AMD CPUs, SoC Rumors and Speculations Temp. thread 2

[jimmysmitty]

If anything we can also blame AMD because instead of trying to push a fully 64bit uArch out, they extended the x86 life by adding the 64bit extensions to it.

Hmm, push a uArch you have a license to produce...or push a uArch you have no rights to produce...?

I think, from a business perspective, we can see why AMD did what they did.

Not Itanium, they could have produced their own 64bit uArch instead is my point. Instead we still have, and will for the foreseeable future, x86 which does have some disadvantages.

 

[ex_bubblehead]

Final reminder that all comments in this thread are to be directly related to the thread topic. The last 2 batches of deleted posts were way off topic. Further off topic posts will result in sanctions being handed out.

 

[crazyperson214]

I'm curious, does anyone have any facts about AMD's new chips fitting the new AM4 socket? I read they might be 14 nm architecture with 2 threads per core, but am wondering if anyone can confirm this? Also, will these new chips be competitive with Intel's new i7 6700k cpu?

 

[sprsk]

Any rumor for new FX series ? or the Wraith Cooler is the last upgrade ?

 

[Cazalan]

Some new evidence for possible 32 core Zen.

http://wccftech.com/amd-zeppelin-32-x86-cores-vega-10/

 

[de5_Roy]

AMD Updates its Desktop Lineup For 2016
http://semiaccurate.com/2016/02/02/amd-updates-it-desktop-lineup-for-2016/
AMD Updates Its Roadmap for 2016 and 2017
http://semiaccurate.com/2016/01/21/38060/
Who Controls the User Experience? AMD’s Carrizo Thoroughly Tested
http://www.anandtech.com/show/10000/who-controls-user-experience-amd-carrizo-thoroughly-tested
AMD Launches Excavator on Desktop: The 65W Athlon X4 845 for $70
http://www.anandtech.com/show/10009/amd-launches-excavator-on-desktop-the-65w-athlon-x4-845-for-70
ASRock A88M-ITX/ac gives AMD APUs a fun-sized foundation
http://techreport.com/news/29682/asrock-a88m-itx-ac-gives-amd-apus-a-fun-sized-foundation
AMD releases new desktop processors for socket FM2+
http://www.cpu-world.com/news_2016/2016020401_AMD_releases_new_desktop_processors_for_socket_FM2_.html

Some new evidence for possible 32 core Zen.

http://wccftech.com/amd-zeppelin-32-x86-cores-vega-10/

every time read it, reminded me of the hindenburg disaster.. 😛
i wonder if nintendo NX will have it... *dreams*

 

[juanrga]

AMD Updates its Desktop Lineup For 2016
http://semiaccurate.com/2016/02/02/amd-updates-it-desktop-lineup-for-2016/
AMD Updates Its Roadmap for 2016 and 2017
http://semiaccurate.com/2016/01/21/38060/

There is not real roadmap update, the slides say the same than at FADD2015. The expected products are the same.

Some new evidence for possible 32 core Zen.

http://wccftech.com/amd-zeppelin-32-x86-cores-vega-10/

every time read it, reminded me of the hindenburg disaster.. 😛
i wonder if nintendo NX will have it... *dreams*

The Nintendo NX will not get that. That is a huge and costly SoC for servers/HPC with TDP rating well above the one that fits inside a gaming console. Not to mention also that no game programmer in the world would be able to generate 64 threads to maintain that beast fully used.

 

[sitandspin]

I thought carrizo was never ever ever coming to desktop athlon 845k must be vaporware

 

[Kulasko]

https://cds.cern.ch/record/2128536
Slide 13
Top zen solution is 32 cores, MCM with 2 chips a 16 cores and quad channel DDR4

 

[jimmysmitty]

https://cds.cern.ch/record/2128536
Slide 13
Top zen solution is 32 cores, MCM with 2 chips a 16 cores and quad channel DDR4

That is probably the server solution, no need for 32 cores on desktop.

 

[de5_Roy]

That is probably the server solution, no need for 32 cores on desktop.

+1.
however, i see some usage for the <16-20c opterons as long as they're priced well. if amd's can deliver halfway decent thermal and single core performance for decent price, those will find their own place among users.

 

[salgado18]

That is probably the server solution, no need for 32 cores on desktop.

+1.
however, i see some usage for the <16-20c opterons as long as they're priced well. if amd's can deliver halfway decent thermal and single core performance for decent price, those will find their own place among users.

These users are better off getting Opterons, which could be ideal for them if compatible motherboards are decent priced and easy to find.
99% of desktop users don't need more than 8c/16t, if not 4c/8t. (I'd get an overclockable 20+ cores Opteron if they exist )

By the way, are there overclockable Opterons? Could there be with Zen? Could it be a viable strategy for gaining new market share?

 

[jimmysmitty]

That is probably the server solution, no need for 32 cores on desktop.

+1.
however, i see some usage for the <16-20c opterons as long as they're priced well. if amd's can deliver halfway decent thermal and single core performance for decent price, those will find their own place among users.

These users are better off getting Opterons, which could be ideal for them if compatible motherboards are decent priced and easy to find.
99% of desktop users don't need more than 8c/16t, if not 4c/8t. (I'd get an overclockable 20+ cores Opteron if they exist )

By the way, are there overclockable Opterons? Could there be with Zen? Could it be a viable strategy for gaining new market share?

Not viable at all. Servers have changed vastly in the past few years. No longer does one server box do one job. Instead most people use VMWare or Hyper-V to make a single server box do multiple jobs. Our corporate office went from 4-6 racks filled with servers to just two racks of much smaller servers.

Now they need more cores that are powerful and low power with tons of RAM.

