Could AMD die shrink thuban and add two cores?

[Maximilian]

Is that even feasible or would that be too difficult/costly at this point?

I doubt they would do that but it wouldn't be a bad idea to keep stars going whitest they fix bulldozer in the meantime. A die shrunk 8 core thuban would likely beat bulldozer in most things if its 45nm 6 core counterpart is anything to go on.

 

[Idontcare]

Is it possible? Heard of Llano?

They already paid to shrink the Stars core to 32nm. And they invested R&D money to improve its IPC another 3-7% over that of the cores in 45nm Thuban.

Two 32nm Llano cores, complete with 2MB L2$ (same as a bulldozer module) are ~30mm^2, just slightly smaller than the size of one bulldozer module complete with 2MB L2$ (~31.5mm^2).

Basically, take Zambezi and replace each bulldozer module with two Llano cores...you get an 8-core llano (sans the iGPU) but with 8MB L2$ and 8MB L3$ (same as zambezi) in the same diesize.

Higher IPC and more FPU cores.

 

[Ancalagon44]

They dont even need that much cache - if I was them I would probably investigate maybe giving each core 256 or 512kb of L2 cache but keeping the 1mb per core of L3 cache. That could save them precious die space.

End result would be something much faster than Zambezi at every task, while also being cheaper to produce and consuming less power.

 

[alyarb]

Glad this became a thread. Everyone seemed to ask for this in the other threads, and it pretty much turns the Bulldozer theory on its head because they claimed more compute resources per area and turned out to be flat out wrong about it(and everything else).

Hopefully they can reverse their strategy before another node comes and goes.

 

[sequoia464]

Is it possible? Heard of Llano?

They already paid to shrink the Stars core to 32nm. And they invested R&D money to improve its IPC another 3-7% over that of the cores in 45nm Thuban.

Two 32nm Llano cores, complete with 2MB L2$ (same as a bulldozer module) are ~30mm^2, just slightly smaller than the size of one bulldozer module complete with 2MB L2$ (~31.5mm^2).

Basically, take Zambezi and replace each bulldozer module with two Llano cores...you get an 8-core llano (sans the iGPU) but with 8MB L2$ and 8MB L3$ (same as zambezi) in the same diesize.

Higher IPC and more FPU cores.

Just for academics, what sort of development time would something like that take?

 

[sangyup81]

I wouldn't increase the cache size since BD seems to get latency penalties from them

 

[Idontcare]

Taking the existing Llano, lopping off the GPU stuff and doubling the core logic (making 8 cores)...no less than 1 yr minimum. Realistically 18-24 months.

If they started today, by the time it got through validation and out the door it would not be relevant. If they started a year ago, say around the time someone decided Dirk had to go, then it could be out as early as Q1 2012.

 

[alyarb]

They *had* to know at least a year ago that Bulldozer was going to be weaker per area per watt than Thuban, and adjusted right then and there.

 

[Ancalagon44]

What amazes me is that Bulldozer got out the door. I mean, if performance has improved with recent steppings, the initial steppings must have been terrible. Why did no one see this was Pentium 4 all over again and put a stop to it? There's a good book in there somewhere "The Failure of AMD's Bulldozer - Why We Didn't Learn From Prescott."

 

[StrangerGuy]

Glad this became a thread. Everyone seemed to ask for this in the other threads, and it pretty much turns the Bulldozer theory on its head because they claimed more compute resources per area and turned out to be flat out wrong about it(and everything else).

Hopefully they can reverse their strategy before another node comes and goes.

A 3.6GHz 6 core Llano with L3 and AVX would have owned, instead AMD chose the craptastic BD design with -20% IPC when they can make Llano with +6% IPC over Deneb. Baffling, really.

 

[Idontcare]

They *had* to know at least a year ago that Bulldozer was going to be weaker per area per watt than Thuban, and adjusted right then and there.

I would think so too. Assuming first silicon was not horribly crippled to the point of not being able to validate instruction latencies and boot a code analyzer, they would have known immediately if something was critically/fundamentally wrong with bulldozer.

