AMD Discusses Next Gen FX CPUs, 20nm GPUs, 5-Way CrossfireX, FreeSync & DDR4

[fojacko]

As long as the new cpus have pretty good single core performance, ill be happy. In my opinion its the one aspect where AMD fails with their cpus. My overclocked 8320 has nowhere near the same fps on single core games (minecraft, source games) as my dads underclocked 2500k (he wanted a "silent" build and the cpu was too hot at stock with the fan speed being low enough to be quiet.).

[SkilledRebuilds]

Still waiting for Multi-APU Socketed motherboards......... and then with 2-4 APU's, in CrossfireX or Hybrid-Xfire would be very nice, rendering or gaming.

[Applepienation]

Seeing how little performance is gained from adding a 4th card, a 5th card would be a complete waste of money.

 

Took the words right out of my mouth. That being said, I think it would be a fun novelty to benchmark framerate, temps, noise, and power consumption.

[Raid]

5-way crossfire makes sense on the technical side of things. It's just that not a single game makes any use of even 4-way that's the problem. Optimization, anyone?

Such a shame we'll never see things like 8 GPU SLI on a single card anymore. Then again, what with modern games being quite a bit more complex than back then it'd make absolutely no sense to even bother spending money to optimize your game engines design for something which only a very, very small amount of your audience will get to experience, so not much point in making the cards for it either.

[deviant88]

If DirectX 12 delivers and fixes the eternal problem as promised, multithreading and cpu overhead, and only if we see an increase in adoption since it doesnt really require new hardware and im only reffering to cpu part of DX12,i know theres some stuff that wont work on current gpu's.If does deliver we wont need more than 4 core cpu's for a long time,all eyes are turning towards GPU's past 2-3 years.

 

If AMD wanted to do something they would have improved crossfire-x not 5 -way bullshit can you imagine how small the 5-way crowd is ? like <0.1% of all PC Master Race gamers.

 

The only thing AMD needs to do is another seriously improved FX 4-6-8 core at least 10-15% IPC not raw processing power added, and top end FX dog must be under <100W default TDP,they can achieve that easily on 14-22nm tech.But no brains or cash for that R&D i guess,i mean comeon intel has half,freaking HALF the fab process 14nm for mid  next year,da fuq are you doing amd,even if they only do APU's those are still 32nm more than 200% fab proc,isnt GF(i think only 28nm GF) and TSMC doing 20 or 22nm process for a long while now? Step up amd derps.

 

I cross my fingers amd will next cpu 16nm finfet and make cpu market crazy again competing with intels 14nm,i cant wait next year hopefully.CPU market right now is disgusting, only overpriced haswell's and aged bulldozer's.

[patrickjp93]

@deviant88

Carrizo is still on 28nm. Its successor could be 20 or 16nm. AMD can't step it up unless TSMC allows them to. TSMC put the proven 20nm nodes towards Samsung and Apple chips due to large orders/bigger money.

Haswell is not overpriced. Gamers are just dumb enough to buy it over SB/IB when it's really a chip designed for raw computational power. 4 ALUs per core is double the density of any AMD chip, and all the complex control logic around the computation units puts out all the nasty heat.

Haswell is not a chip designed to game. It's a chip designed to compute. With DX12 and a total redesign of OpenGL and OpenCL, your old 2500/2600k processors should be fine for another 5 years.

[martinrox1568]

True, a regular gamer would see no benefit, but that doesn't mean it isn't awesome that they are pushing the technology forward. If 5-way becomes a reality, maybe it will help them make 2-way, 3-way, and 4-way more efficient and have a smoother performance curve, rather than the current plateau at 3-way.

Well certainly it's not a good price to performance value (of course we have no idea how 5-way will run - we can guess, but it's just guessing), but this is not for regular gamers. 5-way would be for balls to the wall uber builds.

Why would you want 3 295x2's? I suppose they would consume less power than six R9 290/290x's, but six (or even five) single cards would be cheaper by a fair margin, easily allowing you to get the necessary size of PSU with the remainder.

