Could AMD die shrink thuban and add two cores?

[SickBeast]

They say that you can't just do a straight die shrink of a CPU; it has to be re-designed in a lot of ways.

That said, I think AMD could bring an 8-core "thuban" to the 32nm process if they wanted to. You never know though; they might run into problems by going that route as well.

It probably wouldn't take too many engineering resources to effectively "fix" bulldozer *and* create this 8-core "thuban". It will be interesting to see what AMD does. I could see Bulldozer being an excellent server/workstation CPU if they can get the power consumption in check. The problem is that they made the pipeline too long and effectively created a souped up "hyperthreading" type of setup. The Pentium 4 proved that this type of arrangement creates a hot CPU that has horrible power consumption. Why on earth they chose to emulate that type of design approach is beyond me. Their CEO deserved to be fired. In fact, whoever made those decisions should be fired. I hope for their sake that they are marketing people and not engineers or designers.

 

[Jovec]

AMD does have smart individuals working for them, even if the entire BD team collectively misfired. We all suspect GF has some 32nm issues that should be resolved. None of us know if the BD design is flawed top-to-bottom or merely suffers from a few key issues that can be rather quickly addressed in a respin or Piledriver.

 

[SickBeast]

AMD does have smart individuals working for them, even if the entire BD team collectively misfired. We all suspect GF has some 32nm issues that should be resolved. None of us know if the BD design is flawed top-to-bottom or merely suffers from a few key issues that can be rather quickly addressed in a respin or Piledriver.

I think it's fair to say that by lengthening the pipeline and creating "modules", AMD greatly reduced the efficiency of Bulldozer. What they did was to effectively create a Pentium 4 with a bunch of cores.

Perhaps the marketing trolls at intel behind the P4 got fired and AMD hired them or something.

It's bad. Very bad. AMD is no longer the company that they once were.

 

[3DVagabond]

It's not worth more than that.

Hey if Amd released a 100$ Hex core they'ed corner the damn market.

Intel would still be able to compete on price no matter what AMD does. Neither of them want to earn a living though selling $100 chips as their top bin part.

 

[Jionix]

AMD does have smart individuals working for them, even if the entire BD team collectively misfired. We all suspect GF has some 32nm issues that should be resolved. None of us know if the BD design is flawed top-to-bottom or merely suffers from a few key issues that can be rather quickly addressed in a respin or Piledriver.

There is several "implementation" flaws that have/will be corrected with Piledriver, the biggest being the cache latency, which will provide a relatively big increase in performance by itself.

AMD will *NOT* go back to Thuban. AMD will *NOT* scramble and put together a bigger Llano. It's Bulldozer here on out.

A lot of people are simply being to reactionary to this whole release. Bulldozer sucks in the regard that it isn't a big step forward, or at least, isn't the big step forward that people were expecting.

Bulldozer isn't a bum of a design, it just ran into a lot of issues that will be sorted out. History unfortunately is repeating itself (like Phenom), but Bulldozer will be magnitudes better by this time next year (hell, it will be better early next year with Trinity). Still behind Intel, I would say as a certainty. But it won't be nearly as far behind.

 

[Jovec]

I think it's fair to say that by lengthening the pipeline and creating "modules",

Unless marketing is driving everything (moar speed!), the longer pipeline might have been needed for something else in the design. As a general rule, a longer pipeline means lower IPC which requires higher clocks for the same performance. So that indicates there might be some other reason, else why not just use a shorter pipeline to keep up the IPC? I admit I'm not that informed on the BD arch, but if the CMT design required a longer pipeline with it's inherently lower IPC which in turn required a mature 32nm process for high clocks to make up the performance loss, then AMD made a bad choice trying to launch a new arch on a new process.

As to the modules/CMT, AMD has little chance at beating Intel on IPC (based on recent history), and no chance beating Intel on the process node. They almost have to do something more drastic to compete, and the module/cmt/super-HT concept seems like a natural evolution of the CPU core.

