AMD sheds light on Bulldozer, Bobcat, desktop, laptop plans

[Idontcare]

Do we have confirmation that 16 core Bulldozer is indeed 60-80% faster than 12 core Magny Cours? Is that an estimate from a ambigous graph or is that roughly true?

The only graph/data I have seen AMD use in their presentations that equates Bulldozer (Interlagos) performance to Magny-Cours is the one in slide 9 of this presentation:

http://phx.corporate-ir.net/External.File?item=UGFyZW50SUQ9MjAzMjR8Q2hpbGRJRD0tMXxUeXBlPTM=&t=1

 

[IntelUser2000]

We can't extrapolate 60-80% from that. That's Johan's claim on his server article.

Eyeballing says the Magny Cours gets 27 on FP and 29 on Integer. Looks like its 80% on FP and 60% on Integer. The graph doesn't even pass 45.

27 x 1.8 = 49
29 x 1.6 = 46

I think Johan knows more.

 

[Idontcare]

Thing is neither Johan nor AMD know the clockspeeds of Interlagos...so they can set minimums based on known/planned clockspeeds of Magny-Cours and anticipated Interlagos clockspeeds but they did that with Phenom/Barcelona too...

 

[IntelUser2000]

That is true. So the question I asked before:

60-80%, is it believable? More? Less?

Magny Cours: 2.2GHz, maybe a 2.4GHz version.

 

[Triskain]

Magny Cours Twelve Cores will be 2.1 GHz at a TDP (not ACP) of 140 Watts. I think that compares nicely to Beckton's Eight Cores with 2.26 GHz at a TDP of 130 Watt.

 

[Kuzi]

That is true. So the question I asked before:

60-80%, is it believable? More? Less?

Magny Cours: 2.2GHz, maybe a 2.4GHz version.

I would be extremely surprised if AMD could/would clock Magny Cours higher than 2.2GHz. Looking at the die size of Istanbul, Magny Cours will be double that at over 690mm2. That's huge!

From our guesstimates, an 8-module/16-core Interlogas may end up smaller than 550mm2. So it is 25% smaller (higher clocks), can run 4 more threads (33%), probably has more L3 (16MB vs 10MB usable on Magny), and we have Anand's words about a 2-Module BD being 10-35% faster than Phenom II X4.

I think it is safe to say Interlogas would be "at least" 50% faster than Magny Cours, at the same clocks, but that's only for software that can use all 16 cores. If Interlogas runs at 2.6GHz, then the advantage can jump to at least 70%.

 

[Idontcare]

I would be extremely surprised if AMD could/would clock Magny Cours higher than 2.2GHz. Looking at the die size of Istanbul, Magny Cours will be double that at over 690mm2. That's huge!

From our guesstimates, an 8-module/16-core Interlogas may end up smaller than 550mm2. So it is 25% smaller (higher clocks), can run 4 more threads (33%), probably has more L3 (16MB vs 10MB usable on Magny), and we have Anand's words about a 2-Module BD being 10-35% faster than Phenom II X4.

I think it is safe to say Interlogas would be "at least" 50% faster than Magny Cours, at the same clocks, but that's only for software that can use all 16 cores. If Interlogas runs at 2.6GHz, then the advantage can jump to at least 70%.

True, Magny-Cours will undoubtedly be TDP/socket limited when it comes to clockspeeds whereas Interlagos' clockspeeds will be all the less limited owing to the use of 32nm.

In other words we might be looking at an artificially forced performance gap (TDP related, not microarchitectural) between Magny-Cours and Interlagos in that graph and in the comments so far.

 

[jvroig]

In other words we might be looking at an artificially forced performance gap (TDP related, not microarchitectural)

What do you mean "artificially forced"? Why do you consider a clockspeed limit due to TDP as an artificial performance limit? Isn't TDP always the top clockspeed delimiter, and not the microarchitecture? Or you mean in the case of Magny-Cours, it could perform a heck of a lot better, the microarchitecture is great, but the TDP limit is more cruel versus the TDP limit in Interlagos owing to the newer and smaller process tech?

I'm not trying to be difficult. I just didn't understand what that statement denotes or implies.

 

[GaiaHunter]

What do you mean "artificially forced"? Why do you consider a clockspeed limit due to TDP as an artificial performance limit? Isn't TDP always the top clockspeed delimiter, and not the microarchitecture? Or you mean in the case of Magny-Cours, it could perform a heck of a lot better, the microarchitecture is great, but the TDP limit is more cruel versus the TDP limit in Interlagos owing to the newer and smaller process tech?

I'm not trying to be difficult. I just didn't understand what that statement denotes or implies.