Plus, perhaps it is a technical limitation? They have figured out 5-way, and maybe they couldn't do 6-way until 5-way was proven?

They're not pushing it forward. All they're doing is deleting the line of code that didn't allow this previously. They're not even doing that much because they never stated that they were developing it. Theoretically you can have 10 way crossfire if u have the bandwidth

[flibberdipper]

Meh. I want a 6-way setup. Just to piss case manufacturers off.

[cubercaleb]

Seeing how little performance is gained from adding a 4th card, a 5th card would be a complete waste of money.

Well if you are a professional GPU's happen to scale quite nicely, it's only games that scale like shit

[Trik'Stari]

I'd like to see what they will do to compete with Nvidia's next gpu series (no one seems to be sure at this point if it will be called 800 or 900 series lol)

 

Also, 5 or 6 way crossfire would finally give purpose to the rosewill blackhawk ultra. FINALLY a good reason to slap 9 noctua fans on the side of one of those.

[cubercaleb]

I'd like to see what they will do to compete with Nvidia's next gpu series (no one seems to be sure at this point if it will be called 800 or 900 series lol)

 

Also, 5 or 6 way crossfire would finally give purpose to the rosewill blackhawk ultra. FINALLY a good reason to slap 9 noctua fans on the side of one of those.

Cause you know, its not like people are already putting 6+ single slot Quadros in their systems.

[Trik'Stari]

Cause you know, its not like people are already putting 6+ single slot Quadros in their systems.

That's.....insane

[cubercaleb]

That's.....insane

Not for gaming, Quadros/Firepro GPU's actually scale quite nicely CUDA/OpenCL applications, in fact it's quite linear.

 

Plus you know, there also the people who bitcoin mine

[Trik'Stari]

Not for gaming, Quadros/Firepro GPU's actually scale quite nicely CUDA/OpenCL applications, in fact it's quite linear.

 

Plus you know, there also the people who bitcoin mine

I thought most bitcoin miners just used open benches?

[dalekphalm]

Not for gaming, Quadros/Firepro GPU's actually scale quite nicely CUDA/OpenCL applications, in fact it's quite linear.

 

Plus you know, there also the people who bitcoin mine

You're right about having that many workstation cards, but with Bitcoin mining, you would never use CrossFire/SLI anyway. Each GPU would be configured and mining independently, running it's own workload.

 

I thought most bitcoin miners just used open benches?

You're right. Most crypt coin (in general) miners that still use GPU's to mine would strive as hard as possible to keep their miners OUT of a case, in open air.

[Maslofski]

The new AMD Dumpster boards will be able ...........

 

oh, you mean the motherboard with integrated graphics that holds a APU an 5 extra gpus? now i know wich  one you mean! ASUS Dumpster VII. would be kinda heavy thought with water cooling blocks and all that cooling aid

[Jerakl]

Seeing how little performance is gained from adding a 4th card, a 5th card would be a complete waste of money.

Yes I know, it's ridiculous. 3-way max

 

e-peen

[dizmo]

e-peen

I see it as more of an "I like to waste money, don't have better things to spend it on because I have no life, and really have no idea what I'm doing or what makes sense."

[murratri]

Well now I need 2 295x2's, a r9 290x, a 1600 watt psu, a/c, and probably a 5820k so the cpu doesnt bottleneck, and a x99 mobo

Total: ~$4500

Rank: GabeN and linus's baby, Lukes the uncle

[KeltonDSMer]

Carrizo is still on 28nm. Its successor could be 20 or 16nm. AMD can't step it up unless TSMC allows them to. TSMC put the proven 20nm nodes towards Samsung and Apple chips due to large orders/bigger money.

 

 

Just in case you didn't know, Kaveri is made at GloFo on their "bulk" 28nm SHP, not TSMC. A leak from back in February clearly indicated that Carrizzo will also be a GloFo manufactured chip, (likely 28nm as well). Whatever Apple and Samsung are doing with TSMC should not effect Carrizzo.

