Could AMD die shrink thuban and add two cores?

[3DVagabond]

Perhaps AMD could adapt old K10 into a module like architecture, with 2x K10 derived integer cores (3-issue) sharing 2x 128 FPUs? A Phenom II x6 has 2 extra ALUs to it's benefit.

Or maybe they could release G34 consumer boards and shrink Magny-cores (remove the multi processor and other non essential server circuitry) and release that to us? Then the real nutters could buy Interlagos if they wanted to. :crazy:

 

[Blitzvogel]

Or maybe they could release G34 consumer boards and shrink Magny-cores (remove the multi processor and other non essential server circuitry) and release that to us? Then the real nutters could buy Interlagos if they wanted to. :crazy:

256 bit AVX I think is an important evolutionary step to have, hence why I'm glad to finally see it on an AMD processor.

 

[Ajay]

You work in industry as a professional? (Serious question) Deal much with IP-space in the course of your work?

It could very well be the case that AMD simply avoided SMT technically for reasons of the IP-space having already been thoroughly mined out.

The best mouse-trap may well indeed be SMT, but AMD had to avoid building a mouse-trap lest they run afoul more litigation, so their engineers settled for 2nd best by going CMT but knowing they could do it on their own terms at least.

Not CPU development. Worked professionally as a firmware engineer with custom ASICs and embedded CPUs - needed to understand the architecture of both for max performance. I thought AMD had a cross licensensing deal with Intel (when the lawsuit was settled). Intel was already using SMT, so if AMD didn't get access to those patents, their lawyers should have been lashed. I assumed they had access to SMT and couldn't understand why they wouldn't have used it given it's success in three different CPU lines.

The thing is we know CMT works, even in bulldozer it is pretty good, CMT didn't fail. What failed was that they CMT'ed a core that has the initial IPC of a K8 instead of a K10.

The performance of bulldozer is not indicative of CMT, it is indicative of a cache limited and decode limited microarchitecture. (for example, yes there are 4 decoders but they cannot be used to issue instructions to both cores at the same time - within same clock - within the same module...that's lame even in theory let alone in practice)

Fair enough. From what I've read at RWT, Lostcircuits, etc. there are problems with the decoder (as you mentioned), the int unit's speed and cache performance. I suppose if AMD could improve all three - maybe they'd have something good to run with. I guess will get a clue as what to expect from AMD engineering when we see BD B3 & Trinity.

 

[nenforcer]

 

[Idontcare]

I thought AMD had a cross licensensing deal with Intel (when the lawsuit was settled). Intel was already using SMT, so if AMD didn't get access to those patents, their lawyers should have been lashed. I assumed they had access to SMT and couldn't understand why they wouldn't have used it given it's success in three different CPU lines.

Even if it is true AMD has carte blanche access to ALL of Intel's patents (I find that hard to believe though, that would be unprecedented if true), that doesn't mean that Intel doesn't need anything more than its own patents to enable SMT.

They (Intel) probably have cross-licensed SMT related patents from IBM, HP, and Oracle (SUN), and perhaps a few others by this time.

It gets interwoven pretty quickly, which is where the expense can come in. Intel might have 6 or 7 out of 10 necessary patents, but AMD still needs to secure the remaining 3 or 4 from other IP holders (these numbers are figurative, not literal).

Again, I am just speculating/guessing here that the IP space had something to do with the decision to go CMT.

 

[Ajay]

Even if it is true AMD has carte blanche access to ALL of Intel's patents (I find that hard to believe though, that would be unprecedented if true), that doesn't mean that Intel doesn't need anything more than its own patents to enable SMT.

They (Intel) probably have cross-licensed SMT related patents from IBM, HP, and Oracle (SUN), and perhaps a few others by this time.

It gets interwoven pretty quickly, which is where the expense can come in. Intel might have 6 or 7 out of 10 necessary patents, but AMD still needs to secure the remaining 3 or 4 from other IP holders (these numbers are figurative, not literal).

Again, I am just speculating/guessing here that the IP space had something to do with the decision to go CMT.

Again, fair enough. I know, especially in IBM's case, they have patents galore on all kinds of compute-IP, even stuff that they only worked out in theory or in a lab but didn't get implemented. So it is reasonable that SMT may have been a $$ sink for AMD.

From SemiAccurate, it appears the a number of the BD CPU problems have been worked out in Piledriver - http://semiaccurate.com/2011/10/17/bulldozer-doesnt-have-just-a-single-problem/

For the sake of AMD, investors and consumers - I hope it's true. Maybe Trinity will be able to match SB (w/o HT) more evenly in CPU performance (although IB is coming out around the same time).

