AMD CPU speculation... and expert conjecture

[8350rocks]

http://www.tomshardware.com/news/Xbox-One-Chat-Headset-Buy-Separate-Purchase,23263.html

Just as M$ looked like they were going to be competitive in the consoles again with changes on policy after gamers made an outcry. They shoot themselves in the foot again!

Why does this not at all surprise me...?

Steve Ballmer is clearly a moron.

 

[The Q6660 Inside]

http://www.tomshardware.com/news/Xbox-One-Chat-Headset-Buy-Separate-Purchase,23263.html

Just as M$ looked like they were going to be competitive in the consoles again with changes on policy after gamers made an outcry. They shoot themselves in the foot again!

Why does this not at all surprise me...?

Steve Ballmer is clearly a moron.

Well, M$ has been on a roll with stupid decisions since Windblows 8. That is why this is not surprising at all...............

 

[8350rocks]

http://www.tomshardware.com/news/Xbox-One-Chat-Headset-Buy-Separate-Purchase,23263.html

Just as M$ looked like they were going to be competitive in the consoles again with changes on policy after gamers made an outcry. They shoot themselves in the foot again!

Why does this not at all surprise me...?

Steve Ballmer is clearly a moron.

Well, M$ has been on a roll with stupid decisions since Windblows 8. That is why this is not surprising at all...............

I admit, you got a chuckle for "Windblows 8"....lol.

 

[noob2222]

http://www.tomshardware.com/news/Xbox-One-Chat-Headset-Buy-Separate-Purchase,23263.html

Just as M$ looked like they were going to be competitive in the consoles again with changes on policy after gamers made an outcry. They shoot themselves in the foot again!

Why does this not at all surprise me...?

Steve Ballmer is clearly a moron.

not like the ps4 included headset is anything spectacular.

id say its pretty chincy at best.

 

[8350rocks]

Yes, but M$ is promoting a headset as a "must have" for a system that comes without it, and is already $100 more than the more powerful competitor that includes a flimsy headset.

EDIT: Can you say blatant money grab?

 

[noob2222]

either one is just a marketing gimmick. you don't actually need one since the kinect that comes with the system has an integrated mic. the PS4 comes with a "headset" even though its just a wire and an earpiece. Gamers will opt to buy a real "must-have" headset, but hey, they can brag about including one.

http://reviews.cnet.com/must-have-gadgets/

gotta buy one of each since they are must-haves?

But M$ is stupid with the proprietary mic jack that hooks to the controller.

selling point for me is bluetooth = PS4. Logitech H800 ftw.

XB one went wifi direct and no bluetooth. means all your bt xb 360 devices are useless.

 

[8350rocks]

either one is just a marketing gimmick. you don't actually need one since the kinect that comes with the system has an integrated mic. the PS4 comes with a "headset" even though its just a wire and an earpiece. Gamers will opt to buy a real "must-have" headset, but hey, they can brag about including one.

http://reviews.cnet.com/must-have-gadgets/

gotta buy one of each since they are must-haves?

But M$ is stupid with the proprietary mic jack that hooks to the controller.

selling point for me is bluetooth = PS4. Logitech H800 ftw.

XB one went wifi direct and no bluetooth. means all your bt xb 360 devices are useless.

Well, the issue is interference on the Kinect. M$ is basically all but saying...if you're going to talk while playing, you want a headset because Kinect sucks.

I personally don't care one way or the other...but I know what my kid is getting for Christmas, and it won't be from M$.

 

[8350rocks]

ECS FM2+ mini-ITX board spotted @ Computex 2013

If the boards are already made, then that means that Kaveri is right around the corner. Additionally, this would be a great HTPC platform. It uses an A78 chipset, which correlates to an updated A75 chipset. Unless they got it backwards and it's an A87 chipset, though I don't think the board would be large enough to support full blown CF like the A85 chipset did (though an ATX version would easily be able to do so).

Either way, interesting read. The website is natively German, so google translate helps but the English is rough in a few places.