 

http://www.extremetech.com/computing/176919-amd-leak-confirms-that-excavator-apu-will-be-28nm-and-that-some-production-is-moving-back-to-globalfoundries

^^Feb 2014

 

http://www.kitguru.net/components/cpu/anton-shilov/amd-carrizo-specifications-block-diagram-gets-published/

^^July 2014

 

Of course there is a possibility that Carrizzo ended up being "ported" back to a TSMC node, but I find this highly unlikely due to the time and cost of doing so. The lack of l3 cache and only 2mb of l2 makes me concerned about single threaded performance of this up and coming APU.

 

Did I miss an obvious road map or press release that says otherwise (very possible)?

[patrickjp93]

Just in case you didn't know, Kaveri is made at GloFo on their "bulk" 28nm SHP, not TSMC. A leak from back in February clearly indicated that Carrizzo will also be a GloFo manufactured chip, (likely 28nm as well). Whatever Apple and Samsung are doing with TSMC should not effect Carrizzo.

 

http://www.extremetech.com/computing/176919-amd-leak-confirms-that-excavator-apu-will-be-28nm-and-that-some-production-is-moving-back-to-globalfoundries

^^Feb 2014

 

http://www.kitguru.net/components/cpu/anton-shilov/amd-carrizo-specifications-block-diagram-gets-published/

^^July 2014

 

Of course there is a possibility that Carrizzo ended up being "ported" back to a TSMC node, but I find this highly unlikely due to the time and cost of doing so. The lack of l3 cache and only 2mb of l2 makes me concerned about single threaded performance of this up and coming APU.

 

Did I miss an obvious road map or press release that says otherwise (very possible)?

I believe I read back in June AMD had moved Carrizo back to TSMC due to friction at GloFo, but that landed them behind the queue of Apple and Samsung. It's possible I'm confused. There are many techs I follow closely and I may have gotten lost.

 

That said, I know for a fact Carrizo is still on 28nm. I doubt we'll see much improvement, and if it really does get the stacked eDRAM or HBM onboard, the prices will go higher and make Intel's Iris Pro competitive from a cost standpoint. Iris 6200 vs. Carrizo is going to be an interesting battle. I predict the MBPrs go without a dGPU in late 2015 if Broadwell's top SKU really has 96 EUs and delivers on the promised 2TFlops of peak performance with improved eDRAM. Either that or we might see the first truly "Pro" offering with a small Quadro or FirePro being integrated for actual workstation replacement.

[D3MON]

What do you exactly mean with this? Not sure if you understood me but what I'm trying to say is that AMD should put the 8350's multithreaded performance in a quad core without CMT (so no modules bs). That's only achieveable by improving IPC by 70% orsomething.

Let me put it this way, a new "FX2 4600" just a quad core nothing special performs multithreaded wise exactly the same as the 8350 in every benchmark. The only difference being the single threaded performance is massive that way, kicking them who knows above intels IPC.

yeah i got you and agree (to an extent), if i had to chose between for instance an i3 dual core with hyperthreading vs an i5 quad core that performed the same id chose the i5, however if i had to chose between an 8 core with no hyperthreading (like a 5960x with HT off) and a 6 core with HT (like 5820k) then id chose the 6 core, because as i found out with my phenoms, having 4 powerful cores is great but when you have a game like bf3 that requires 3 cores +ideally (2 minimum, 4 prefferable) and an app like fraps that requires 2 cores (or threads atleast) then you can run into issues, i ran bf3 on 2 cores with no loss in avg frames but both cores at 100% and ive run it with 3 cores at 75% and set the remaining core to fraps but fraps maxed it out and recording proved tricky, hence i moved to an fx8 which is in my mind just a phenom with hyperthreading in all effectiveness (i know its not, i know about the CMT design and architecture but fundimentally its a great way to describe it).

i can have bf4 run on modules 1, 2 and 3 with 2 threads apiece and have another module with 2 threads for fraps running, not only that but I can run bf4 or hardline on 2 modules 4 threads (and have done as seen in my "stuff" gallery), and have 1 module for fraps as mentioned and have another one in the background transcoding previously recorded gameplay, and my 8320 cost the same as some i3's do...

