6 or 8 core Steamroller based AMD CPU likely?

[Fjodor2001]

Hi,

I just wonder if you think it's likely that we'll see a 6 or 8 core Steamroller based CPU from AMD? Or in AMD terminology 3 Modules / 6 Threads, or 4 Modules / 8 Threads if you prefer that.

I mean if they already have spent the R&D money to develop the Steamroller uarch CPU core, making a 6/8 core chip based on that should not require that much more work, right?

Also, the current 8 core FX-8350 is one of the AMD CPUs that is selling the most. So is there any reason to not release a similar Steamroller based version?

 

[ShintaiDK]

AMD shows 2M/4T top on their roadmaps for SR. And AM3+ platform is dead.

The FX series is AMDs least selling CPU. And the FX8xxx series doesnt even sell close to a million a year.

 

[mrmt]

Also, the current 8 core FX-8350 is one of the AMD CPUs that is selling the most. So is there any reason to not release a similar Steamroller based version?

Source? Because the last info that got public is from one year ago and 8C FX was selling like 100.000 per quarter, servers included. In other words, peanuts, even for AMD.

 

[nismotigerwvu]

Source? Because the last info that got public is from one year ago and 8C FX was selling like 100.000 per quarter, servers included. In other words, peanuts, even for AMD.

Don't feed the trolls. You know he doesn't have a source and he's presenting a classic strawman. The high-end has always been a small market for every provider and will continue to be. The specific numbers are nearly meaningless and the bulk of sales are in the entry level to mid range offerings. Back on topic, there hasn't been any confirmation or denial of 6-8 core FX chips from AMD.

 

[Fjodor2001]

Source? Because the last info that got public is from one year ago and 8C FX was selling like 100.000 per quarter, servers included. In other words, peanuts, even for AMD.

Well, sorry I might be mistaken about that. I know it's one of the AMD CPUs that is usually among the most popular at consumer sites ranking CPUs, so I assumed it would sell pretty well for being a mainstream higher end AMD CPU. If it's not and you've got some source to prove that, then I stand corrected on that point.

 

[Eeqmcsq]

Basically, we don't know. But the signs so far have pointed to a 3M/6T or a 4M/8T as being NOT likely, that the FX-83xx isn't being replaced by an 8-core Steamroller, and that AMD will concentrate on APUs instead.

 

[Centauri]

It'll be interesting if AMD does back off of the 8T CPUs because it'll mean the 83xx chips will reign supreme over future AMD CPUs in multi-threaded work for a good while longer.

 

[AtenRa]

Source? Because the last info that got public is from one year ago and 8C FX was selling like 100.000 per quarter, servers included. In other words, peanuts, even for AMD.

That was for the Bulldozer FX81xx, not the Vishera FX83xx.

Im not saying that the FX83xx is the highest selling AMD CPU though.

 

[Abwx]

They could easily do a MCM CPU using two Kaveri dies
with non functional or badly damaged GPUs , it would cost
almost nothing to implement.

 

[Exophase]

They could easily do a MCM CPU using two Kaveri dies
with non functional or badly damaged GPUs , it would cost
almost nothing to implement.

That would mean Kaveri would have to be SMP capable which is probably not the case.

Actually it's even worse than that. Both dies would have to keep using their own memory controllers, and the socket wouldn't support a dual controller setup. Maybe if the dies let you disable a memory channel each it could happen, but I'm pretty skeptical about this.

 

[Yuriman]

That would mean Kaveri would have to be SMP capable which is probably not the case.

Actually it's even worse than that. Both dies would have to keep using their own memory controllers, and the socket wouldn't support a dual controller setup. Maybe if the dies let you disable a memory channel each it could happen, but I'm pretty skeptical about this.

And then you go back to the context-switching penalties we had in early "duals" and "quads" that didn't share cache.

 

[isawthedevil]

Hi,

I just wonder if you think it's likely that we'll see a 6 or 8 core Steamroller based CPU from AMD? Or in AMD terminology 3 Modules / 6 Threads, or 4 Modules / 8 Threads if you prefer that.

I mean if they already have spent the R&D money to develop the Steamroller uarch CPU core, making a 6/8 core chip based on that should not require that much more work, right?

Also, the current 8 core FX-8350 is one of the AMD CPUs that is selling the most. So is there any reason to not release a similar Steamroller based version?

http://www.youtube.com/watch?v=6aOhPncuj4g

 

[mrmt]

Well, sorry I might be mistaken about that. I know it's one of the AMD CPUs that is usually among the most popular at consumer sites ranking CPUs, so I assumed it would sell pretty well for being a mainstream higher end AMD CPU. If it's not and you've got some source to prove that, then I stand corrected on that point.

You know, assumptions are the mothers of all f.... you know the drill.

Check this old thread from this forum:

http://forums.anandtech.com/showthread.php?t=2284188

This was before the Q412 crash, so of anything the current numbers for derpdriver should be even worse than these.

 

[Fjodor2001]

So what would have to be done to go from a completed Steamroller and Kaveri design to a corresponding 8 core CPU version simlar to the FX8350? What parts could be reused, and what would have to be redesigned and/or added? E.g. would a new memory controller, socket, or similar be needed?

