Manufacturing Process and SB VS Phenom II

[azadali]

Sandy Bridge clearly wins over Phenom II in almost each and every respect. However, the manufacturing process of SB is 32nm and that of Phenom II is 45nm. I was wondering, how Phenom II would have performed if it was manufactured also on 32nm process. So the question arises whether it is only the physical limitation of Phenom II or SB really has a superior architecture?

Azad

 

[khon]

Nehalem was also superior to Phenom II, even though it was made with a 45 nm process. And we know that Sandy Bridge is an improved architecture compared to Nehalem.

Or in short:

Sandy Bridge > Nehalem > Phenom II

 

[Concillian]

AT declared the Phenom II slower, clock for clock, compared with Core2 architecture:
http://www.anandtech.com/show/2702/4

The comparison isn't any better 2 generations of Intel down the road.

We will be able to compare directly when Llano is released, since it will essentially be a revised Phenom II on 32nm. Don't set your expectations too high, though. Performance at a given clockspeed doesn't really change with a die shrink, and AMD is using additional space for the GPU rather than CPU architecture improvement.

 

[Idontcare]

We will be able to compare directly when Llano is released, since it will essentially be a revised Phenom II on 32nm. Don't set your expectations too high, though. Performance at a given clockspeed doesn't really change with a die shrink, and AMD is using additional space for the GPU rather than CPU architecture improvement.

Yeah, Llano vs PhenonII should be similar to the 65nm Athlons versus the 90nm ones. 65nm Athlons were developed in the shadow of Phenom's release and were scaled with similar goals in mind - to be cost sensitive and for laptop markets.

 

[Soleron]

Yeah, Llano vs PhenonII should be similar to the 65nm Athlons versus the 90nm ones. 65nm Athlons were developed in the shadow of Phenom's release and were scaled with similar goals in mind - to be cost sensitive and for laptop markets.

Former AMD employee redpriest was saying better than Conroe -> Penryn IPC increase...

 

[podspi]

Former AMD employee redpriest was saying better than Conroe -> Penryn IPC increase...

But won't it also be hit with the lack of L3 cache? I realize in some applications it won't matter, but overall performance will suffer. The increase in IPC could just counteract the loss of L3 cache.

If they really are leaving AM3 w/out BD, they should release PhIII with the IPC improvements and cache. I bet they'd sell if they could clock them > 4ghz+.


Probably could make them for peanuts given how small the K10 core is...

I'm... mostly kidding.

 

[Idontcare]

Former AMD employee redpriest was saying better than Conroe -> Penryn IPC increase...

Llano's IPC increase over Thuban is going to be greater than penryn's IPC over Conroe!?

Kudos to AMD if they pulled that off, I'm intentionally delaying my laptop upgrade this year waiting for Llano...but it seems like an strange allocation of R&D dollars for AMD to have invested so much into improving an EOL microarchitecture at the same time they are spreading those R&D dollars thinner and thinner while trying to kick-off two totally brand new microarchitectures (Bulldozer and bobcat).

 

[drizek]

Llano is the launch vehicle for mainstream Fusion. The Future is Fusion, and I think AMD wanted to start off on the right foot.

 

[Triskain]

Llano's IPC increase over Thuban is going to be greater than penryn's IPC over Conroe!?

Kudos to AMD if they pulled that off, I'm intentionally delaying my laptop upgrade this year waiting for Llano...but it seems like an strange allocation of R&D dollars for AMD to have invested so much into improving an EOL microarchitecture at the same time they are spreading those R&D dollars thinner and thinner while trying to kick-off two totally brand new microarchitectures (Bulldozer and bobcat).

Goto-San has a nice picture summing up the improvements that went into Llano:

click here

 

[Tuna-Fish]

It is very hard to theoretically compare different manufacturing processes -- there are many independent factors, and you don't know how much weight should be placed on any of them. So take all this with a large mountain of salt -- I openly admit that I don't think this kind of analysis can be very accurate.

Now that the disclaimer is out of the way, the various foundries released some interesting data at IEDM 2008. GF, AMD's foundry, is part of the Common Platform Alliance that does process research, so the IBM SOI numbers are likely representative of the process AMD will use. Llano and BD will be made on the "IBM 32nm SOI", and Phenom 2 was made on the "IBM 45nm SOI" that is found from lower left. The second "IBM 45nm SOI" with much better transistor performance is a research HKMG process that afaik has never been used in production.

The first chart is about transistor performance, both values count (I don't know at what weighs, NFET is probably more important) and upper right is better. The first thing to notice is that as far as transistor performance is concerned, AMD has been operating at a huge disadvantage for years. IBM 45nm is closer in performance to Intel 65nm than 45nm, and Intel has also went for the smaller process faster. This is about to change -- if only for the time until Ivy Bridge and 22nm is out, for the first time in years AMD will be producing processors on a process that is fundamentally competitive with Intel.