Of course there are still some situations where you might use one server for one purpose but considering that a file server doesn't need 32 cores and 384GB of system RAM it is a waste of power.

 

[cdrkf]

One thing I find interesting is the 8 channel ram to 32 cores ratio. That would suggest only 1 memory channel on a quad core, 8 thread consumer part, which surely would be where the volume sales are. My only thought is they could construct such a part using 2 quad core blocks with half the cores disabled? I can't see consumers being very impressed with a mid to high end part relying on single channel ram.....

 

[de5_Roy]

One thing I find interesting is the 8 channel ram to 32 cores ratio. That would suggest only 1 memory channel on a quad core, 8 thread consumer part, which surely would be where the volume sales are. My only thought is they could construct such a part using 2 quad core blocks with half the cores disabled? I can't see consumers being very impressed with a mid to high end part relying on single channel ram.....

that's linear math, not how the cpus or memory channels are configured. yes, you can derive a 2-4c cpu/soc with a single channel memory controller by binning and/or lasering off bits. but amd will very likely be sellling processors with dual channel IMC unless they really, really want to get into ULP socs.

 

[crazyperson214]

If this new mystery chip for the AM4 socket competes with Intels 6th gen i series, I just hope AMD maintains its excellent pricing points.

 

[cdrkf]

One thing I find interesting is the 8 channel ram to 32 cores ratio. That would suggest only 1 memory channel on a quad core, 8 thread consumer part, which surely would be where the volume sales are. My only thought is they could construct such a part using 2 quad core blocks with half the cores disabled? I can't see consumers being very impressed with a mid to high end part relying on single channel ram.....

that's linear math, not how the cpus or memory channels are configured. yes, you can derive a 2-4c cpu/soc with a single channel memory controller by binning and/or lasering off bits. but amd will very likely be sellling processors with dual channel IMC unless they really, really want to get into ULP socs.

I was ony basing this on the previous info suggesting that Zen is built in quad core blocks. Maybe the memory controllers aren't comtained within those blocks (in which case the memory config it totally independent).

 

[salgado18]

I was ony basing this on the previous info suggesting that Zen is built in quad core blocks. Maybe the memory controllers aren't comtained within those blocks (in which case the memory config it totally independent).

I don't think it makes sense from a marketplace perspective to offer a four-core processor at near Haswell speeds with only one memory channel. Many lower dual-core CPUs would have more bandwidth than that. The memory controller has to be independent from core or block count.

 

[juanrga]

https://cds.cern.ch/record/2128536
Slide 13
Top zen solution is 32 cores, MCM with 2 chips a 16 cores and quad channel DDR4

About one year ago I wrote in the old thread:

[...]
For the Zen die:

8 Zen cores.
Dual DDR4 channel

The K12 core has same size and shape than Zen but is ~15% faster at same power consumption thanks to ARMv8.

Top HPC/server configurations:

* SoC size ~500mm² (Glofo 14FF) on quad-die configuration over interposer

* CPU SoC of 32 core (64 threads, octo-channel DDR4). This is four Zen xPU.
* APU SoC of 16 core (32 threads) plus 2048 SP (~250W TDP, 2x8GB HBM, quad-channel DDR4). This is two Zen xPU plus two Greenland xPU.
* GPU of 4096 SP (4x8GB HBM). This is four Greenland xPU.

The old thread was deleted by mods, but luckily I wrote the same in another forum,

http://semiaccurate.com/forums/showpost.php?p=235170&postcount=219

 

[juanrga]

One thing I find interesting is the 8 channel ram to 32 cores ratio. That would suggest only 1 memory channel on a quad core, 8 thread consumer part, which surely would be where the volume sales are. My only thought is they could construct such a part using 2 quad core blocks with half the cores disabled? I can't see consumers being very impressed with a mid to high end part relying on single channel ram.....

I expect 8-core and 6-core dies. Fusing off 4 cores from a 8-core die would be a complete disaster. Glofo dies cannot be that bad. I expect posterior 2--4core APUs using a different die than servers.

 

[juanrga]

If this new mystery chip for the AM4 socket competes with Intels 6th gen i series, I just hope AMD maintains its excellent pricing points.

Nope. Zen will be expensive by reasons stated many many times in this thread.

 

[CooLWoLF]

Expensive as compared to AMD's current offerings, sure. Expensive as related to Intel's prices? I doubt it.

 

[de5_Roy]

I was ony basing this on the previous info suggesting that Zen is built in quad core blocks. Maybe the memory controllers aren't comtained within those blocks (in which case the memory config it totally independent).

all of that is fine. but the only memory that directly serves the cpu(cores) are caches. those can be directly tied to individual clusters depending on design config. the IMC block is different.

yes, the IMC is seperated from the cpu cluster.

I don't think it makes sense from a marketplace perspective to offer a four-core processor at near Haswell speeds with only one memory channel. Many lower dual-core CPUs would have more bandwidth than that. The memory controller has to be independent from core or block count.

hehe, but it does make sense.
<insert power efficiency hyping chip designer name here> can claim superior(!!) power efficiency thanks to heavily neutered IMC. choke the memory access, choke the cpu load, save power, hype the gains - the ignorant parrot the claims without ever understanding what's going on under the heatspreader. well, snipping the IMC is only one of the many tricks though.

 

[juanrga]

Expensive as compared to AMD's current offerings, sure. Expensive as related to Intel's prices? I doubt it.

I expect octo-core Zen CPU to be priced somewhat between six-core and octo-core Broadwell-E.

 

[jdwii]

Do you guys think Amd might give 16 cores to the desktop market? The more cores its rumored to have the more i'm worried about single core performance(which matters more to everyone in this forum even if they don't think so)