A bit OT...

Honestly, if BD had matched the performance of Gulftown, but for 1/2 the price or less and had a bit more overclocking headroom - I would be considering it for my next build (which won't be any earlier than 4Q12). I use my main CPU on and off for folding (plus a dedicated machine); the equivalent of a couple more cores w/HT would have been worth my while and decently priced. Right now, BDs suck for folding.

 

[podspi]

Even if it is true AMD has carte blanche access to ALL of Intel's patents (I find that hard to believe though, that would be unprecedented if true), that doesn't mean that Intel doesn't need anything more than its own patents to enable SMT.

They (Intel) probably have cross-licensed SMT related patents from IBM, HP, and Oracle (SUN), and perhaps a few others by this time.

It gets interwoven pretty quickly, which is where the expense can come in. Intel might have 6 or 7 out of 10 necessary patents, but AMD still needs to secure the remaining 3 or 4 from other IP holders (these numbers are figurative, not literal).

Again, I am just speculating/guessing here that the IP space had something to do with the decision to go CMT.

Doesn't BD use SMT for the FlexFPU, though? I don't think patents are behind AMD's decision to use CMT. I think they believe(d) CMT would be superior to SMT for high-threaded integer server workloads.

 

[bryanl]

Bulldozer will be magnitudes better by this time next year...

1 magnitude = factor of 10.

2 magnitudes = factor of 100.

Do you have an estimate of the power requirement for the 300 GHz Bulldozer you promise will make its appearance in 2012?

 

[Idontcare]

Doesn't BD use SMT for the FlexFPU, though? I don't think patents are behind AMD's decision to use CMT. I think they believe(d) CMT would be superior to SMT for high-threaded integer server workloads.

If true then you have to ask the other side of the question - why does Intel avoid CMT in favor of SMT? Why does IBM avoid CMT in favor of SMT?

Intel is using it in every processor line, from Atom to Sandy to Itanium.

And Power7 is not a slouch.

These are companies with far more revenue on the line if they whiff in their architecture choices, and yet they are certainly fielding the best of the best and enabling it with SMT, not CMT.

 

[podspi]

You don't know they're actively avoiding CMT, IDC


Somebody has to be the first in trying a novel type of architecture. Sure, you could make the argument that because IBM and Intel don't do it, it shouldn't be done, but don't forget that Intel's first attempt at SMT didn't go over too well either. Research into SMT at Intel is probably quite advanced, and it is implemented in Intel's architecture's quite well. To start developing a CMT architecture might not be cost effective (when compared to developing an improved SMT one).


Evaluating CMT is difficult at this point because we only have one example, and a poor one at that. We might try to consider the following:

1) If ST performance in BD (1M/1T) was equal to SB, how would we view its performance in its (4M/8T) actual configuration?

2) Is the huge die caused by the CMT configuration (in which case it is a non-starter) or something else?


Anyway, my point really was, doesn't AMD use SMT in BD's FlexFPU? If it does, it clearly can't be a patent issue. If they don't, it potentially could be (although they still could consider CMT a viable/superior alternative).

 

[Maximilian]

1 magnitude = factor of 10.

2 magnitudes = factor of 100.

Do you have an estimate of the power requirement for the 300 GHz Bulldozer you promise will make its appearance in 2012?

The power requirement will be magnitudes greater than what it currently is... duh

 

[Idontcare]

You don't know they're actively avoiding CMT, IDC


Somebody has to be the first in trying a novel type of architecture. Sure, you could make the argument that because IBM and Intel don't do it, it shouldn't be done, but don't forget that Intel's first attempt at SMT didn't go over too well either. Research into SMT at Intel is probably quite advanced, and it is implemented in Intel's architecture's quite well. To start developing a CMT architecture might not be cost effective (when compared to developing an improved SMT one).


Evaluating CMT is difficult at this point because we only have one example, and a poor one at that. We might try to consider the following:

1) If ST performance in BD (1M/1T) was equal to SB, how would we view its performance in its (4M/8T) actual configuration?

2) Is the huge die caused by the CMT configuration (in which case it is a non-starter) or something else?


Anyway, my point really was, doesn't AMD use SMT in BD's FlexFPU? If it does, it clearly can't be a patent issue. If they don't, it potentially could be (although they still could consider CMT a viable/superior alternative).

You have an excellent point here! AMD did pursue a number of things that then became the new norm for the rest of the MPU industry.

The IMC, 64bit x86, integrated graphics, CMT...sure AMD isn't the first, but they've done some things that were outside the norm at the time and we all benefited from it in time.

Who knows where CMT goes from here.