This repeats with density numbers (remember that density also affects speed because signals will have less distance to travel). IBM 45nm SOI is not really all that competitive with Intel 45nm, but 32nm SOI is actually denser than Intel 32nm for memory, while still losing out in logic density.

This is a *very* narrow view of only a few of the parameters affecting the speed of a process, leaving out many important factors (for example, how fast are the upper metal layers?), but based on what I see, I don't think the "bigger advance than Penryn" to be at all questionable. If anything, it's conservative -- on transistor speed for AMD moving from 45nm to 32nm is like Intel moving from 65nm to 32nm.

 

[Idontcare]

<snip>...based on what I see, I don't think the "bigger advance than Penryn" to be at all questionable. If anything, it's conservative -- on transistor speed for AMD moving from 45nm to 32nm is like Intel moving from 65nm to 32nm.

I thought the reference to penryn over conroe was in regards to IPC...which is entirely architecture dependent and not process technology dependent (insofar as we hold process-induced minimum latencies to be constant).

Operating clockspeeds and absolute performance/watt metrics are process dependent (as you know, I know I'm not telling you anything new here)...they go towards giving Llano really good performance/watt possibilities compared to Phenom II (which we all expect) but the real kicker is the prospects of considerable IPC improvements above and beyond the clockspeed and power-consumption benefits of the new process tech.

 

[Tuna-Fish]

I thought the reference to penryn over conroe was in regards to IPC

Yes, it appears to be. So it's going to get both plenty of IPC improvements and a very nice boost thanks to process tech. Who needs bulldozer?

 

[Martimus]

Now that the disclaimer is out of the way, the various foundries released some interesting data at IEDM 2008. GF, AMD's foundry, is part of the Common Platform Alliance that does process research, so the IBM SOI numbers are likely representative of the process AMD will use. Llano and BD will be made on the "IBM 32nm SOI", and Phenom 2 was made on the "IBM 45nm SOI" that is found from lower left. The second "IBM 45nm SOI" with much better transistor performance is a research HKMG process that afaik has never been used in production.

The first chart is about transistor performance, both values count (I don't know at what weighs, NFET is probably more important) and upper right is better. The first thing to notice is that as far as transistor performance is concerned, AMD has been operating at a huge disadvantage for years. IBM 45nm is closer in performance to Intel 65nm than 45nm, and Intel has also went for the smaller process faster. This is about to change -- if only for the time until Ivy Bridge and 22nm is out, for the first time in years AMD will be producing processors on a process that is fundamentally competitive with Intel.

This repeats with density numbers (remember that density also affects speed because signals will have less distance to travel). IBM 45nm SOI is not really all that competitive with Intel 45nm, but 32nm SOI is actually denser than Intel 32nm for memory, while still losing out in logic density.

This is a *very* narrow view of only a few of the parameters affecting the speed of a process, leaving out many important factors (for example, how fast are the upper metal layers?), but based on what I see, I don't think the "bigger advance than Penryn" to be at all questionable. If anything, it's conservative -- on transistor speed for AMD moving from 45nm to 32nm is like Intel moving from 65nm to 32nm.

Here is the same data updated from IEDM 2010: http://realworldtech.com/page.cfm?ArticleID=RWT021511004545&p=7

 

[Idontcare]

Here is the same data updated from IEDM 2010: http://realworldtech.com/page.cfm?ArticleID=RWT021511004545&p=7

Unfortunately there appears to be little, if any, process tech data published regarding GloFo's 32nm SOI HK/MG for Llano and BD.

 

[Martimus]

Unfortunately there appears to be little, if any, process tech data published regarding GloFo's 32nm SOI HK/MG for Llano and BD.

Very true. This seems strange to me, as I would think that information would be available since GF would want potential customers to know what they are getting. Perhaps AMD takes up all 32nm SOI production, so they don't make that info available to anyone else?

 

[Vic Vega]

Sandy Bridge clearly wins over Phenom II in almost each and every respect. However, the manufacturing process of SB is 32nm and that of Phenom II is 45nm. I was wondering, how Phenom II would have performed if it was manufactured also on 32nm process. So the question arises whether it is only the physical limitation of Phenom II or SB really has a superior architecture?

Azad

Was your entire reason for registering here to post this statement?

 

[betasub]

^ Well, to be fair, he did also cross-post it into Highly Technical too.

 

[bryanW1995]

seems like a good reason to register to me. AT is a technical website, he had a technical question, it's like a match made in internet heaven!