Future to Bulldozer architecture?

[William Gaatjes]

IMHO at the point of adding AVX2 to Jaguar, you may as well use the Zen core, perhaps minus the L3 cache. Zen is already very efficient, it should be more then able to scale down to 25W for 4C/8T with lower clocks.

Those are interesting numbers. What is the power consumption of zen during a prime95 test with 8 cores/16 threads at around 2.5 GHz ?
Can you test that ?


I doubt they will redesign the ccx.
Reoccurring theme with AMD is a 4 core cpu complex + cache.
I am sure that in the future consoles will have some Zen derivate.
But for now zen is too expensive to be added next to all other new features.
I am sure Microsoft and Sony did some comparison on how the whole system is balanced in computation when it comes to cpu, gpu and ram.


edit:

It seems a 4 core puma+ (jaguar) + gpu has a tdp of 15W with clocks of 2.5GHz.
I have no idea what the real power consumption is.

I think the newer jaguars are more puma+ and of course with all the HSA features.
http://www.anandtech.com/show/9246/amds-carrizo-l-apus-unveiled-12-25w-quad-core-puma

 

[Insert_Nickname]

Those are interesting numbers. What is the power consumption of zen during a prime95 test with 8 cores/16 threads at around 2.5 GHz ?
Can you test that ?

Best guesstimate? Somewhere between 35-45W. 1700non-X is using ~70W full package power at 3.2GHz, and 15W at 1550MHz idle. So somewhere in the middle seems reasonable. I haven't the time to thoroughly test presently unfortunately. I might get around to it sooner or later though.

I am sure that in the future consoles will have some Zen derivate.

Xbox Scorpio is confirmed to use "modified Jaguar" cores. That smells of Puma+ to me. But it is only conjecture.

 

[William Gaatjes]

Best guesstimate? Somewhere between 35-45W. 1700non-X is using ~70W full package power at 3.2GHz, and 15W at 1550MHz idle. So somewhere in the middle seems reasonable. I haven't the time to thoroughly test presently unfortunately. I might get around to it sooner or later though.

The 1700 seems more interesting every time i read about it.
I will give it a 6 more months to let all the AGESA modifications crystallize out into stable bios for at least 2933MHz RAM.

Xbox Scorpio is confirmed to use "modified Jaguar" cores. That smells of Puma+ to me. But it is only conjecture.

I kind of meant the consoles after scorpio and neo may be using a zen derivate.

But i too think modified puma + cores for scorpio. Strange thing is that carrizo-l which is puma+ based does not have HSA features.
So that may be where the modified jaguar expression comes from. puma+ and HSA support.

 

[coffeemonster]

I think the newer jaguars are more puma+ and of course with all the HSA features.
http://www.anandtech.com/show/9246/amds-carrizo-l-apus-unveiled-12-25w-quad-core-puma

actually none of the consumer puma+ apu's have HSA implementation. It's only been in the construction core APUs so far.

 

[William Gaatjes]

actually none of the consumer puma+ apu's have HSA implementation. It's only been in the construction core APUs so far.

That is true. I think maybe the new modified jaguar cores for scorpio are puma+ just with HSA support.
puma+ allows for easily higher clocks up to 2.5GHz and other easy enhancements on 28nm. Wild guess is that scorpio cores are puma+ based and not the old jaguar. And the new scorpio soc will be on 16nm or 14 nm if i am not mistaken.

I mean it is unlikely that the older jaguar can be made to fit to have HSA features, but the newer puma+ which is jaguar with tweaks for higher clocks at same TDP could not.

 

[darkswordsman17]

That is true. I think maybe the new modified jaguar cores for scorpio are puma+ just with HSA support.
puma+ allows for easily higher clocks up to 2.5GHz and other easy enhancements on 28nm. Wild guess is that scorpio cores are puma+ based and not the old jaguar. And the new scorpio soc will be on 16nm or 14 nm if i am not mistaken.

I mean it is unlikely that the older jaguar can be made to fit to have HSA features, but the newer puma+ which is jaguar with tweaks for higher clocks at same TDP could not.