of course as always this is because i chose it for my specific purposes, if i was only playing retro games on a htpc id be chosing a pentium k because old games only use 1 core so i'd have core 0 set for windows and core 1 for the game.

id imagine id have a use for an i3/fx4300 somewhere, maybe if i was a casual gamer and played lots of badly made games like "kinda" use 4 threads but mainly rely on 2 cores. but thats not me.

id happily have an fx6300 apu but only with a powerful gpu part to take over the transcoding duty's once HSA becomes more prevelant. 3 modules 6 threads for gaming, and 10 Compute Units (7770) for transcoding or recording gameplay on the fly, doing on the cpu what shadowplay does on the gpu. also with ddr4 now available and likely to drop in price over the next year (as it always does with modern tech) the doubling in memory bandwidth should lead to some nice improvements and possibilities.

imagine that fx6300/7770 apu (20nm) being on the fm2+ playform on a board like this...

with 2x r9-290's with the 3m/6t cpu part playing the game and the iGPU recording gameplay, transcoding it into a small file saved to your system so it can either record the whole game or the last 15 minutes, or last 10 minutes, or just using the whole apu to transcode a large file down but instead of having "8 cores" like i currently do its 6 cores+10 CU's, bring it on

[KeltonDSMer]

I believe I read back in June AMD had moved Carrizo back to TSMC due to friction at GloFo, but that landed them behind the queue of Apple and Samsung. It's possible I'm confused. There are many techs I follow closely and I may have gotten lost.

 

 

Ahh okay I probably missed it, I'm likely to be the one confused here .

 

Either way, I'm not holding my breath for the 8850K or whatever they are going to call the DT Excavator APU. Without more memory bandwidth, I just don't see what they could possibly do to "wow" the industry. The last road map I saw showed ddr3 (not ddr3/ddr4 like the early ones), and I feel we would have heard about a leak if eDRAM or HBM was actually in the works. The 65w designation doesn't help me get very jazzed about it either...

 

My brain has trained itself to not get excited about these releases anymore. The hype train has not been kind to AMD releases for the better part of the decade, so I guess I expect very little in order to not be let down once benchmarks start rolling out. I do, however, want them to blow me away with something competitive sometime in the next couple years.

 

I wish a few things went a little better for them 5 years ago, we might have 28nm steamroller 4 module fx parts right now waiting for an excavator refresh. Not that it would have been groundbreaking, but it would have made it easier to hang on as they get to the end of the bulldozer CMT "experiment".

[patrickjp93]

Ahh okay I probably missed it, I'm likely to be the one confused here .

 

Either way, I'm not holding my breath for the 8850K or whatever they are going to call the DT Excavator APU. Without more memory bandwidth, I just don't see what they could possibly do to "wow" the industry. The last road map I saw showed ddr3 (not ddr3/ddr4 like the early ones), and I feel we would have heard about a leak if eDRAM or HBM was actually in the works. The 65w designation doesn't help me get very jazzed about it either...

 

My brain has trained itself to not get excited about these releases anymore. The hype train has not been kind to AMD releases for the better part of the decade, so I guess I expect very little in order to not be let down once benchmarks start rolling out. I do, however, want them to blow me away with something competitive sometime in the next couple years.

 

I wish a few things went a little better for them 5 years ago, we might have 28nm steamroller 4 module fx parts right now waiting for an excavator refresh. Not that it would have been groundbreaking, but it would have made it easier to hang on as they get to the end of the bulldozer CMT "experiment".

We have a lot of "leaks" regarding stacked onboard memory, but how that manifests we do not know.

[KeltonDSMer]

We have a lot of "leaks" regarding stacked onboard memory, but how that manifests we do not know.

 

I worded my post poorly, and I am aware HBM was developed in a partnership by AMD and Hynix. I was referring to information "leaks" about HBM on Carrizzo, not just the existence of such a memory technology.