 

[The Q6660 Inside]

ECS FM2+ mini-ITX board spotted @ Computex 2013

If the boards are already made, then that means that Kaveri is right around the corner. Additionally, this would be a great HTPC platform. It uses an A78 chipset, which correlates to an updated A75 chipset. Unless they got it backwards and it's an A87 chipset, though I don't think the board would be large enough to support full blown CF like the A85 chipset did (though an ATX version would easily be able to do so).

Either way, interesting read. The website is natively German, so google translate helps but the English is rough in a few places.

I would assume Kaveri will release in October or November. Still, I prefer a retooled, old FX-6X or Core 2 Quad/Extreme in an HTPC case :3.

 

[Cazalan]

Kaveri is supposedly just waiting on GF 28nm to ramp. Both AMD/GF have been dead silent.

July 18th is AMDs next earnings announcement. Expect some news then.

 

[mlscrow]

Okay, so the ultimate conclusions of this thread can be:

A) hajifur is a complete waste of worldly resources.
B) The 220W TDP of the 5GHz FX9590 is spot-on, as it's simply an overclocked FX8350, which we all know is based on Piledriver, which we all know requires exponentially more power as you overclock it beyond 4GHz. 220W is actually surprisingly low for a 5GHz OC (compared to an FX8350 OC'd to 5GHz), but we can attribute that to process maturity and AMD lowballing.
C) High-end performance enthusiasts don't care about power consumption anyway.
D) The FX-8350 provides better performance per dollar than its Intel counterparts.
E) The FX-8350 provides worse performance per watt than its Intel counterparts.
F) The point of this thread is supposed to be Steamroller anyway, so it's dumb to battle the troll in this virtual stadium, lets get back on topic.
G) Steamroller should put AMD back in direct competition with Intel in the high-end.
H) AMD only shows a quad core Steamroller available for 2014, which is very, very sad and hopefully they'll realize their mistake sooner than later, because we all want an 8-core Steamroller FX for 2014, don't we?

B) The FX9590 isn't simply an overclocked FX8350. We don't have details, but either it is a chip with extraordinary thermal and electric tolerances preselected from the main FX8350 line or includes some 2.0 revision like Richland. In any case, the FX9590 will consume less power than an OC FX8350.

E) Compared with ordinary i7 line such as the 3770k or 4770k, compared with extreme series such as the 3960x or 3970x I am not so sure.

H) I don't know any updated desktop roadmap for 2014. I only know server roadmap and the desktop for 2013. If steamroller is 2 threads per module I would wait some 4 module FX, but if steamroller is 4 threads per module as speculated before, then a 2 module FX makes more sense for general public.

B) Agreed, we don't have the details, but all Richland was, outside of the GPU, was higher clocks due to the maturing of the process and slight power management tweaks. In the end, it's still the same arch, hence my reasoning that all the FX9590 will be, is an overclocked FX8350. I would still prefer 28nm SR cores for a new FX, wouldn't you?

E) I did say "counterparts". To me, the counterpart to an AMD 4m/8t APU is a 4c/8t APU from Intel, thus the 2700K, 3770K, 4770K series. Once you go on to the E series (with the 3820 as maybe the exception), AMD doesn't have any counterpart.

H) Historically speaking, the FX line were based on AMD's 1P server platform offerings. If there was an 8-core part made for the 1P server platform, we would eventually see an 8-core Desktop FX version. So, with that, one can speculate that since AMD shows no 8-core 1P server part with SR cores, but only shows a 4-core part, then the best we can expect to see, in terms of SR core parts available for desktops are 4-core, meaning 2 module parts for 2014. And although I agree that a 4-core part does make sense for the general public, it's the enthusiasts, such as myself and most people here, who are sorely disappointed that there is no mention of an 8-core part. Fingers crossed for the future.

 

[hcl123]

All right guys, it was fun to battle and all, but we really should get back on topic. How about that Steamroller?

