AMD CPU speculation... and expert conjecture

[Cazalan]

I have everything set right and i'm sorry typing to you right now i'm using 95 watts with cool and quiet on that means what juan has to say is some what right you know i'm no fan to him but that doesn't make some of his claims not correct

If you want low wattage get a laptop.

 

[-Fran-]

Which is irrelevant when discussion DX vs Mantle performance metrics.

It isn't. A bad game will carry out bad results for whatever you want to see around it. Well, it's not a "bad programmed game", but it's the "1st try" at MANTLE. Its bound to have quirks and poor decisions given the rush and/or beta state of the API. Also, consider they didn't do MP tests.

Take a look at Thief, which is a "2nd try" on MANTLE. It looks great IMO and shows the direction MANTLE is taking and should be taking.

We just need more games to actually get a more broad view on how it's evolving.

Cheers!

 

[Embra]

Which is irrelevant when discussion DX vs Mantle performance metrics.

It isn't. A bad game will carry out bad results for whatever you want to see around it. Well, it's not a "bad programmed game", but it's the "1st try" at MANTLE. Its bound to have quirks and poor decisions given the rush and/or beta state of the API. Also, consider they didn't do MP tests.

Take a look at Thief, which is a "2nd try" on MANTLE. It looks great IMO and shows the direction MANTLE is taking and should be taking.

We just need more games to actually get a more broad view on how it's evolving.

Cheers!

"Evolving" being the key word. We tend to shoot first and ask questions later too much.

 

[de5_Roy]

I have everything set right and i'm sorry typing to you right now i'm using 95 watts with cool and quiet on that means what juan has to say is some what right you know i'm no fan to him but that doesn't make some of his claims not correct

If you want low wattage get a laptop.

i am not arguing juan's claim, mind you. however, i have very strong doubt if juan has ever even tried to tune a pc. that's not the point anyway.
in idle mode, the cpu clocks down and undervolts itself. so the cores aren't using that much power. but the other devices still stay on and use power, such as storage drives, o.d.d., gfx card and networking devices. in os., background processes some times prevent low idle power use. that's why i said that you need tuning.

as for web browsing, i dunno about chrome, but i've seen firefox cause spikes in processor usage even in simple webpage refreshing. i think scripts have something to do with this.

laptop cpus are already clocked lower, has far less i/o ports (2-3 vs 10-14 in dt), only one storage drive and one o.d.d. i.e. less connected devices and low power parts result in even less power use. performance desktops have more power phases in the cpu power delivery which also may cause high idle/active idle power use as more phases are activated. i think that the amd fx platform typically idles higher. richland and kaveri should idle at lower power level.

 

[szatkus]

Have a nice Google Translating:

http://chinese.vr-zone.com/121197/no-ondie-stack-ram-amd-carrizo-soc-structure-shown-07162014/

 

[de5_Roy]

Have a nice Google Translating:

http://chinese.vr-zone.com/121197/no-ondie-stack-ram-amd-carrizo-soc-structure-shown-07162014/

fortunately, i know of a parrot site that rips off original leaks:
http://wccftech.com/amd-carrizo-apu-mobile-platform-details-unveiled-30-performance-uplift-15w-tdp-4-x86-excavator-512-gcn-cores-soc/


edit:
i don't know what to make of it. on one hand, it reinforces the previous rumor of mobile carrizo having fch on die. on another hand, it adds several points that make sense for mobile but not the apu itself. for example, i understand why 2MB L2 cache had to go - cache subsystem musta been improved and amd traded off cache area for other stuff. btw, is smaller L2 cache indication for on-package dram? also snipping off pcie gen 3 lanes musta helped with keeping power consumption down.

 

[szatkus]

Have a nice Google Translating:

http://chinese.vr-zone.com/121197/no-ondie-stack-ram-amd-carrizo-soc-structure-shown-07162014/

fortunately, i know of a parrot site that rips off original leaks:
http://wccftech.com/amd-carrizo-apu-mobile-platform-details-unveiled-30-performance-uplift-15w-tdp-4-x86-excavator-512-gcn-cores-soc/

Yeah, but as you can see they transformed few paragraphs to quite long article, so I usually prefer to check source. I don't trust WCCF's "expert analyze".

edit:
i don't know what to make of it. on one hand, it reinforces the previous rumor of mobile carrizo having fch on die. on another hand, it adds several points that make sense for mobile but not the apu itself. for example, i understand why 2MB L2 cache had to go - cache subsystem musta been improved and amd traded off cache area for other stuff. btw, is smaller L2 cache indication for on-package dram? also snipping off pcie gen 3 lanes musta helped with keeping power consumption down.

