6 or 8 core Steamroller based AMD CPU likely?

[NaroonGTX]

So...disable CMT or get a performance penalty.

The performance penalty is only in effect when both cores in a module are at full load. Lightly-threaded/loaded tasks won't see any huge bottleneck or anything.

I'm not running a BD or PD chip right now, but the only things that max out cores for me right now (consistently for long periods of time) are PCSX2 emulation and video rendering. I'd never buy an octocore only to disable half my total threads for a perf boost in MT workloads. I'd just buy an i5 or lower-tier i7 at that point.

 

[PPB]

So is all companies because everyone are not willing to pay the full price when they want a different product, yet the production cost of creating an entire new product is too high. Its how capitalism works.

You come from the position that there should be an i5 to begin with.

HT functionality is intrinsic to the core design in intel's current uArch. You can't make a true 4c/4t part without redesigning the architecture as a whole.

The HT in your i5 was arbitrarily fused off, it's there and would work if Intel didnt chop it off in order to make even more money off the i7s.

Same die for both products, a 50% price premium for one of them.

Keeping spinning to economics only proves that some company's user base became very tame during these years.

 

[NostaSeronx]

In general, Steamroller-derived SKUs should go up to 24 cores.

Zero to ten being APU domain and twelve to twenty-four being CPU domain.

 

[monstercameron]

In general, Steamroller-derived SKUs should go up to 24 cores.

Zero to ten being APU domain and twelve to twenty-four being CPU domain.

seems too fantastical to be real...

 

[NaroonGTX]

Are you actually saying that you expect 10-core APU's to exist in the near future? Because that's pure lunacy.

 

[erunion]

You come from the position that there should be an i5 to begin with.

HT functionality is intrinsic to the core design in intel's current uArch. You can't make a true 4c/4t part without redesigning the architecture as a whole.

The HT in your i5 was arbitrarily fused off, it's there and would work if Intel didnt chop it off in order to make even more money off the i7s.

Same die for both products, a 50% price premium for one of them.

Keeping spinning to economics only proves that some company's user base became very tame during these years.

You're correct about how Intel disables chips, but incorrect about the economics.

Intel is taking a $300 chip, fusing off parts, then selling for it $220. The only one losing money here is Intel, they are literally destroying $80 of value from their product.

Its cheaper for them to disable functional high end chips than to design a unique mid range die.

The reason they do that is simple:volume. They can sell more $220 chips than they can $300 dollar chips. Margins are lower but revenue is higher. They need cheap products in addition to expensive products to maximize revenue to fund their fabs.

Its the same across all semiconductor companies. AMD sold functional quad core dice as x3, when demand for x3s out stripped the supply of harvested dice.

 

[Homeles]

You're correct about how Intel disables chips, but incorrect about the economics.

Intel is taking a $300 chip, fusing off parts, then selling for it $220. The only one losing money here is Intel, they are literally destroying $80 of value from their product.

Its cheaper for them to disable functional high end chips than to design a unique mid range die.

The reason they do that is simple:volume. They can sell more $220 chips than they can $300 dollar chips. Margins are lower but revenue is higher. They need cheap products in addition to expensive products to maximize revenue to fund their fans.

Its the same across all semiconductor companies. AMD sold functional quad core dice as x3, when demand for x3s out stripped the supply of harvested dice.

Bingo.

 

[NostaSeronx]

Are you actually saying that you expect 10-core APU's to exist in the near future? Because that's pure lunacy.

2.5+ years after Kaveri launches and should be on a fully depleted or a fully thin depleted process.

 

[Exophase]

In general, Steamroller-derived SKUs should go up to 24 cores.

Zero to ten being APU domain and twelve to twenty-four being CPU domain.

What will come first, the ten core APUs or the zero core ones?

 

[krumme]

What will come first, the ten core APUs or the zero core ones?

Lol.
That is true scalability!

 

[NaroonGTX]

World's first zero-core APU?

Sign me up!

 

[NostaSeronx]

What will come first, the ten core APUs or the zero core ones?

The 10 core APUs, since the zero cores ones are in development hell.

 

[Exophase]

Here's what I think:

a) AMD isn't going to start releasing multiple big core APU dies with different CPU core arrangements
b) AMD isn't going to spend substantially more die area on CPU than they currently do
c) AMD will in general favor more GPU area over more CPU area because it has more obvious returns

For a 10 core APU to even be on the table I think it'd have to be on GF (or TSMC?) 10nm, and even that's a stretch. I really doubt anyone will use FD-SOI in addition to FinFETs on that (much less the presence of an FD-SOI only process). H2 2016 seems too optimistic for 10nm, although I don't actually know their timeline - if the shrink work is happening in parallel with the FinFET stuff rolled out at 16/14nm then maybe. Although that assumes GF will have 20nm ready around when TSMC does which I'm highly doubting given how much longer it took 28nm (for Samsung too). Something at this point would surely at least be using Excavator, nothing you would consider Steamroller derived.

Me, I think AMD may bump it up to 6 cores for 20nm and 16/14nm and that's it.