 

[SickBeast]

I don't see the module as the natural evolution of the CPU core, and evidently Intel doesn't either.

For that matter, I don't think IBM even uses that type of approach with their supercomputers.

IMO it was a marketing decision created to make it look like the CPU had more "cores" than it really did. Essentially Bulldozer is a 4 core CPU with "enhanced hyperthreading".

 

[3DVagabond]

The modular concept is to be able to use them as building blocks in other products as well as expandability.

 

[Ajay]

Bulldozer isn't a bum of a design, it just ran into a lot of issues that will be sorted out. History unfortunately is repeating itself (like Phenom), but Bulldozer will be magnitudes better by this time next year (hell, it will be better early next year with Trinity). Still behind Intel, I would say as a certainty. But it won't be nearly as far behind.

'Magnitudes better'

AMD's own slides show BD getting 50% by 2014. Hopefully they'll do better, because 50% isn't nearly enough to become more competitive with Intel.

 

[SickBeast]

The modular concept is to be able to use them as building blocks in other products as well as expandability.

The modules have nothing to do with a modular concept. CPUs are already "modular" in nature. The modules are modular as well, but I fail to see their use aside from artificially inflating core count.

It's a marketing gimmick. Marketers had too much say in the creation of Bulldozer, and the end result was a failed product.

I'm all about my FPU power in my CPUs, as it has typically been the limiting factor in gaming performance since the days of the original Pentium. I don't want my FPUs being "shared" with anything.

 

[dakU7]

How long will it take for AMD to just shrink the Thuban to 32nm?

 

[Rifter]

Not only could they but they SHOULD have. They would have had more than enough time to do it since finding out that BD sucked but decided to screw us over with BD anyways i guess due to super bad management.

 

[Maximilian]

Maybe they could slap two of the llanos together and create an 8 core MCM without the graphics. That might be easier...

 

[Idontcare]

Not only could they but they SHOULD have. They would have had more than enough time to do it since finding out that BD sucked but decided to screw us over with BD anyways i guess due to super bad management.

Consider that if their entire lineup planned for the next 4-5 yrs is all based on the bulldozer CMT microarchitecture then they've got to get those zambezi systems seeded in the wild or they risk never gaining critical mass in terms of marketshare for the software eco-system to care to follow-through with delivering "AMD optimized" software.

Falling back to stars core would be disastrous long-term. They've got to keep moving forward.

 

[Ajay]

Consider that if their entire lineup planned for the next 4-5 yrs is all based on the bulldozer CMT microarchitecture then they've got to get those zambezi systems seeded in the wild or they risk never gaining critical mass in terms of marketshare for the software eco-system to care to follow-through with delivering "AMD optimized" software.

Yeah, I guess they will be stuck for 4-5 years, just because of server side market needs anyway. Aside from fixing the three apparently big problems (slow int units, low sustained decode rate and horrible caches) - I wonder if they can pull off a redesign in, say, three years.

Maybe instead of adding more cores, they could just turn the 8 CMT cores into 8 fully fledged cores - and get over their SMT issues to beef of each of those cores even more.

It sort of boggles the mind that if SMT is working for Sparc, Power and Intel HT processors - that AMD thought CMT would be a better solution (maybe it will turn out so for virtualized server applications - we don't know anything about that yet).

Actually, maybe they should just do the Above for TrinityII and then AMD would 'just' have to segment out the server side CPUs from the desktop CPUs. This adds more expense, but being stuck with CMT on the desktop could be more expensive in the long run and even cause AMD to lose the desktop market all together.

Eh, nutz!

 

[Idontcare]

It sort of boggles the mind that if SMT is working for Sparc, Power and Intel HT processors - that AMD thought CMT would be a better solution (maybe it will turn out so for virtualized server applications - we don't know anything about that yet).

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.

Rather than pay royalties, they opted to strike out for what they knew to be fertile IP space with unknown performance rewards. A roll of the dice.

Consider the fact that the reason dresdenboy was able to track AMD's CMT efforts so well and for so long was because of the IP-space they were mining and documenting with patent applications.