I don't know. But strange, that there are nothing about stacked DRAM on the diagram.

 

[Fidgetmaster]

No man, I tried everything with that Dang CPU....its a turd...and turn off ALL of that crap power saving junk....its just bad man....really getting crappy frames with that cpu....

 

[Fidgetmaster]

Mantle optimized for 280x/290x.... runs really bad on 270/270x though...sucks I have a 270....R9 Series I swear have had the most Hellish drivers of all time....

 

[gamerk316]

NVIDIA K1 benches, and not even from NVIDIA!

http://www.digitalstormonline.com/unlocked/nvidia-tegra-k1-benchmark-performance-review-idnum313/

 

[esrever]

I just want to see civilizations, all the games they have so far don't even have much CPU bottleneck. Civilizations actually requires more cpu that dx11 didn't seem to be able to allow.

 

[esrever]

NVIDIA K1 benches, and not even from NVIDIA!

http://www.digitalstormonline.com/unlocked/nvidia-tegra-k1-benchmark-performance-review-idnum313/

I can't take the graphs seriously with that digital storm thing on top of it.

 

[gamerk316]

I just want to see civilizations, all the games they have so far don't even have much CPU bottleneck. Civilizations actually requires more cpu that dx11 didn't seem to be able to allow.

Civs problem isn't GPU or driver side though; its the AI processing that kills performance. Think about it, you have 8+ Civs to processes, national/local production decisions, unit deployments (and re-processing based on the outcomes), and you have to do it serially. Yeah, you're going to be CPU bound. Mantle isn't going to help with this. The problem is algorithmic, not driver/API based.

 

[8350rocks]

Have a nice Google Translating:

http://chinese.vr-zone.com/121197/no-ondie-stack-ram-amd-carrizo-soc-structure-shown-07162014/

fortunately, i know of a parrot site that rips off original leaks:
http://wccftech.com/amd-carrizo-apu-mobile-platform-details-unveiled-30-performance-uplift-15w-tdp-4-x86-excavator-512-gcn-cores-soc/


edit:
i don't know what to make of it. on one hand, it reinforces the previous rumor of mobile carrizo having fch on die. on another hand, it adds several points that make sense for mobile but not the apu itself. for example, i understand why 2MB L2 cache had to go - cache subsystem musta been improved and amd traded off cache area for other stuff. btw, is smaller L2 cache indication for on-package dram? also snipping off pcie gen 3 lanes musta helped with keeping power consumption down.

You are asking some really interesting questions...the answers will become more clear soon enough...

"There were 2 divergent paths..."

Perhaps AMD is taking mobile far more seriously... I cannot comment...

 

[jdwii]

Really love Toms article on mantle. Hating me however those thief results should not be downplayed we are talking about a 4170(100$) performing as a high-end intel in gaming. If directx 12 has these same benefits then i have faith in the future. GTA 5 is really going to be nice with mantle i assume. With just a API we are seeing CPU's from the bottom move up to the top

 

[de5_Roy]

You are asking some really interesting questions...the answers will become more clear soon enough...

"There were 2 divergent paths..."

Perhaps AMD is taking mobile far more seriously... I cannot comment...

don't assume i've forgotten:, what's gonna be the dt apu's code name - prairie dog town fork red? blanka? cimarron? i'm getting strong vibes from the first one. it's got "red" in the name, geddit?. if it's not, then it's most definitely k12's x86 sister core's name.