That said, we can't discount project momentum, group psychology, denial, and the universal trait that humans share called hope.

Who knows how long they languished, paralyzed like a deer caught in the headlights, before someone pulled together a Plan B. They may have simply not.

Recall we all had similar discussions about Phenom versus an MCM Athlon X2, wondering why they delayed and delayed Phenom, hyped up the "native" quad-core design crap (while knowing it was crap), and came out with a chip that couldn't hit the needed clockspeeds.

Plan B then would have been an MCM'ed X2 to make a quad. But they didn't do it.

So who knows if they have any plans for the Llano cores beyond what we've already seen. History says not to count on it.

 

[Vesku]

Thing is they need something with future instructions and accelerators like bulldozer for servers, not sure why they plugged it into the enthusiast space. For the high end consumer CPUs though, would it really take a year+ to make a GPU-less 4 core Llano with 2MB L2 per core and 6-8MB of FAST L3? They could then put out a 6 core a bit later ala Thuban.

 

[jvroig]

Thing is they need something with future instructions and accelerators like bulldozer for servers, not sure why they plugged it into the enthusiast space. For the high end consumer CPUs though, would it really take a year+ to make a GPU-less 4 core Llano with 2MB L2 per core and 6-8MB of FAST L3? They could then put out a 6 core a bit later ala Thuban.

I think the reason is that they don't really care about the enthusiast space. They willingly screwed over the enthusiast market (what % is this? Compared to low-end and mainstream covered by Ontario, Zacate, and Llano, probably small enough to disregard) for a grab at the server market, thereby covering what they recognized as the most profitable markets they can handle given their limited resources: low-end and mainstream consumer parts, plus server parts.

Whatever cannot ship as Opterons, they just badge as Zambezi/FX products and sell off in the enthusiast space to reduce wastage, but it does not seem like a real effort (aside from marketing) to answer the need of the enthusiast / high-end space.

I've said this months before, but it seems very apt now that Bulldozer has proven itself ineffective in the high-end / enthusiast space: AMD is better-off focusing on server and higher-volume parts than trying to go head-to-head with Intel in the high-end market. It's a far better business decision, enthusiast market be damned, if it only contributes the smallest to the bottomline compared to the other markets available.

 

[Vesku]

That may indeed be it jvroig, sell off all the chips that are too leaky for server as "enthusiast" chips. Those guys all have big power supplies, amiright?

 

[alyarb]

If they are dumb enough to buy an 8-core CPU for their 4-threaded game, then they will be dumb enough to get the 700w Sunbeam PSU for $39.

 

[Rhoxed]

If we compare to the original Phenom launch, we will see a slightly better SKU in the coming months. And if any indication it looks like 18 months plus until we see a better revision (going by PHII timeline)

Looks like I will be going back to Intel if AMD can't pull something off. But I feel I can last a good 18 months + on my current system, so time will tell.

Originally I felt I should have waited until BD to upgrade when I bought my x6. Though looking back I am glad I bought it close to release, and got a good chip.

 

[IntelUser2000]

Just for academics, what sort of development time would something like that take?

Pentium D took 9 months to develop the core, and it likely took another 12 months for validation and get it to market.

On Llano, the mobile chips seem pretty power efficient(but requiring low clock speeds to do it), but the desktops aren't, and in both chips, CPU is making quite a difference in power consumption.

3.6GHz with 8 cores and reasonable power efficiency? I don't think that would have been possible in the timeframe they had.

 

[gevorg]

I wonder if its too late for AMD to change the next iteration of Bulldozer, the Pilerdriver. Maybe reduce the cache and increase core size, optimize/change architecture a bit, and do whatever it takes to shrink the gap with Intel. With Windows 8 out, it might be a little better for Piledriver. I hope AMD's R&D won't spend the whole 2012 making the same mistakes over and over. 10% gain by Piledriver is not enough, won't even catch up with Sandy Bridge, let alone Ivy Bridge (a tick+).

