Future to Bulldozer architecture?

[VirtualLarry]

That it's considerably faster at the same clock and scales better than Excavator. something something AMD lying to cover up these supposed facts. I dunno, what do you make of it? nonsense?
https://www.reddit.com/r/Amd/comments/61b947/asus_fx550iuwsfx_gaming_laptop_amd_fx_9830p_rx/dfdid5r/

Basically true, from what I've observed. My understanding, is that Jaguar (is that before, or after, Puma?) cores have near-Core2Quad IPC, only that they were hobbled by low clocks. Hence my comment, above, about getting a new 14nm APU based on Jaguar that would clock at 3-3.5 Ghz. I think that it would be great.

 

[NostaSeronx]

There was a guy on AMD reddit recently that come into a Bristol Ridge thread talking about how AMD should have kept developing jaguar. That it's considerably faster at the same clock and scales better than Excavator. something something AMD lying to cover up these supposed facts. I dunno, what do you make of it? nonsense?
https://www.reddit.com/r/Amd/comments/61b947/asus_fx550iuwsfx_gaming_laptop_amd_fx_9830p_rx/dfdid5r/

Uhhh.
Steamroller started to wreck Cat. Excavator assured the wrecking. BR/SR are the best implementations of that.

There is only this for overclock...
http://valid.x86.fr/pfpg2k
http://hwbot.org/submission/3319873_namegt_cpu_frequency_a12_9800_4798.88_mhz
http://hwbot.org/hardware/processor/a12_9800/

=====
Other things should be considered... the Steamroller we got wasn't actually Steamroller.
(There was a more significant reason for Steamroller-related dies to have one extra module: HEDT - 5 modules / nDT - 3 modules)

The focus was put on mobile-optimized cores that were significantly more cheap to "improve" ~180m budget to ~40m budget, etc:
FX-7500/7600 => FX-8800P -> FX-9800/FX-9830...
This in turn allowed for a higher return and bigger budget for Zen. While, Zen wasn't solely alone in the development basket. Cat-derived successors(which moved to Samsung via Mongoose) and Bulldozer-derived successors(has yet to pop-up in retail products) were still in the making. Sometime 2014-2016, the plan was to shuffle successor Cat teams into the successor Bulldozer development. So, the future of Bulldozer might outcome into it becoming a phat 8-core Cat-based module eventually.

 

[bakyt115]

how many mm2 will be die of Beema/Mullins in 14nm?

http://www.anandtech.com/show/7974/...hitecture-a10-micro-6700t-performance-preview

in 28nm estimated 107 mm2

in 14nm 4 times less ~ 26.75mm2?

 

[NTMBK]

how many mm2 will be die of Beema/Mullins in 14nm?

http://www.anandtech.com/show/7974/...hitecture-a10-micro-6700t-performance-preview

in 28nm estimated 107 mm2

in 14nm 4 times less ~ 26.75mm2?

The foundries' "14nm" isn't really that dense... it's more like "20nm plus FinFETs". Not to mention there is stuff like IO which doesn't scale with process.

 

[VirtualLarry]

If the 14nm hypothetical "cat" core quad-core APU would be that small, in silicon area, then why not make a line of 8-core "cat" core APUs. Could shake the market up, and move the lower-powered market forward as far as core counts go. Though the weakness of the "cat" cores, was in single-threaded speed, especially in 64-bit code.

 

[NTMBK]

If the 14nm hypothetical "cat" core quad-core APU would be that small, in silicon area, then why not make a line of 8-core "cat" core APUs. Could shake the market up, and move the lower-powered market forward as far as core counts go. Though the weakness of the "cat" cores, was in single-threaded speed, especially in 64-bit code.

Precisely. There's a reason why AMD replaced 4-core Cat APUs with 2-core Excavator APUs (Stoney Ridge)- it has slightly lower multithreaded performance but way higher single threaded performance.

 

[coffeemonster]

Uhhh.
Steamroller started to wreck Cat. Excavator assured the wrecking. BR/SR are the best implementations of that.