 

[Fjodor2001]

Here's what I think:

a) AMD isn't going to start releasing multiple big core APU dies with different CPU core arrangements
b) AMD isn't going to spend substantially more die area on CPU than they currently do
c) AMD will in general favor more GPU area over more CPU area because it has more obvious returns

For a 10 core APU to even be on the table I think it'd have to be on GF (or TSMC?) 10nm, and even that's a stretch. I really doubt anyone will use FD-SOI in addition to FinFETs on that (much less the presence of an FD-SOI only process). H2 2016 seems too optimistic for 10nm, although I don't actually know their timeline - if the shrink work is happening in parallel with the FinFET stuff rolled out at 16/14nm then maybe. Although that assumes GF will have 20nm ready around when TSMC does which I'm highly doubting given how much longer it took 28nm (for Samsung too). Something at this point would surely at least be using Excavator, nothing you would consider Steamroller derived.

Me, I think AMD may bump it up to 6 cores for 20nm and 16/14nm and that's it.

But what if they make a Steamroller or Excavator CPU without iGPU? Then they'll have lots of die area to add additional CPU cores.

 

[red_222]

From my understanding the modules will carry out not only through Steamroller but Excavator too:
http://assets.vr-zone.net/17088/AMD_ArchiRoadmap.jpg

Its a great idea but I think I will hang onto my Phenom 1090T a little longer, plus I hear Intel will have a true 8 core coming out next year.

I have an A6 3650 Llano APU and often wondered why AMD didn't use the Llano cores (derived from the K10 architecture AKA the Phenoms) and add 4 more Llano cores to make a true 8 core??

 

[Exophase]

But what if they make a Steamroller or Excavator CPU without iGPU? Then they'll have lots of die area to add additional CPU cores.

In that case I think we could see 10 or maybe 12 cores - again, I don't think AMD will make multiple dies if they even make a non-APU at all. This would mean that they'd continue to prefer MCMing them together for higher core server parts. AFAIK there were supposed to be 10 core parts as early as the first Bulldozer generation but they got scrapped, that's kind of telling of how aggressive AMD wants to (or can) be here.

One thing to keep in mind is that you can't just keep adding as many cores as you want to dedicate die area to, the interconnects need to support it. Thuban topped out at a 6 core cross-bar. Here modules are basically playing the same roles to the interconnect as the cores were, so I think they'd have similar difficulty putting more than 6 or maybe 8 modules on a die. They could make the equivalent of the MCM configuration on a single monolithic die, but that means doubling the L3 controllers and memory controllers in the desktop space where it'd be a poor fit.

 

[NaroonGTX]

I have an A6 3650 Llano APU and often wondered why AMD didn't use the Llano cores (derived from the K10 architecture AKA the Phenoms) and add 4 more Llano cores to make a true 8 core??

Firstly because GloFo were doing horribly with the 32nm process for Llano, and were having bad yields. That's kinda why Llano wasn't such a great overclocker and didn't have impressive clocks, even though it was originally projected to clock higher than Deneb/Thuban.

Second because the K8 uarch (Agena, Deneb, Thuban, etc. were all still K8 family, K10 was actually Bulldozer and onward but it's become a misnomer that 10h = K10, so everyone just adopted that...) hit its performance wall and AMD knew it, which is why Bulldozer was rushed out in 2011. Llano was about 6~7% faster than PII clock-for-clock even without L3 cache, but that was about all they could squeeze out of an ancient uarch like that.

APU '13 starting later today. New roadmaps should be out either today or one of the following days, it'll clear up whether there will be a SR successor to Vishera (I doubt it) or not, amongst other things.

Edit: Alatar from OCN posted this in a SR thread: http://cdn.overclock.net/9/99/9933b803_1464715_10151756108733946_1811656005_n.jpeg

Dunno where he got it from. CPU-Z screen of a Kaveri ES.

 

[schmuckley]

Idk..let's see what this does 1st.. https://scontent-b-dfw.xx.fbcdn.net/hphotos-prn2/1471077_10151756127648946_1689011012_n.jpg

Also..this pretty much says it all: http://www.extremetech.com/computin...ation-rumor-claims-big-cores-still-a-priority

LOL..I can't say anything ;x

 

[NaroonGTX]

^What are you talking about?

 

[Erenhardt]

Also..this pretty much says it all: http://www.extremetech.com/computin...ation-rumor-claims-big-cores-still-a-priority

LOL..I can't say anything ;x

I doesn't say a thing!

hat said, Hughes comment only referred to Steamroller. It’s possible that AMD has killed Excavator in favor of an entirely different design. It could be planning to scale Jaguar’s (second-generation Brazos) CPU core upwards to fill a gap in the lower end of the desktop market that Trinity currently addresses.

SR is slower than scaled jaguar?
So, SR is maxed out and they abandon bulldozer arch to find something better?

 

[schmuckley]

^What are you talking about?