I'll start. Steamroller is apparently only announced on FM2(+) currently, judging from the server roadmaps. The server die is used in the high-end desktop as well and 2014 server is going to be a tweaked Piledriver, not Steamroller. Thoughts?

FAKE... as it seems...

http://www.abload.de/img/9gtasti18.jpg

Its the first time i saw the logic part get an astounding shrink, the BLACK BOXES in that image is ~50% shrink (edit) for the same functionality structures,* yet the SRAM macros of the cache (L2 in this case) doesn't scale at all* ( that 2MB L2 should be quite visible smaller than in PD, and 4 MB is no likely at all, it will mean 50% overall shrink and that is not 28nm for sure) LOL

And its not due to a superimposition, the "PINK" first L2 arrays, 16 in total, where already present in the first *Module* image in the same position they are in the above, of a ST module on top of a PD module.

https://securecdn.disqus.com/uploads/mediaembed/images/500/9389/original.jpg

It would be nice anyway, to think 4 ALUs + 4AGUs, but what the hell for if not for 4 threads module, the old ALU+AGU scheme of Athlon K10 was deprecated when OoO (out-of-order) memory ops, STL (store to load) and "Data Speculation" was introduced( not really need to calculate every address)... is also to scratch head. The double FlexFPU is more nicer, i think AMD will end do it, but not 4MMX + 4 FMAC each as in that picture (2MMX + 4FMAC is more likely on those high density libraries).

Also i think AMD will end up introducing "2 fetch front-end" engines, decode is not the most of concern if you have a small decoded cache after decode, and a stack offloading pipe after fetch(like K10 did). Besides the rumor based on revelations and analyzed at RWT, of the developer guides is that there will be the same 4 decode pipes(not 8 like that picture presents)... only perhaps arranged in a different way in the "vertical multithreading scheme of things", 2 dedicated input buffers(1 per thread), and 2 dedicated output dispatch buffer can reasonably simulate 2 decode engines.

In the end we got to applaud, who ever wasted weeks doing that god job, is very good at computer graphics arts.

 

[hcl123]

Personally if they are going to be doing a new chipset (1090X/FX) I would like to see AM4. It would like to see a Steamroller processor for AM3+ 990FX, but top of the line Steamroller processors running AM4 1090FX. If they are going to come out with a new chipset, requiring a new motherboard anyway I would like to see AM4 which would hopefully mean "high end" FX processors through excavator. It would be a shame if AMD dropped its high end FX line and only produced APUs.

I know I read somewhere awhile back that AMD was "unifying" its line which I took it to mean producing only APUs and the FX line would end. I personally want to see the high end FX line continue and keep nipping at Intel's heels until they are able to surpass Intel. A commitment like AM4 1090FX would in my mind show AMD is committed to continuing its FX line, hopefully through excavator.

Yes it usually means a new motherboard lol... yet they can use the same sockets... one doesn't invalidate the other.

And unifying is a good suspect... Berlin APU for servers, if the I/O block is out of the die, could function in the same scheme of a AM3+ socket, the central interconnect for HSA is HyperTransport, the base of IOMMU v2.5( an HSA standard)... not crap PCIe...

Perhaps AM4 could have CPUs and APUs in the same socket, with no difference... an have PCIe3 + HTX4.0 combo slots directly from the socket dies (ITS IN THE PATENTS)... much better than pcie alone like intel.

 

[hcl123]

This is completely irrelevant to the topic at hand.

I'd say he's just a regular spammer, haha.

Anyway, in regards to SR without iGPU talk... Is it going to exist at all? If so, they should move it to a timeframe that allows to have DDR4 (mainly) in their MoBos. Man, I wish they do just that for the "high end" SR parts.

APUs would be the most benefited from it, but the move to GDDR5 will be enough for a good time. In all honesty, 8GB or RAM for Windows computers is still enough and Win7 (not sure about 8) goes around the 2GB mark and 1.5GB when "optimzed", so there's still a lot of memory pool for games and all the rest of the stuff. I wonder if they'll use GDDR5 sticks or what, haha.