I would say it is more related to the desktop parts - the phenom II are already clocked decently - if BD architecture doesn't create that performance delta at same clocks speeds than we would have to be talking of 5+ GHz zambezi to achieve 20-35% improvements over phenom II.

 

[Idontcare]

What do you mean "artificially forced"? Why do you consider a clockspeed limit due to TDP as an artificial performance limit? Isn't TDP always the top clockspeed delimiter, and not the microarchitecture? Or you mean in the case of Magny-Cours, it could perform a heck of a lot better, the microarchitecture is great, but the TDP limit is more cruel versus the TDP limit in Interlagos owing to the newer and smaller process tech?

I'm not trying to be difficult. I just didn't understand what that statement denotes or implies.

What Gaia said, and what you posted that I bolded, all together answer the question.

It is not artificial in a laws-of-device-physics sense but rather I mean within the context of what we are trying to leverage the performance gap information to accomplish (divine some expectations regarding performance gap between PhII and Zambezi).

 

[JFAMD]

60-80% over Magny Cours is accurate for Interlagos.

 

[jvroig]

60-80% over Magny Cours is accurate for Interlagos.

It seems to me that AMD is hell-bent on making a big bang on the server front, if not beat Intel altogether in raw performance aside from price/power/performance metric.

That's all well and good, but I wonder how much of this translates to the desktop front? I mean, from here, it looks like the server front will be well-served by the Bulldozer offerings, in fact even just by Magny Cours before it. But what about the desktop? Will AMD's next offerings make as big a splash on the desktop as they will on the server?

(I know you're in charge of the server business, so perhaps you might not want to comment on desktops as that is probably a different division to yours. If so, thanks anyway)

 

[IntelUser2000]

I would be extremely surprised if AMD could/would clock Magny Cours higher than 2.2GHz. Looking at the die size of Istanbul, Magny Cours will be double that at over 690mm2. That's huge!

From our guesstimates, an 8-module/16-core Interlogas may end up smaller than 550mm2. So it is 25% smaller (higher clocks), can run 4 more threads (33%), probably has more L3 (16MB vs 10MB usable on Magny), and we have Anand's words about a 2-Module BD being 10-35% faster than Phenom II X4.

I think it is safe to say Interlogas would be "at least" 50% faster than Magny Cours, at the same clocks, but that's only for software that can use all 16 cores. If Interlogas runs at 2.6GHz, then the advantage can jump to at least 70%.

I agree with the TDP, but the Magny Cours is two Istanbul dies, not really 690mm2.

I think looking at the graph between Magny Cours and Istanbul on that graph, the performance comparison is with similar clocks.

 

[JFAMD]

It seems to me that AMD is hell-bent on making a big bang on the server front, if not beat Intel altogether in raw performance aside from price/power/performance metric.

That's all well and good, but I wonder how much of this translates to the desktop front? I mean, from here, it looks like the server front will be well-served by the Bulldozer offerings, in fact even just by Magny Cours before it. But what about the desktop? Will AMD's next offerings make as big a splash on the desktop as they will on the server?

(I know you're in charge of the server business, so perhaps you might not want to comment on desktops as that is probably a different division to yours. If so, thanks anyway)

I wouldn't say that we are "hell bent" on anything. We're trying to deliver the best performance per dollar and the best performance per watt.

I can't speak to desktops at all and wouldn't want to try.

 

[cbn]

It seems to me that AMD is hell-bent on making a big bang on the server front, if not beat Intel altogether in raw performance aside from price/power/performance metric.

That's all well and good, but I wonder how much of this translates to the desktop front? I mean, from here, it looks like the server front will be well-served by the Bulldozer offerings, in fact even just by Magny Cours before it. But what about the desktop? Will AMD's next offerings make as big a splash on the desktop as they will on the server?

(I know you're in charge of the server business, so perhaps you might not want to comment on desktops as that is probably a different division to yours. If so, thanks anyway)

That is how I feel about AMD. They seem to have a strategy for making impressive large chips for servers but then I wonder how good Bulldozer will be for the rest of us?

But then again for gamers using Eyefinity the spotlight really isn't on the CPU anymore since the GPU(s) effectively become the bottleneck.

 

[Idontcare]

That is how I feel about AMD. They seem to have a strategy for making impressive large chips for servers but then I wonder how good Bulldozer will be for the rest of us?

But then again for gamers using Eyefinity the spotlight really isn't on the CPU anymore since the GPU(s) effectively become the bottleneck.

They beefed up the integer ALUs quite handedly, that will help desktop customers nicely.