 

[AtenRa]

I think the reason is that they don't really care about the enthusiast space. They willingly screwed over the enthusiast market (what % is this? Compared to low-end and mainstream covered by Ontario, Zacate, and Llano, probably small enough to disregard) for a grab at the server market, thereby covering what they recognized as the most profitable markets they can handle given their limited resources: low-end and mainstream consumer parts, plus server parts.

Wrong, they do care about enthusiasts markets

Whatever cannot ship as Opterons, they just badge as Zambezi/FX products and sell off in the enthusiast space to reduce wastage, but it does not seem like a real effort (aside from marketing) to answer the need of the enthusiast / high-end space.

Im not 100% sure but i believe it's the opposite, Zambezi/FX are the highest binned dies due to higher operating frequency (3.6GHz/4.2 Turbo), what dies cannot be an FX CPUs goes to be an Opteron, due to lower frequency (2.3GHz).

I've said this months before, but it seems very apt now that Bulldozer has proven itself ineffective in the high-end / enthusiast space: AMD is better-off focusing on server and higher-volume parts than trying to go head-to-head with Intel in the high-end market. It's a far better business decision, enthusiast market be damned, if it only contributes the smallest to the bottomline compared to the other markets available.

As slide 4 shows, enthusiasts are the minority but they spend more and that translates to higher ASPs. Same for Server market, fewer server than desktops/laptops but higher ASPs = higher profit

 

[AtenRa]

First gen Bulldozer will not make a difference in Desktop, but it will start to make a big difference in Server market.

If AMD can raise Server(Bulldozer) and laptop(APUs) market share, they will make more profit and next gen BD will be even more competitive(Performance) in desktop too.

This is a new architecture, it will only get better in time

 

[jvroig]

@AtenRa

That's a marketing slide, yet again. The same types that hyped up how BD would perform, and we know that's just a half-hearted attempt.

What would certainly prove if enthusiasts are worth the effort or not compared to Servers, Zacate, Ontario and Llano would be a breakdown of profits from those different markets.

First gen Bulldozer will not make a difference in Desktop, but it will start to make a big difference in Server market.

As a server guy myself, this is what I've been waiting to see. Unfortunately, I have not seen such reviews yet.

 

[AtenRa]

@AtenRa
That's a marketing slide, yet again. The same types that hyped up how BD would perform, and we know that's just a half-hearted attempt.

The numbers come from Jon peddie research, no matter if it is inside a marketing slide

 

[jvroig]

The numbers come from Jon peddie research, no matter if it is inside a marketing slide

I don't think we're understanding each other.

It doesn't matter that it came from JPR. It could have come from Gartner, IDC, etc and it still wouldn't mean what you think it means.

AMD is using it as a marketing bullet point. It does not offer "proof" that the enthusiast market is important to their bottomline.

As I mentioned, what would be "proof" of the high-end/enthiusiast market's importance to their bottomline is a breakdown of their revenue, or a breakdown of the total, world-wide revenue per segment (low-end / mainstream / high / server) to see if that 16B actually matters or not. If the total market is 200B, and high-end is only 16B, then it is only 8%.

 

[AtenRa]

You said they (AMD) dont care about enthusiast space, i said they do care (lets agree at least for the Gaming market) because enthusiasts spend 41% of the total 16.6 Billion in hardware.

From that 16.6B AMD would very much like to grab the most they can buy selling CPUs, Motherboard Chipset, GPUs and more. So yes they care about enthusiasts and that is why they show gaming benchmarks in High resolutions in their marketing Slides

 

[Ancalagon44]

You said they (AMD) dont care about enthusiast space, i said they do care (lets agree at least for the Gaming market) because enthusiasts spend 41% of the total 16.6 Billion in hardware.

From that 16.6B AMD would very much like to grab the most they can buy selling CPUs, Motherboard Chipset, GPUs and more. So yes they care about enthusiasts and that is why they show gaming benchmarks in High resolutions in their marketing Slides

yeah but they made a CPU that is better suited to servers than to enthusiasts and gaming. They obviously dont care that much - if they did, it wouldnt have sucked as hard.