There is only this for overclock...

that's what I thought. The guy seemed to think die shrinking puma+ would allow 3Ghz+ and give the cat core the edge it needed to be relevant again

The focus was put on mobile-optimized cores that were significantly more cheap to "improve" ~180m budget to ~40m budget, etc:
FX-7500/7600 => FX-8800P -> FX-9800/FX-9830...
This in turn allowed for a higher return and bigger budget for Zen. While, Zen wasn't solely alone in the development basket. Cat-derived successors(which moved to Samsung via Mongoose) and Bulldozer-derived successors(has yet to pop-up in retail products) were still in the making. Sometime 2014-2016, the plan was to shuffle successor Cat teams into the successor Bulldozer development. So, the future of Bulldozer might outcome into it becoming a phat 8-core Cat-based module eventually.

wasn't AMD's plan to scale down to 1 core architecture for all segments going forward?

 

[Abwx]

that's what I thought. The guy seemed to think die shrinking puma+ would allow 3Ghz+ and give the cat core the edge it needed to be relevant again

Actually it would have clocked above 3GHz and that was on AMD s plans but GF never delivered the adequate node, and Intel contra revenues rendered the whole thing risky, so the project was cancelled despite being about finalized.

 

[Headfoot]

Maintaining 2 large x86 cores is not going to happen. It makes zero sense technically, economically, really in any way at all. Bulldozer is going to go join Netburst and Itanium at the bottom of the processor ocean, where it belongs.

Zen will appear everywhere. I doubt even Jaguar and the Cat line will survive into a new iteration, other than as licensable IP

 

[William Gaatjes]

A bulldozer based module and successors do have a lot less transistors in comparison to a zen module.
As Jaguar replacement it could work fine when running at clocks > 3.5GHz.
I would not be surprised if in the near future, jaguar would be replaced and a fully synthesizable modernized excavator module would appear.
8 modules soc with a capable on die or on mcm gpu could work very well for game consoles with optimized code.

 

[krumme]

No Soironx is (at best) on Mary Jane.
It makes zero sense.
Br is ended this year. It marks the death of the 15 usd oem apu.
7nm is comming in 2 years. There is plenty room for zen transistors no room for a bastard 8 usd apu in 2018.

 

[KompuKare]

Actually it would have clocked above 3GHz and that was on AMD s plans but GF never delivered the adequate node, and Intel contra revenues rendered the whole thing risky, so the project was cancelled despite being about finalized.

Of topic and probably needs it own thread, but during this ~4$billion period was there no big company affected by this contra revenue and willing to lodge a formal complaint about dumping?
Think I read that in the US it wouldn't be considered dumping (or at least illegal dumping) but surely in other jurisdictions?
Well, anyway one thing is for sure: all the smaller players must be breathing a sigh of relieve that contra revenue is over.

 

[scannall]

Maintaining 2 large x86 cores is not going to happen. It makes zero sense technically, economically, really in any way at all. Bulldozer is going to go join Netburst and Itanium at the bottom of the processor ocean, where it belongs.

Zen will appear everywhere. I doubt even Jaguar and the Cat line will survive into a new iteration, other than as licensable IP

I doubt the cat cores will be around in 3 years. 4 core, 8 thread custom Zen SOC's for the Playstation and Xbox will replace them.

 

[Azuma Hazuki]

"Zen" can mean "everything" in Japanese. I doubt they're going to do much of anything with the old Cat cores any longer...and, if they can get Zen down to single-digit TDP, with its almost Broadwell-level IPC, why would they?

What I want to see is what AMD can do in 10W with Zen Something like a 2-core/4-thread Ryzen R3 derivative tuned for low TDP maybe. Would kick seven shades of impure silicon out of Apollo Lake for sure.

 

[NostaSeronx]

Bulldozer-derived successors would largely be targeting ultra low power. Zen-derived architecture would be targeting high performance. Even though the core(&module) sizes would be similar, each architecture would be using different: nodes, macros, cells, etc. Primarily to achieve a niche region for those specific architectures.