Cheers!

Don't count on GDDR5 for now. If you do the right accounts, 3 channels of DDR3 2400 have the double bandwidth of 2 channels of DDR3 1866 ( the official supported)... and AMD could sell those with AMP profiles. It would be grand if Kaveri could have 2x the performance of a Trinity (which i doubt will happen).

Its quite possible Kaveri will have ESRAM on board like Wuii or XB one... its more scalable... if not 3 channels of DDR3 at higher freq/bandwidth(with perhaps a small L3 block...4MB maybe)(edit).

 

[hcl123]

In the end, it's still the same arch, hence my reasoning that all the FX9590 will be, is an overclocked FX8350. I would still prefer 28nm SR cores for a new FX, wouldn't you?

Not if it comes with Turbo core 3.0... and probably tweaked at memory controller also. Centurion could be based on Richland that got average 5% better IPC than Trinity on the CPU side (the GPU is more specially at games, and compute where it can go >20%)... Then Centurion will be Vishera 2.0.

http://m.hexus.net/tech/news/cpu/56653-amd-set-release-5ghz-clocked-fx-9590-cpu/

( 5% clock to clock will be half hasfail lol... not bad for such a minor revision.. plus 20% more due to clock lol)

 

[-Fran-]

Don't count on GDDR5 for now. If you do the right accounts, 3 channels of DDR3 2400 have the double bandwidth of 2 channels of DDR3 1866 ( the official supported)... and AMD could sell those with AMP profiles. It would be grand if Kaveri could have 2x the performance of a Trinity (which i doubt will happen).

Its quite possible Kaveri will have ESRAM on board like Wuii or XB one... its more scalable... if not 3 channels of DDR3 at higher freq/bandwidth(with perhaps a small L3 block...4MB maybe)(edit).

But going tri or quad channel DDR will make the expense of the boards skyrocket. AMD doens't need expensive MoBos IMO. Specially when they're not perf kings of the hill.

As it stands now, dual channel DDR3 is more than plenty for CPUs without iGPU (yes, including Intel's -E series), so DDR4 or GDDR5 only makes sense for APUs; or makes sense to rush it in that territory, which don't crave for enthusiasts that much, so going tri or quad channel doesn't make any sense to me, to be honest, from a price performance point of view. Nice to have, but not really needed IMO.

AMD APUs are still horribly constrained by heat draw, so including an even bigger IMC isn't a great idea at first glance.

Cheers!

 

[hcl123]

NO it could even make it cheaper... offer boards with only 3 DIMM slots instead of the usual 4... or have the same 4, in case only one channel is expanded to 2 slots, which should be the same price.

DRAM upgrade could be in the DIMMS themselves not in more DIMMs, AMD has good support for dual bank (both sides) DIMM toplogies... perhaps you could match 4GB with 8GB DIMMS without problem.

Besides Intel done it before with Nehalem... the problem then was that DRAM was much more expensive per GB, an then you needed all 6 DIMM slots yet usually restricted to 16GB total... now you can have, and not very expensive, 24GB with only 3 DIMMS... 48GB if this supports the 8Gbit DRAM devices in dual rank/"dual-bank".

UPGRADE
DDR4 will be much more expensive, a price nightmare for expansion, you need a Buffer Chip (as if it were "registered" DRAM) in those DIMMS for expansion... otherwise you'll get only 1 DIMM per channel. DDR4 has a point-to-point topology that don't even support the "clam-shell" topology of GDDR5.

Besides the DDR4 that is to be in 2014 is 2133 and 2400.. perhaps 2666 by end of 2014... hardly any improvement in bandwidth compared to DDR3, only power(but latencies are worst comparatively), and very far from the top 4266 that is to be certified by JEDEC.

 

[The Q6660 Inside]

I'm just posting this to sub to this.

Gah, what is with these flags and faces! Any explanation? On the topic of FM2b APUs, I think they may be waiting for 28nm FD-SOI from GF to ramp up.