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)

 

[Rifter]

Consider that if their entire lineup planned for the next 4-5 yrs is all based on the bulldozer CMT microarchitecture then they've got to get those zambezi systems seeded in the wild or they risk never gaining critical mass in terms of marketshare for the software eco-system to care to follow-through with delivering "AMD optimized" software.

Falling back to stars core would be disastrous long-term. They've got to keep moving forward.

Right but what im saying is they should have went with what they had and worked and wasnt a total flop(shrinking thuban)

And scrapped BD completly and worked on starting a completly new CPU.

Obviously BD is a complete flop and fails on all major metrics(IPC, Perf/watt, Value/$)

Even if they "fix" some of those fails on the next gen and lets say they even do very well and get 10% better performance and 10% better power numbers as well. Which IMO would be a miracle for not even being a die shrink(look at bloomfield to gulftown/westmere not a huge gain even with with a die shrink so AMD pulling that off really would be a miracle but lets say they do it) they are still going to have a chip that still sucks when compared to Ivy and its super low TDP and high performance.

BD is not the answer AMD needs and they should have realized that and scrapped it while they still could. They had better decide to scrap BD before haswell or they are really in trouble.

They could obviously cut prices by $50-100 per SKU and that would put them in a very competative price/performance bracket even if they still get killed in perf/watt but im unsure given BD's collosal die size if that would allow them to make any money.

Edit spelling and clairity.

 

[3DVagabond]

How long will it take for AMD to just shrink the Thuban to 32nm?

I doubt they will ever do it. You don't "just shrink" a chip It's not like opening up photoshop and clicking resize (sic).

 

[dakU7]

I doubt they will ever do it. You don't "just shrink" a chip It's not like opening up photoshop and clicking resize (sic).

Theoretically.

 

[Rifter]

I doubt they will ever do it. You don't "just shrink" a chip It's not like opening up photoshop and clicking resize (sic).

That would actually be pretty cool

But really it is easier than designing a whole new Chip(BD) and i think we can all agree at least from a performance standpoint they would have been better off and spent alot less cash shrinking Thuban than going forward with BD.

 

[mosox]

Since they decided modules, a 4-core module would have been better for Bulldozer? Or worse?

 

[3DVagabond]

That would actually be pretty cool

But really it is easier than designing a whole new Chip(BD) and i think we can all agree at least from a performance standpoint they would have been better off and spent alot less cash shrinking Thuban than going forward with BD.

You are using 20/20 hindsight, though. AMD didn't have that benefit 5yrs ago when they started designing BD.

AMD needs something better than PhII to be able to compete with Intel. I'm pretty sure they thought that BD was their best option to do that. It hasn't worked out that way, but I really doubt they realized all along that shrinking PhII would yield better performance.

For the sake of competition all we can hope is that AMD can somehow get out of BD what they saw in it when they started designing it.

 

[Rifter]

For the sake of competition all we can hope is that AMD can somehow get out of BD what they saw in it when they started designing it.

I think pigs will fly before this happens but i really hope im wrong on that.

But even if i am right and BD never matures into what they had hoped there is still hope that someone will buy AMD.

 

[Blitzvogel]

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.

 

[Blitzvogel]

The modules have nothing to do with a modular concept. CPUs are already "modular" in nature. The modules are modular as well, but I fail to see their use aside from artificially inflating core count.

It's a marketing gimmick. Marketers had too much say in the creation of Bulldozer, and the end result was a failed product.

I'm all about my FPU power in my CPUs, as it has typically been the limiting factor in gaming performance since the days of the original Pentium. I don't want my FPUs being "shared" with anything.

*Sorry for double post*

But FPU performance isn't as important in server workloads, no? Even still, perhaps AMD could've adapted K10.5 with a 256 bit FPU per core, and just went with 8 cores? That would be some killer FPU performance, and knowing AMD, it would be cheap. Content creators could have a ball with such a chip.