"According to AMD, this simply isn’t the case. We spoke with Phil Hughes, senior AMD spokesperson, who told us “Kaveri with the Steamroller core is still a product in development at AMD and we continue to develop big cores.”
Obviously, AMD has enormous reasons to want to control this type of leak and only announce a product cancelation on its own terms, but we’re inclined to believe the company in this case."


source: http://www.extremetech.com/computin...ation-rumor-claims-big-cores-still-a-priority



RIF

 

[NaroonGTX]

No, I asked what you were talking about because in your post, there wasn't any context. "This pretty much says it all" Says what about what?

If we're talking about whether or not SR or EX were canceled, it should be beyond obvious at this point that neither of them are canceled. Kaveri is on track for its launch and Carrizo -- the Kaveri follow-up using EX x86 cores and presumably a Pirate Islands-derived GPU -- is a Q1 2015 part.

SR is slower than scaled jaguar?
So, SR is maxed out and they abandon bulldozer arch to find something better?

I'm not sure where you got this from... There is TONS of room for improvement within the Bulldozer uArch. It is not 'maxed' out. Jaguar was never meant to scale higher than its intended segment: low-power devices, hence it using 'little cores'. It's the successor to the Bobcat parts.

AMD is in no position to just go "oop! time for a new arch!" and throw away all the R&D they've done on EX. The sheer amount of time, planning, and money that goes into a microarchitecture is insane. I'm pretty sure I recall public talks of a uarch design very much like what we got with Bulldozer way back in 2004. I'm sure that whatever succeeds Excavator will continue the modular design ethos.

 

[ajnorth22]

I've been wondering about 8-core Steamroller for a long time. My theory is that we won't see six or eight-core CPUs again until Excavator, and they'll be rebranded. Here's why:

1) Traditionally, server processors are on a longer replacement cycle and don't adhere to the annual cadence that's been used for consumer products. AMD is sticking with Vishera/Piledriver through 2014 on the server market. Given the paltry number of enthusiast many-core CPUs they sell, it makes little sense to design, produce, and market a product strictly for the consumer space that may only sell into the low hundreds of thousands of units in the upcoming year. Waiting until the next generation of products economizes production and promotion by allowing AMD to hit the enterprise and consumer markets with the same eight-core product at the same time.

2) AMD's current microarchitecture demonstrates very poor L3 cache performance. They spoke previously about identifying a fix but considered it a low priority until Excavator's release. Why? I believe it's because the Kaveri generation is all-APU and does not include L3 cache in any SKU. Mark Papermaster talked about this subject in 2012, IIRC.

3) AMD's current restructuring initiative is harsh but simple; the executive leadership has elected to focus exclusively on the products that can make AMD money right now, and justifiably so. That means focusing on inexpensive, lower-power APUs with one hand (Jaguar, Kaveri, and semi-custom APUs) and pushing an aggressive GPU strategy (Radeon rebranding and Hawaii) with the other. Until they're convinced that they have the ability to sell a capable CPU that will make money, FX and its derivatives will wither on the vine. Steamroller apparently is not the product for beating Intel and winning back market share, although it should definitely hold its own. Excavator might be what it takes to really win hearts and minds above the $150 price point.

4) Power consumption is still the ultimate shackle for AMD. Kaveri will include HSA-oriented optimizations and a better graphics microarchitecture but if you isolate the CPU side, Steamroller alone likely does not push the bar forward far enough to make it a truly effective alternative to a comparable Intel product.
I suspect that any power savings we'll see in the Kaveri design are due largely to GCN being more efficient than VLIW on the GPU side; very little in the way of added efficiency may come from the Steamroller cores alone. 28nm to 22nm bulk isn't exactly a huge improvement, and since AMD is beefing up the front end with more decode hardware and larger caches, they may be offsetting their power savings entirely. Excavator on the other hand is supposed to include the automated design tools AMD is leveraging from the GPU branch of the company, from which they expect to gain improvements equal to a full process node. This is another thing Papermaster spoke about previously.

---

I think there's a pretty compelling case for expecting AMD to just target the HSA/APU space for the upcoming year while reloading their other products for 2015. I hope I'm wrong and that we'll see an enthusiast-grade CPU from AMD for 2014, but I believe they're out of this market until they retrench, rebrand (to escape the FX stigma), and come back with a truly competitive product. If they're not 100% certain that Steamroller is up to the task, there's no reason for them not to stall until Excavator is ready to go.

Sorry for writing a novel.

 

[bgt]

I think there's a pretty compelling case for expecting AMD to just target the HSA/APU space for the upcoming year while reloading their other products for 2015. I hope I'm wrong and that we'll see an enthusiast-grade CPU from AMD for 2014, but I believe they're out of this market until they retrench, rebrand (to escape the FX stigma), and come back with a truly competitive product. If they're not 100% certain that Steamroller is up to the task, there's no reason for them not to stall until Excavator is ready to go.

Nice read:thumbsup:

 

[maddie]

But what if they make a Steamroller or Excavator CPU without iGPU? Then they'll have lots of die area to add additional CPU cores.

Looking at this page gives the impression that an 8 core steamroller even with added L3 cache would be about the same size as a Kaveri die.

http://www.anandtech.com/show/7507/amd-kaveri-apu-launch-details-desktop-january-14th