Now if they get turbo-boost working and they make it a healthy clockspeed boost for those single-threaded instances then it will be go time.

 

[Kuzi]

I agree with the TDP, but the Magny Cours is two Istanbul dies, not really 690mm2.

I think looking at the graph between Magny Cours and Istanbul on that graph, the performance comparison is with similar clocks.

Two Istanbul dies in MCM right? It will still hit TDP/heat constraints even if it's two dies, instead of one large one. I mean if AMD can clock Magny Cours the same as Istanbul (was it 2.6GHz?), the TDP will be much higher, that's for sure.

 

[Kuzi]

They beefed up the integer ALUs quite handedly, that will help desktop customers nicely.

Now if they get turbo-boost working and they make it a healthy clockspeed boost for those single-threaded instances then it will be go time.

I'm wondering how much of an improvement in TDP or clocks would AMD will get from adding HiK/MG to their future processors. IIRC, ever since Intel started using HiK/MG they got a noticeable TDP advantage over AMD.

 

[JFAMD]

Two Istanbul dies in MCM right? It will still hit TDP/heat constraints even if it's two dies, instead of one large one. I mean if AMD can clock Magny Cours the same as Istanbul (was it 2.6GHz?), the TDP will be much higher, that's for sure.

Similar to Istabul dies, but different. New features in them.

 

[Idontcare]

Two Istanbul dies in MCM right? It will still hit TDP/heat constraints even if it's two dies, instead of one large one. I mean if AMD can clock Magny Cours the same as Istanbul (was it 2.6GHz?), the TDP will be much higher, that's for sure.

As JFAMD alluded too, I believe the new chips are called Lisbon (not sure if that is the CPU codename or the platform).

Magny-Cours is supposed to be based on Lisbon, as is Thuban. At least that is the popular lore at the moment, maybe JFAMD can clarify this.

I'm wondering how much of an improvement in TDP or clocks would AMD will get from adding HiK/MG to their future processors. IIRC, ever since Intel started using HiK/MG they got a noticeable TDP advantage over AMD.

They stand to gain a lot from HK/MG, and for a lot of different reasons.

Now Intel has locked up a lot of IP in their replacement-gate integration strategy, so it makes it all the more challenging to create unique IP and integration approach that is every bit as good (if not better) while not tanking the rest of the integration (thermal budgets, yields, etc).

There is a reason the industry waited absolutely as long as it could (collectively) before transitioning to HK/MG.

 

[piesquared]

I'm betting on big things coming out of AMD in the near future and their 32nm node is one of them. And I don't think their process engineers take a back seat to anybody (and they are now at GloFo). Most recently intel, or at least their marketing, were making claims how HiK/MG was such a breakthrough, and the only way forward from 65nm. And since AMD never had it, it was going to be such a huge liability......Move ahead a couple years and here we are with AMD's 45nm SOI using immersion (before intel I might add) without HiK/MG performing quite nicely, even getting withing 4 degrees of absolute zero and functioning. Historically they were the first to impliment copper interconnects and the fabs themselves employ APM, the best of it's kind in the industry.
But, it all comes down to numbers in the end! So we'll have to say what they are early next year.

 

[piesquared]

Two Istanbul dies in MCM right? It will still hit TDP/heat constraints even if it's two dies, instead of one large one. I mean if AMD can clock Magny Cours the same as Istanbul (was it 2.6GHz?), the TDP will be much higher, that's for sure.

I believe they are going to be Lisbon die's, not Istanbul. I wonder what the chances are of them getting new instructions like SSE4.1

 

[JFAMD]

Can't comment on instructions, partially because I don't know. A Magny Cours is composed of 2 Lisbon dies that are connected via HyperTransport links. This allows all interdie traffic between the 2 dies to stay on-package. So the scalability is as close to perfect linear scaling as you can get.

And Thuban is a desktop design, Lisbon is a server design. You will probably see different features between the two.

 

[cbn]

They beefed up the integer ALUs quite handedly, that will help desktop customers nicely.

Now if they get turbo-boost working and they make it a healthy clockspeed boost for those single-threaded instances then it will be go time.

I just hope this part is good enough to warrant investment for faster die shrink transitioning.

People say die shrinks do nothing for performance but in reality the smaller chips just seem flat out better.

 

[CTho9305]

Move ahead a couple years and here we are with AMD's 45nm SOI using immersion (before intel I might add)

idontcare, would you take it as a good or bad thing to be going to immersion earlier? I had the impression it was a trick you'd keep in your back pocket for as long as possible and that dry was preferable, even if it required double-exposure (but I'm not a process guy, and I may have been mislead by marketing-disguised-as-research).