Microsoft has already said that, they're using TSMCs 16nm.

It's difficult to tell what exactly they are. They're better than Jaguar, but how much, only the people with the actual chip design right now can say, and Microsoft is being somewhat vague. I doubt they've done too much tweaking (that would make a lot of difference in CPU power at least) since working to lessen CPU bottlenecks is a big reason why they put that DX12 command processor hardware in the GPU. But Microsoft also has to account for their OS/software setup, although perhaps that's also why it was necessitated, due to the virtual environment, they needed something that would keep the graphics command queue running smoothly because the virtual OS adds latency due to the abstraction of the software (which then kinda makes the point of changing graphics APIs to lessen abstraction a bit odd, Microsoft seems to be trading one abstraction for another, although I think their decision makes more sense, as developers will get better control of the hardware and if they screw something up it won't crash everything, and there's the usual security benefits to being able to sandbox things; I think it also could have big benefits for compatibility, where "compatibility mode" would be a full on virtual environment without having to saddle the main OS with all the extra legacy baggage).

 

[NostaSeronx]

Microsoft has already said that, they're using TSMCs 16nm.

Puma in Xbox One & S use body-biasing, and while Tiger in Xbox Scorpio uses back-biasing(Body biasing region in 16nm FinFET). So, 16h already support body/back biasing. Their is an issue though and it is that bulk nodes have very poor body/back effect potential. SSRW(DDC)/UTBB SOI have the highest potential for body-back potential.

Sony uses the standard architecture forms. Microsoft uses the enhanced models built semi-internally.

Specific to Bulldozer, anything in 16h can be ported back to 15h. (Same applies to 17h).

 

[Flash831]

Puma in Xbox One & S use body-biasing, and while Tiger in Xbox Scorpio uses back-biasing(Body biasing region in 16nm FinFET). So, 16h already support body/back biasing. Their is an issue though and it is that bulk nodes have very poor body/back effect potential. SSRW(DDC)/UTBB SOI have the highest potential for body-back potential.

Sony uses the standard architecture forms. Microsoft uses the enhanced models built semi-internally.

Specific to Bulldozer, anything in 16h can be ported back to 15h. (Same applies to 17h).

I don't really follow. According to you, what effect will above changes do to the Puma core in Xbox Scorpio versus the ones in Sony PS4 Pro?

 

[coffeemonster]

That is true. I think maybe the new modified jaguar cores for scorpio are puma+ just with HSA support.
puma+ allows for easily higher clocks up to 2.5GHz and other easy enhancements on 28nm. Wild guess is that scorpio cores are puma+ based and not the old jaguar. And the new scorpio soc will be on 16nm or 14 nm if i am not mistaken.

I mean it is unlikely that the older jaguar can be made to fit to have HSA features, but the newer puma+ which is jaguar with tweaks for higher clocks at same TDP could not.

Yeah I would think they would be using puma cores even in the xbox one S but I do not know for sure. from what I've read about Puma, it's basically Jaguar with turbo core and more advanced power management, which seems like exactly what they'd want in the slim model.

I wonder how amazingly efficient a puma soc shrunk to 16nm in a consumer laptop would have been.

 

[William Gaatjes]

Yeah I would think they would be using puma cores even in the xbox one S but I do not know for sure. from what I've read about Puma, it's basically Jaguar with turbo core and more advanced power management, which seems like exactly what they'd want in the slim model.

I wonder how amazingly efficient a puma soc shrunk to 16nm in a consumer laptop would have been.

I am too very curious.

 

[amd6502]

I am too very curious.

Those who have a xbox 1s (and compiler if such a thing is available) might be able to estimate the wattage: subtract system idle wattage from system wattage with fully loaded CPUs (8) and idle GPU.

 

[amd6502]

William Gaatjes said: ↑
Those are interesting numbers. What is the power consumption of zen during a prime95 test with 8 cores/16 threads at around 2.5 GHz ?

