AMD Zambezi news, info, fans !

[MaddMutt]

Hey Guys,

Here is test of my rig

If I understand the GREAT KNOWLEDGE that has been shared in here. The INTEL 4c/8t 2600 is doing 5.48, this is 4 REAL cores @ 4.00 and 4 (FAKE) SMT cores @ 1.48. The real cores are probably doing more but to make it simple math, Each SMT core is doing .37% additional work as compared to a real core. If AMD's BD can do as chew said earlier (think of it as a 4c/8t proc) 4 real cores @ 4.00 and if each of the 4 CMT cores can work at 50% or better then I think AMD does have a winner.

But back to my original thought, a PH II moved to 32nm process with some good optimizations should still have some life left in it's arch???? I think that my test shows that this arch can still compete if AMD can do a small respin on it.


Let the debate begin.


And yes I am A AMD LOVER

MaddMutt

[BeepBeep2]

No.
Single thread on AMD "STARS" sucks.

[MaddMutt]

True....

But more and more software are becoming multi-threaded.

Does this not make AMD's CPU's better????

[dess]

The INTEL 4c/8t 2600 is doing 5.48, this is 4 REAL cores @ 4.00 and 4 (FAKE) SMT cores @ 1.48. The real cores are probably doing more but to make it simple math, 4 SMT cores are doing .37% additional work as compared to real cores.

There are no real + SMT cores in Intel's uarch'es. There are cores with SMT support.

If AMD's BD can do as chew said earlier (think of it as a 4c/8t proc) 4 real cores @ 4.00 and 4 CMT cores at 50% or better

Neither are real + CMT cores in Bulldozer. There are CMT-type modules.

I'm sorry, but this wording is just wrong and technically inappropriate.

Otherways, you're comparing it this way: x * 1.37 vs. x * 1.5
But, it really looks like this: x * 1.37 vs. y * 1.5 -- and we don't yet know the ratio between x and y...

[MaddMutt]

Thank You,

This is why I STATED earlier about all this knowledge you people on this forum are freely sharing

I do not have a Intel system in my house but at work when I go to set the affinity on a program, it is listed as CPU 0 through CPU 11. It does not make a distinction between a real core and a smt one. It looks as if I have 12 cores in my CPU.

For some of us that do not have the full grasp, you are willing to explain it instead of flaming us.

This is why I joined

MaddMutt

[Mechanical Man]

Thank You,

This is why I STATED earlier about all this knowledge you people on this forum are freely sharing

I do not have a Intel system in my house but at work when I go to set the affinity on a program, it is listed as CPU 0 through CPU 11. It does not make a distinction between a real core and a smt one. It looks as if I have 12 cores in my CPU.

For some of us that do not have the full grasp, you are willing to explain it instead of flaming us.

This is why I joined

MaddMutt

Win 7 and maybe Vista are aware of SMT. That means OS scheduler can take advantage when scheduling threads so that first it puts heavy ones to cores and then adds light ones to cores and SMT cores. Xp cant do that, basically that means Win7 can use SMT arch more efficiently.

Same could be done to BD. Depending on how Turbo works it could be best in games to put heaviest ones to other half of module and use other to run lighter threads.

[dess]

at work when I go to set the affinity on a program, it is listed as CPU 0 through CPU 11. It does not make a distinction between a real core and a smt one. It looks as if I have 12 cores in my CPU.

Again, there are no "real" and "smt" cores here. Those are the threads (of execution) this CPU is capable of working on at the same time. Every two represents one core that runs two threads in SMT, the way most of resources are shared between these two threads, so they are slowing each other down. The benefit is that the slow-down is usually less than 50% per thread. The net value can be somewhere between ~90% (worst case) and ~140% (best case), depending on the task. If there are only one thread running on the core -> no slow-down. We can speak about primary and secondary threads here. (I don't think there are prioritization between them, though, HW wise.)

In the other hand, Bulldozer has CMT based modules, where only certain resources are shared (there are almost two full cores - narrower than before, though), that means minimal slow-down. The net value here can be somewhere between ~150% and ~200%. So, there are less or even no benefit in running only one thread per module. Better yet, if there are no many threads to run, run them on less modules with elevated clocks, and shut down the rest. I think we are waiting for an update to Win 7 for this to work better.

For some of us that do not have the full grasp, you are willing to explain it instead of flaming us.

Well, it's more useful, isn't it? I guess this forum (and some others) is not for the martial teenagers...

[DedEmbryonicCe1]

Interview yesterday:
http://www.youtube.com/watch?v=PlwIFFHrNW4
They are talking about the Middle East market and say FX processors will be available in October.

[sviola]

The main problem is the 32 nm process tech which is not matured enough and there is a capacity problem too. I think both of these issues will be solved in few months.
It's always difficult to start a brand new microarchitecture on a brand new and complex (SOI + gate first HKMG) process tech.

So, AMD shouldn't have done it this way, should have used Intel tic-toc strategy: new architecture on matured process, new process on matured architecture. Would clearly make things smoother. If they had transitioned ph 2 a couple of months ago to 32 nm, they could have improved its lifecycle a bit and now would be releasing BD with the process sorted out.

But, this doesn't matter anymore. Just waiting on BD to make a decision on which way to go for my new rig (and, I'm leaning for AMD this time).

[Opteron146]

So, AMD shouldn't have done it this way, should have used Intel tic-toc strategy: new architecture on matured process, new process on matured architecture. Would clearly make things smoother. If they had transitioned ph 2 a couple of months ago to 32 nm, they could have improved its lifecycle a bit and now would be releasing BD with the process sorted out.