 

[hcl123]

AMD APUs are still horribly constrained by heat draw, so including an even bigger IMC isn't a great idea at first glance.

You have a Richland A10-6700 at 65W for the all chip yet above 4.0Ghz... the 6800 is less than 100W for sure ( the K versions are always deliberately over rated )

I think facts contradict your statement. Besides the IMC is not the worst offender, not even close, not even with 6 channels would be, the GPU is. period

 

[hcl123]

Gah, what is with these flags and faces! Any explanation? On the topic of FM2b APUs, I think they may be waiting for 28nm FD-SOI from GF to ramp up.

It could be, the plans at GF are clearly future going in this respect, down to the 10nm node.
http://www.advancedsubstratenews.com/2013/04/gfs-two-flavors-of-fd-soi-kengeri-explains-exclusive-asn-qa/
(THAT IS AN EXCLUSIVE INTERVIEW WITH A TOP EXECUTIVE OF GF)

Besides porting a design from "planar" bulk to "planar" FD-DOI is the less complicated of all transitions so far... that has been one strong point of STMicro on the marketing front. Planar "bulk" and "FD-SOI" share the same BEOF(back end of line) steps ... and some middle to...

But i think it was already revealed (don't remember where but i read it) in an exclusive interview with AMD that Kaveri will debut on "bulk" process. If its low power low clock, mobile parts, as Richland also launched mobile first, is not a bad bet.

EDIT:
50% better at Vdd= 1 volt and 200% at 0.6v.. kaveri could gain much more than 60% perf/power, clearly ahead of intel, cherry on top of cake to "knock your socks off", "forward back gate biasing" was able to get 40% more clock than without it ... simple (it is a relative simple process) is beautiful lol

http://www.advancedsubstratenews.com/2013/06/fully-depleted-soi-and-more-at-vlsi-kyoto-some-knock-your-socks-off-papers-2/
(just read the head lines of those VLSI (kyoto) conference presentations)

And they will catch intel at 14nm, they even plan to reach there first (finfet is quite harder to push and do right)... a bold plan by any metric...

http://www.advancedsubstratenews.com/2013/06/which-will-hit-the-14nm-jackpot-first-fd-soi-or-finfet-gauntlet-down-race-on-2/

 

[-Fran-]

AMD APUs are still horribly constrained by heat draw, so including an even bigger IMC isn't a great idea at first glance.

You have a Richland A10-6700 at 65W for the all chip yet above 4.0Ghz... the 6800 is less than 100W for sure ( the K versions are always deliberately over rated )

I think facts contradict your statement. Besides the IMC is not the worst offender, not even close, not even with 6 channels would be, the GPU is. period

Well, my own fact is sitting under my TV right now, underclocked for the very same reason I told you about, haha.

I agree with the quad channel 4 dimms approach though. I wonder if AMD could do that to extend the use of DDR3. On the other hand, wasn't DDR4 like the second coming of Jesus "DDR" Christ? Why make it closer to ECC RAM than keeping it unbuffered? Are you sure they'll keep Server and Desktop grade RAM buffered? You got me confused.

Cheers!

 

[hcl123]

For server the "buffer" is no problem... they need it... the idea was to prop up the market at all levels "client/DT" included, but too late for Elpida that went belly up.

DDR4 will have plenty of success in servers.. more cores need more bandwidth. Client specially with the push of integrated GPGPUs, not even DDR4 will be enough, 4 year ago "top GPU" level of performance are clearly at hand, so this ones will go for HBM (wide I/O, Memory Cube) instead(that or very large caches on die like ESRAM, CW).

 

[de5_Roy]

Spoiler

Don't count on GDDR5 for now. If you do the right accounts, 3 channels of DDR3 2400 have the double bandwidth of 2 channels of DDR3 1866 ( the official supported)... and AMD could sell those with AMP profiles. It would be grand if Kaveri could have 2x the performance of a Trinity (which i doubt will happen).