Best guesstimate? Somewhere between 35-45W. 1700non-X is using ~70W full package power at 3.2GHz, and 15W at 1550MHz idle. So somewhere in the middle seems reasonable. I haven't the time to thoroughly test presently unfortunately. I might get around to it sooner or later though.

Xbox Scorpio is confirmed to use "modified Jaguar" cores. That smells of Puma+ to me. But it is only conjecture.

It seems Zen is incredibly efficient slightly under 2.5ghz and "efficient AF" at 2GHz:
If this reddit poster (below) is accurate dual core sub-10watt Zen mobile will clock surprisingly high ~2ghz:

https://www.reddit.com/r/Amd/comments/698fwc/did_a_bit_of_math_on_ryzen_power_consumption/

Perf/Watt on Puma on equivalent manufacturing could actually be inferior? Maybe the only selling point of Puma would be higher performance/transistors or low cost to manufacture due to its low area per unit of performance.

 

[NostaSeronx]

Intel might be having a MCMT core based upon Bulldozer. It should be OoO and Big. It's development coincides with Skylake and Cannonlake. With a successor currently in progress, so we should start seeing that core soon-ish. Intel might even throw it on 22FFL to minimize risk, before slotting it post-Cannonlake.

SNC-Gen1 22FFL?
WLC-Gen2 14FFL?

 

[VirtualLarry]

Intel might be having a MCMT core based upon Bulldozer.

Of all of your posts that I have read, and thought that they were rubbish, this post has to take the cake as far as the MOST rubbish post I have seen yet.

Intel? Bulldozer? Excuse me while I laugh out loud.

 

[Shivansps]

The only future BD should have is a cold grave, just look forward and dont look back.

 

[Azuma Hazuki]

@NostaSeronx ...how would that work? I mean how would Intel even legally be allowed to just lift the BD floorplan wholesale, and why would they move from their Core lineage onto a design that was misbegotten from the start?

I can theoretically think of ways to improve it massively, and Intel's node superiority might make them feasible, but even that wouldn't be better than the current core lineup. If they doubled the FPU width, massively increased L1 and L2 caches, chopped a hell of a lot of latency off of them, and paired the cores with a decent IMC, they might have an interesting synthesizable product for the low-end market, but...no, I can't even see how this would be worth the effort.

Orochi is dead. It was never properly implemented, and it had some good ideas, and it produced my favorite chips ever before Zen (8320e and A8-7600), but...it's dead. Let it rest. It fought a fight it was never able to and did yeoman's work.

 

[Insert_Nickname]

If that's accurate, it is seriously impressive. 7W for a Zen dual @ 2.1GHz with full AVX2? Even if you add a few W for graphics and I/O, Atom-based cores are finished.

 

[William Gaatjes]

It seems Zen is incredibly efficient slightly under 2.5ghz and "efficient AF" at 2GHz:
If this reddit poster (below) is accurate dual core sub-10watt Zen mobile will clock surprisingly high ~2ghz:

https://www.reddit.com/r/Amd/comments/698fwc/did_a_bit_of_math_on_ryzen_power_consumption/

Perf/Watt on Puma on equivalent manufacturing could actually be inferior? Maybe the only selling point of Puma would be higher performance/transistors or low cost to manufacture due to its low area per unit of performance.

If correct, very interesting. Since the frequency goes down, the voltage can go down as well and that is were the real power savings come from. Power consumption and dissipation formula increases with quadratic voltage. And static power consumption goes down as well.

 

[NostaSeronx]

@NostaSeronx ...how would that work? I mean how would Intel even legally be allowed to just lift the BD floorplan wholesale, and why would they move from their Core lineage onto a design that was misbegotten from the start?

I see it more in replacing Atom cores.

Clustered Multithreading will always be better than pure CMP(2+) and extreme SMT(SMT2+(2x)).

Eight Atom cores require eight power controllers.
Two CMT modules with eight cores combined require only two power controllers.

Two SMT cores with four threads have only resources of two cores.
Two CMT modules with four cores have resources of four+ cores. While consuming less area if SMT was moved into a quad-core.