 

[escrow4]

AMD should dump AM3+, just take it out back and shoot it, then exclusively focus on FM2+ and its APU's. I'd forget the whole FX performance angle. AMD's time is past. Catching up and providing a half decent APU solution with the few resources AMD currently has would be the ticket, not trying to compete with something that would smash an i5 or i7 which I seriously doubt will happen anymore.

 

[bononos]

AMD should dump AM3+, just take it out back and shoot it, then exclusively focus on FM2+ and its APU's. I'd forget the whole FX performance angle. AMD's time is past. Catching up and providing a half decent APU solution with the few resources AMD currently has would be the ticket, not trying to compete with something that would smash an i5 or i7 which I seriously doubt will happen anymore.

Thats probably what they are doing despite keeping steamroller on the glossy slides. After slashing headcount, its doubtful that they are going to churn out a minor miracle.

 

[Exophase]

And then you go back to the context-switching penalties we had in early "duals" and "quads" that didn't share cache.

AMD's big core APUs already don't share any cache between modules..

Even if they could cobble together APUs with defective parts that'd just mean doing that instead of selling them as lower capability parts individually, which there's probably much higher demand for.

 

[AtenRa]

You know, assumptions are the mothers of all f.... you know the drill.

This was before the Q412 crash, so of anything the current numbers for derpdriver should be even worse than these.

I thought you said assumptions are the mothers of all f.....

 

[NTMBK]

They could easily do a MCM CPU using two Kaveri dies
with non functional or badly damaged GPUs , it would cost
almost nothing to implement.

Not possible. The old dual-die Intel parts were only possible because they could share a memory controller in the Northbridge via the FSB- but Kaveri has an integrated memory controller. Joining multiple chips with an IMC requires NUMA capabilities, like Hypertransport or QMI- and FM2(+) chips have no support for that. The AM3+ platform does, which is why you could get MCM Opterons based on the chips.

 

[guskline]

Basically, we don't know. But the signs so far have pointed to a 3M/6T or a 4M/8T as being NOT likely, that the FX-83xx isn't being replaced by an 8-core Steamroller, and that AMD will concentrate on APUs instead.

Well said. Like every other owner of a AMD3+ mb (Asus Sabertooth 990FX) I wish and hope and pray that the Vishera is not the last generation for this socket BUT I agree that the likelihood is that it IS the end.

For Fjodor2001, I offer this song to keep your hopes alive:
http://www.youtube.com/watch?v=ycbgHM1mI0k

 

[Abwx]

Not possible. The old dual-die Intel parts were only possible because they could share a memory controller in the Northbridge via the FSB- but Kaveri has an integrated memory controller. Joining multiple chips with an IMC requires NUMA capabilities, like Hypertransport or QMI- and FM2(+) chips have no support for that. The AM3+ platform does, which is why you could get MCM Opterons based on the chips.

They can use half an IMC for each chip and then all that is needed
is for the individual chips to have two HT links , one being used for
cross CPUs routings , this is how a MCM opteron works actualy ,
the chips are internaly HT linked , they dont use the external links
other than for cross sockets communication in MS systems.

Of course this would still be for an FM2+ plateform.

 

[NTMBK]

They can use half an IMC for each chip and then all that is needed
is for the individual chips to have two HT links , one being used for
cross CPUs routings , this is how a MCM opteron works actualy ,
the chips are internaly HT linked , they dont use the external links
other than for cross sockets communication in MS systems.

Of course this would still be for an FM2+ plateform.

But FM2+ still doesn't use Hypertransport links, so...

 

[nenforcer]

AMD has been selling PCI-E 3.0 graphics cards for nearly 2 years now and was in fact first to market with these cards.

Yet their very own high end x86 chipset, the 990FX, doesn't support this interface and is functionally identical to the previous chipset, the 890 series which is now like 4 years old.

They still make server chipsets and there is a lot of money to still make piggy backing off of Intel as the only other x86 licensee in the world. (VIA doesn't count)

I'm hoping they have a 1090FX with DDR4 support late next year but that's probably a pipe dream. They only just recently added PCI-E 3.0 support to FM2+ chipsets with on-die PCI-E controller on Kaveri APU's. Chips that aren't even available at retail yet.

 

[ShintaiDK]

AMD has been selling PCI-E 3.0 graphics cards for nearly 2 years now and was in fact first to market with these cards.

Yet their very own high end x86 chipset, the 990FX, doesn't support this interface and is functionally identical to the previous chipset, the 890 series which is now like 4 years old.

They still make server chipsets and there is a lot of money to still make piggy backing off of Intel as the only other x86 licensee in the world. (VIA doesn't count)

I'm hoping they have a 1090FX with DDR4 support late next year but that's probably a pipe dream. They only just recently added PCI-E 3.0 support to FM2+ chipsets with on-die PCI-E controller on Kaveri APU's. Chips that aren't even available at retail yet.

They dont. Their serverline is Pilediver only. And 2M/4T on the 1 socket designs.

 

[bononos]

I thought you said assumptions are the mothers of all f.....

Ok, so what are the numbers(or what do you think they are) are for Piledriver after 3 more quarters?