked for more than 2 CPUs. They had to share it. Now with QPI, Intel just adds more links so they don't have to share as much while they could only have one FSB.

As for true cores, still debatable. We will see when it hits. Should get a nice long article explaining it.

 

[fazers_on_stun]

There was a lot of discussion years back about whether Intel MCM quad cores were "real quad cores" or not. One could easily argue that this discussion was egged on by AMD itself with this little gem: AMD sends "Multi-Core Processing for Dummies" to IDF.

http://img.tomshardware.com/us/2006/09/25/amd_mc_processing.jpg

Given that 2 Bulldozer cores share FP resources, one has to expect that some people are going to say they're "not 2 real cores".

But your point is well made - what defines a "core" is ever changing, and what resources those cores should/should not share will always be a cost/benefit tradeoff.

* Not speaking for Intel Corporation *

LOL - I remember that stunt by AMD - sorta backfired on them when Barcelona was so late and slow and then got recalled almost immediately due to the TLB bug. I'm sure the Intel engineers had a huge belly laugh after that 😛..

And now AMD is doing things like sending fake boxes of chocolates to Intel over the SB chipset problem, and a new advertising campaign "Ready, Willing & Stable". Given their track record, you'd think they would wait and make sure Bulldozer doesn't blow up in their faces first 😗 ..

 

[EXT64]

Yeah, this is a bit early for them to be gloating (actually, way too early).

 

[dogman_1234]

..or, they have nailed it without us knowing it!

Just sayin'. 😀