AMD & Intel are the only two companies that have originator patents/IP for the modern CMT core. <== Spans 1998 to 2005, before freezing then everything going ham.

It has been researched as more processors, CMT's advantage starts to become negliable.

Eight SMT cores is only trumped by eight CMT modules with a negligible 10%-20% EPI improvement.
While extreme CMT modules can replace a single SMT8 core with a massive 50% EPI improvement. Basically, for the same IPC the CMT module consumed half the power of the SMT core. Since, Atom used hyperthreading and the newer revisions aren't as power efficient. It only makes sense to push a new core that supersedes Atom, for less complexity, lower cost, and longer marketability.

 

[amd6502]

Dozers were not that power efficient. I think SMT wins in efficiency and single thread. Drawbacks of SMT are design complexity and possibly area. Dozer was not that great in area though thanks to the long number of stages.

I think area and cheapness is still something Puma could beat Zen at. Lower performance at lower power. Sometimes you need a golf cart rather than an automobile or racecar.

Maybe a modernized Puma could beat Zen in the under 5W SoC. If they can get 40% higher freq at the same wattage it'd be close: 1.7GHz Puma 4c/4t vs ???GHz Zen 2c/4t .

http://www.cpu-world.com/CPUs/Puma/AMD-A10-Series A10 Micro-6700T.html

I still think the next atom/cat replacements should be a hybrid complex of cores, involving puma+excavator or puma+zen.

 

[Abwx]

Power consumption and dissipation formula increases with quadratic voltage. And static power consumption goes down as well.

Here the total rate is a 2.53 exponent polynomial, this was easily deducible using the frequency/voltage curve posted by The Stilt in his technical thread.

 

[William Gaatjes]

Here the total rate is a 2.53 exponent polynomial, this was easily deducible using the frequency/voltage curve posted by The Stilt in his technical thread.

Math is not my strongest point...
I only have an Integer alu.

 

[NTMBK]

I see it more in replacing Atom cores.

Clustered Multithreading will always be better than pure CMP(2+) and extreme SMT(SMT2+(2x)).

Eight Atom cores require eight power controllers.
Two CMT modules with eight cores combined require only two power controllers.

Two SMT cores with four threads have only resources of two cores.
Two CMT modules with four cores have resources of four+ cores. While consuming less area if SMT was moved into a quad-core.

AMD & Intel are the only two companies that have originator patents/IP for the modern CMT core. <== Spans 1998 to 2005, before freezing then everything going ham.

It has been researched as more processors, CMT's advantage starts to become negliable.

Eight SMT cores is only trumped by eight CMT modules with a negligible 10%-20% EPI improvement.
While extreme CMT modules can replace a single SMT8 core with a massive 50% EPI improvement. Basically, for the same IPC the CMT module consumed half the power of the SMT core. Since, Atom used hyperthreading and the newer revisions aren't as power efficient. It only makes sense to push a new core that supersedes Atom, for less complexity, lower cost, and longer marketability.

To be honest I think small SMP cores clustered at the L2 level do the job fine. Both AMD and Intel have used this (Intel it on Knights Landing with 2 core tiles, AMD with Jaguar with 4 core tiles). It reduces the number of nodes in the on-die network, simplifying things nicely. Tighter integration between the cores just seems to add a lot of complexity.

 

[Abwx]

Math is not my strongest point...
I only have an Integer alu.

This means that if frequency is increased by 1.5x then power will increase by 1.5^(2.53) = 2.79x, and reciprocaly when frequency is reduced.

That s for frequencies below 3.3GHz, at wich point said exponent is increased because more voltage/Hz become necessary.

 

[William Gaatjes]

This means that if frequency is increased by 1.5x then power will increase by 1.5^(2.53) = 2.79x, and reciprocaly when frequency is reduced.

That s for frequencies below 3.3GHz, at wich point said exponent is increased because more voltage/Hz become necessary.

I see. That looks easy. Thank you.
I know standard English, but i could really need an update what all the math expressions mean.
When you type it like that it makes immediately sense.
I really need to put some effort in remembering translations of more (to me atleast) uncommon words.