You surely missed Llano .. that is K10@32nm.

[FlawleZ]

Win 7 and maybe Vista are aware of SMT. That means OS scheduler can take advantage when scheduling threads so that first it puts heavy ones to cores and then adds light ones to cores and SMT cores. Xp cant do that, basically that means Win7 can use SMT arch more efficiently.

Same could be done to BD. Depending on how Turbo works it could be best in games to put heaviest ones to other half of module and use other to run lighter threads.

Do you have any documentation on this? Not denying just would like to know exactly how Win7 & Vista are able to discern between logical and physical cores.

[AKM]

spotreba.png

Power consumption looks very good considering it's 125w vs 95w TDP

[xBanzai89]

No name attention grabber pic

Power consumption looks very good considering it's 125w vs 95w TDP

I suggest taking that down. You cannot post stuff from that guy here on this forum.

[AKM]

I know the guy is a monumental , but I thought this was worth posting since no power numbers were leaked. I'll remove it if the mods tell me to.

[-Boris-]

You surely missed Llano .. that is K10@32nm.

But not meant to be released before Bulldozer, and with very large changes (GPU), so it's an simultanious Tick/Tock for both Llano and BD. They said they had BD ready for 45nm but delayed it for 32nm to put more goodies in it, I think they should have released a quad on 45nm aswell. But as it looks now they maybe skipped that since it wouldn't be able to compete with Phenom II at that time with only 4 cores.

[Manicdan]

I know the guy is a monumental , but I thought this was worth posting since no power numbers were leaked. I'll remove it if the mods tell me to.

heres another thing to consider, CB11.5 is rather weak on BD, thus its possible power consumption could be much higher if the chip was able to perform better at it.
also CB11.5 has a HUGE range of power consumption spikes, trust me i did testing with it before. its not reliable with reporting consumption

[sviola]

You surely missed Llano .. that is K10@32nm.

From what I read, the process is still not matured enough to put out a new architecture on it (I read somewhere that there might be yeald issues with it), but these might be fake rumors.

By the looks of it, BD needed the smaller process to allow for higher clocks with the same TDP (it would have been bad for them to release BD, with these clocks on 45nm and have a TDP of 150+, specially on a marketed-less-power-the-better environment).

[Smartidiot89]

From what I read, the process is still not matured enough to put out a new architecture on it (I read somewhere that there might be yeald issues with it), but these might be fake rumors.

By the looks of it, BD needed the smaller process to allow for higher clocks with the same TDP (it would have been bad for them to release BD, with these clocks on 45nm and have a TDP of 150+, specially on a marketed-less-power-the-better environment).

The rumors of yields might be correct, and if they are those rumors revolve around Llano.

Bulldozer was designed specifically for power efficiency, and efficiency in die space. Or what most people on XS would call it performance/watt, the architecture has high clocks cause it was designed this way. I am quite sure AMD would've gone down the route of higher IPC/lower clocks if that would've been more beneficial.

Now die space, we can call that performance/mm2. The whole die seems to take up LOTS of space, but one Bulldozer module surely is quite small even for 32nm.

[Olivon]

[Manicdan]

The rumors of yields might be correct, and if they are those rumors revolve around Llano.

Bulldozer was designed specifically for power efficiency, and efficiency in die space. Or what most people on XS would call it performance/watt, the architecture has high clocks cause it was designed this way. I am quite sure AMD would've gone down the route of higher IPC/lower clocks if that would've been more beneficial.

Now die space, we can call that performance/mm2. The whole die seems to take up LOTS of space, but one Bulldozer module surely is quite small even for 32nm.

i think the high clocks thing could be to help market it. so few people know the IPC and so if they see 2 options and one is a few hundred mhz faster, they might pick that because they think its faster, even though it could be the same or even slower

and the space for a module is small, but the chip as a whole is still big. which dosnt concern me too much personally. i think intel goes for as much space saving as possible because they are still most of the market and increasing revenue per wafer has a bigger impact than it does for AMD. most of that space is L2/L3 related, and with BD that extra space we still are quite curious about between the modules. i worry about the arguments people will have when the look at perf/mm2, since the module perf might be incredible, the whole chip perf might be very low compared to SB. it might not make a huge difference to end users, but it sure gives alot of fuel for a fire.

[EniGmA1987]

Very interesting Olivon. What source is that video from?

It would be great is that is a real video of BD like it claims and the CPU is running with turbo at 4.2GHz. Even with nearly all cores loaded it is still hitting really high speeds.

[TESKATLIPOKA]

EniGmA1987 the source is OBR.

[Oliverda]

EniGmA1987 the source is OBR.

That bastard should be hunted down by AMD's commando.

[imamage]

Some important info by admin of semiaccurate
http://semiaccurate.com/forums/showp...postcount=1825

I have a feeling that Bulldozer will be out right Before Thanksgiving

[dess]

So, AMD shouldn't have done it this way, should have used Intel tic-toc strategy: new architecture on matured process, new process on matured architecture. Would clearly make things smoother.

I think the original plan was to release it on a mature process: 45nm. But it seems it wasn't working well on that process. So, they had no choice but to take some risk...

If they had transitioned ph 2 a couple of months ago to 32 nm, they could have improved its lifecycle a bit and now would be releasing BD with the process sorted out.

As it was pointed out already, they've did, with Llano. Although, the two projects were running parallelly, so I wonder if there was much new knowledge to transfer.

I think Llano has born rather because it would have been just too much risk to take if it was already with Bulldozer CPU cores.