BD => Extremely* low cost to produce... thus high profit margin.
ZN => Extremely* high cost to produce... thus low profit margin.
*Reflected to effort needed to produce each processor.

Then, there is thing in the room. Does the Bulldozer successor take what is in Zen and remove SMT and double certain resources cuz Clustered Multitheading? (Same IP means even lower cost to produce)
//
8-way L1i -> Dual-threaded Branch Prediction (Really, no change what so ever here... other than more L1i)
Two decodes -> Two micro-op caches
Two integer slices -> Increased width(FP128/128-bit load/store to FP256/256-bit load/store) FPU.
2x L1D/LSQ? etc (Way the architecture is built data communication between slices can be done... *cough*rSMT/CSMT*cough*)
Single interfaces L2 16-way(One huge 1 MB slice) or Dual interfaced L2 2x8-way(Two smaller 512 KB slices that act like one huge 1 MB slice)?
Instruction windows double what is in standard Zen to clustered Zen. (Which means effectively, some should be looking at 2x IPC or 2x EPI over standard Zen) // Only way to increase IPC in modern architecture is to increase the window size.

 

[IllogicalGlory]

Excavator Mk.2 on 14nm is coming in September. It was nice knowing ya, Zen!

Ninja'd

 

[SPBHM]

new developments to bulldozer would be as logical as Intel keeping Netburst after 2006...

 

[NostaSeronx]

Excavator Mk.2 on 14nm is coming in September. It was nice knowing ya, Zen!

Production: Q2 2017

HPA = Bristol/Stoney Node. (Why build on FinFET when FDSOI is there: Bulldozer SOI/Bulk but not FinFET)

 

[Flash831]

I assume as BD was originally developed for a SOI process, it might make a comeback when newer SOI processes are competitive with comparable FinFET process.

Would it be impossible to port Zen to use FDSOI?

 

[NostaSeronx]

Would it be impossible to port Zen to use FDSOI?

The easy aka cheap port solutions:
PDSOI => FDSOI w/o BB
Bulk => FDSOI w/ BB or w/o BB
FinFET => FDSOI w/o BB

 

[dark zero]

Bulldozer and Cat are dead and Intel Atom with VIA ISAIAH will follow them on the bottom of the sea.... The future is Intel Core, AMD x86 Zen/ARM K12 and VIA upcoming uarch.

 

[amd6502]

I wouldn't mind seeing some new AM1 CPUs on 14nm rather than 28nm; certainly, they should be able to clock higher within the same TDP envelope. TBH, AM1 wasn't a horrible platform - it had SATA6G and USB3.0. Only, the CPUs released for it clocked kind of low, and had around Core2 IPC, so they didn't have stellar ST benchmarks.

I would also like to see a high clocked 25W A9 Stoney Athlon X2 CPU breathe a little more life into AM1. This would offer AM1 users a drop in upgrade via such a CPU plus a budget card like RX460. I don't know if Stoney SoC is capable of this; it may still have the required DDR3 controller (like Carrizo). (Can anyone here make an educated guess on this???)

As far as the 2016-2018 roadmap, the Stoney refresh is puzzling (unless it is a dirt cheap project). Stoney is great for single and sparse thread but lacks energy efficient multithread (two threads in 2018 is near laughable). Software in the near future will be 4+ threads, so it'll take a context switching hit as well.

 

[waltchan]

Any love with the K10 architecture today? An excellent lower-cost alternative to Bulldozer that's not clustered into threads at all.

 

[DrMrLordX]

Any love with the K10 architecture today? An excellent lower-cost alternative to Bulldozer that's not clustered into threads at all.

No.

 

[NTMBK]

Any love with the K10 architecture today? An excellent lower-cost alternative to Bulldozer that's not clustered into threads at all.

No. The K10 derived Llano was clearly slower than the Piledriver based Trinity.