amd taking mobile seriously - has been obvious since the a8 3500m apu came out. technically brazos came out first but llano was the proof that amd was serious about mobile pcs. now, carrizo seems like another turning point/crossroads moment. at least that what the rumors have pointed to so far. i mean, the first e series apus (brazos) had around 20w rating and 2 weak cpu cores and double digit shaders. later, richland couldn't even scale down to 15w. now this rumored carrizo is here, claiming to aim for 15-35w with integrated fch and full 4 cpu cores and 8CU igpu. it's amd's endgame of bd uarch. i wonder if glofo can even fab something like this considering how badly they seem to be doing with kaveri. others may disagree and may be right even, but i think glofo shoulda done better with their 28nm shp node.
as for the die, i think shaving the L2 array and may be by shrinking the suspiciously large memory controller area amd has opened up space for the fch.
http://media.bestofmicro.com/0/P/418633/original/kaveri.jpg
my concern is about the igpu and power limits. for full 8 cus and 15w tdp, i'm guessing the psp would be turned off.
kaveri has only two 17w skus - single module cpu and 1MB L2 cache.
so who is the 15w counterpart that the leaked slide is claiming 30% supremacy over? kabini or mullins socs? or that 35w carrizo is faster than 35w kaveri fx7600p? marketing shenanigans? what about that 12w number? is that sdp? how is amd power gating the fch? the unb and fch will stay on and use power as long as the pc will run, so how fine-tuned is the power management circuitry?

 

[8350rocks]

Dynamic configurable TDP is part of the equation. That is the only thing I can say.

 

[Cazalan]

my concern is about the igpu and power limits. for full 8 cus and 15w tdp, i'm guessing the psp would be turned off.

how is amd power gating the fch? the unb and fch will stay on and use power as long as the pc will run, so how fine-tuned is the power management circuitry?

There's no reason to believe the lowest TDP variant will have all the CUs enabled.

If they put the PCH on it's own power island and gate off the power it's basically zero power drain.

http://www.edn.com/design/integrated-circuit-design/4318458/Dual-threshold-voltages-and-power-gating-design-flows-offer-good-results

 

[con635]

Mantle optimized for 280x/290x.... runs really bad on 270/270x though...sucks I have a 270....R9 Series I swear have had the most Hellish drivers of all time....

Is it mantle/drivers though? tbh I thought it was bf4 that had the problem, vram memory leak, star swarm works perfect for me but my vram fills almost instantly on mantle bf4 then it stutters, I know I've only 1gb but dx uses just 400-600mb on the same settings others on the bf4 forum have exact same problem some only after the last 1gb patch. Thief and PvZ I've not much interest in so havent tried.

 

[szatkus]

Carrizo on 28 nm still isn't confirmed by AMD, right?

 

[colinp]

Carrizo doesn't currently feature in AMD's official roadmap, so in that sense, no.

 

[jdwii]

http://www.fool.com/investing/general/2014/07/16/will-amd-supply-chips-for-nintendos-next-handhel-2.aspx

Maybe another win for Amd maybe it will be some Arm+GNC 2.0 core that would be great

 

[szatkus]

http://www.fool.com/investing/general/2014/07/16/will-amd-supply-chips-for-nintendos-next-handhel-2.aspx

Maybe another win for Amd maybe it will be some Arm+GNC 2.0 core that would be great

There's nothing about ARM in this article.
Did you forget about Discovery prototype?

 

[jdwii]

http://www.fool.com/investing/general/2014/07/16/will-amd-supply-chips-for-nintendos-next-handhel-2.aspx

Maybe another win for Amd maybe it will be some Arm+GNC 2.0 core that would be great

There's nothing about ARM in this article.
Did you forget about Discovery prototype?

hence why i said maybe meaning i just claimed something with no proof and i except that(something others rarely do)
It would be amazing if it was Arm+GCN 2.0. Talk about a powerful GPU option with little TDP. Me and juan do not agree on most things but i agree with him on some things.

 

[de5_Roy]

TSMC reportedly losing Apple and Qualcomm orders
http://www.fudzilla.com/home/item/35255-tsmc-reportedly-losing-apple-and-qualcomm-orders
amd(gpu) and nvidia aren't going anywhere yet. i sense a future blamethrowing if gm200/204 yields suck!

There's no reason to believe the lowest TDP variant will have all the CUs enabled.

i agree. the possibility is intriguing though. the performance outlook won't be good for such part as it'd be thermally limited as well as memory-bottlenecked. other variants such as single module with full igpu will be cpu-bottlenecked too.

i want to see how a 15-25w, 1.2-2.5GHz, 4c/256-512 shaders(ddr3/L 1866) carrizo performs against a 2.0 GHz beema soc in an amd soc vs amd soc(ulp).

If they put the PCH on it's own power island and gate off the power it's basically zero power drain.

http://www.edn.com/design/integrated-circuit-design/4318458/Dual-threshold-voltages-and-power-gating-design-flows-offer-good-results

some of the stuff went over my head but overall it's very interesting.