Its quite possible Kaveri will have ESRAM on board like Wuii or XB one... its more scalable... if not 3 channels of DDR3 at higher freq/bandwidth(with perhaps a small L3 block...4MB maybe)(edit).

But going tri or quad channel DDR will make the expense of the boards skyrocket. AMD doens't need expensive MoBos IMO. Specially when they're not perf kings of the hill.

As it stands now, dual channel DDR3 is more than plenty for CPUs without iGPU (yes, including Intel's -E series), so DDR4 or GDDR5 only makes sense for APUs; or makes sense to rush it in that territory, which don't crave for enthusiasts that much, so going tri or quad channel doesn't make any sense to me, to be honest, from a price performance point of view. Nice to have, but not really needed IMO.

AMD APUs are still horribly constrained by heat draw, so including an even bigger IMC isn't a great idea at first glance.

Cheers!

i keep finding information saying sram is much more expensive than dram(and gddr5) and wiiu has edram (the one on the package). won't it be better add just 4-6 MB L3 cache(afaik, made of sram) instead of adding 32MB esram? 32MB sram seems like 6-8 times the cost among other issues.

apus started at entry level consumer devices which also migrated to servers. they're still consumer devices in essence. tri/quad channel seems like a move suitable for high end enthusiast level hardware that will increase motherboard costs and may not help with mini itx or micro atx form factors. it's bad enough that there is so little demand for amd mini itx motberboards...

Spoiler

NO it could even make it cheaper... offer boards with only 3 DIMM slots instead of the usual 4... or have the same 4, in case only one channel is expanded to 2 slots, which should be the same price.

DRAM upgrade could be in the DIMMS themselves not in more DIMMs, AMD has good support for dual bank (both sides) DIMM toplogies... perhaps you could match 4GB with 8GB DIMMS without problem.

Besides Intel done it before with Nehalem... the problem then was that DRAM was much more expensive per GB, an then you needed all 6 DIMM slots yet usually restricted to 16GB total... now you can have, and not very expensive, 24GB with only 3 DIMMS... 48GB if this supports the 8Gbit DRAM devices in dual rank/"dual-bank".

UPGRADE
DDR4 will be much more expensive, a price nightmare for expansion, you need a Buffer Chip (as if it were "registered" DRAM) in those DIMMS for expansion... otherwise you'll get only 1 DIMM per channel. DDR4 has a point-to-point topology that don't even support the "clam-shell" topology of GDDR5.

Besides the DDR4 that is to be in 2014 is 2133 and 2400.. perhaps 2666 by end of 2014... hardly any improvement in bandwidth compared to DDR3, only power(but latencies are worst comparatively), and very far from the top 4266 that is to be certified by JEDEC.

did you count the slowly rising cost of desktop ddr3 ram? for budget/entry level, higher cost could become an issue. in addition to that, all the quad/tri channel memory kits i've seen cost a lot more than dual channel kits. i assume it's because of the validation process.

edit:
Kabini laptops available in US
http://www.fudzilla.com/home/item/31811-kabini-laptops-available-in-us
AMD leaks model numbers of future Kabini APUs
http://www.cpu-world.com/news_2013/2013062701_AMD_leaks_model_numbers_of_future_Kabini_APUs.html
Microsoft Xbox One Cracked Opened: First Photos of Hardware Inside Xbox One Emerge.
http://www.xbitlabs.com/news/multimedia/display/20130627222108_Microsoft_Xbox_One_Cracked_Opened_First_Photos_of_Hardware_Inside_Xbox_One_Emerge.html

 

[The Q6660 Inside]

http://techcrunch.com/2013/06/26/windows-8-1s-start-button-isnt-a-start-button/ Oh here we go again!

 

[-Fran-]

I love fresh MS bashing in the morning, hahaha.

And in regards to the RAM convo... SRAM is like on-die RAM or packaged RAM? Packaged RAM works wonders, but limits your upgrade path, since making external modules would be a lot slower and prolly slow down more than speed up things; that's a totally without-deep-knowledge assumption.

Cheers!