AMD CPUs, SoC Rumors and Speculations Temp. thread 2

Page 18 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.


My understanding (can't remember sources) was that the new Zen FX and a new Carrizo based desktop APU (as it's only mobile atm) will share socket AM4, which will support DDR4 memory.

The FX chip is rumoured to be (up to) 8 core / 16 thread (thanks to HT) whilst the Carrizo APU will cover the low end. It makes a lot of sense for AMD to release an integrated socket as it gives a good upgrade path not currently available with split sockets.

Note, if desktop carrizo is just the laptop design at higher tdp, then it would be a soc- which would suggest the zen fx cpu would also be a soc design. That said, without a GPU (the real die hog) and on 14nm, that would still leave loads of room for cores, and I think for desktop 16 threads is kinda overkill already so this wouldn't bother me.
 


Papermaster mentioned several SoCs are not APUs at FAD 15. The official roadmap also explicitly mentioned DDR4 compatibility for AM4 platfform.
 

in the latest official roadmaps.
http://www.anandtech.com/show/9231/amds-20162017-x86-roadmap-zen-is-in

none afaik. but expect the latest consumer tech on the high end parts, e.g. m.2, that interface that takes up two sata ports forgot what it was..), usb 3.1, 3.0, type c, freesync and so on.

some of the processors with zen cores will be full socs.

14nm is supposed to open up plenty of die area without compromising cpu performance. it will depend on how zen based processors are designed.

 
@de5_Roy this looks like some really promising information then.

downside to a unified socket for apu and cpu platform will be non dedicated motherboards. all will have hdmi and DP out when only apu's could use them and id rather see the space on the rear IO for maybe 10 gbe or dual gbe ports more usb you know usefull stuff.

maybe we can see something like the old crosshair V Formula without any rear IO.

as far as 14nm FF+ design I think this speaks for all cpu's, yes you can fit more things on it if we maintain the size of the cpu cores and mobo componets such as sata controller etc... but what if these extras were moved off die and the cores were made larger? wouldn't we see more performance as we know the pipeline to southbridge isn't a weakpoint so there isn't an advantage performance wise to move it on die. why not just pump up that cpu and only move other componets on to make an SoC when they would be bottlenecking the system if not on the cpu die?

I understand how in a lappy this can reduce power comsumption and theory can increase sata perf pci perf etc... but we already max out the sata bus bandwith with our old designs why move that old bus on die? we will get no performance and lose valuable cpu core space on die. 8 core 16 thread will be stupid OP but I have a feeling the single core perf will leave much to be desired. we could use this space to make cores physically larger and increase single core perf. resulting in better multiplier ratios etc...

I just don't see how moving to an SoC could possibly increase performance, only hamper the cpu possibility.
 


On a CPU (no iGPU) the lrgest parts go like this:

Cache -> Cores -> Memory Controller/interlink/everything else.

Removing the other components would not only make no difference to the overall CPU performance but would actually drop performance of those parts. The reason why the memory controller was moved on die was due to performance limitations of it being on the Northbridge. Intel gained almost 4x the memory bandwidth on Nahelem compared to Yorksfield.

And no, not all will have HDMI/DP/DVI. Some don't. Either way a HDMI and DP doesn't take much space and 10Gbe is not even out as a onboard NIC yet, otherwise Intel would have one but not even they do yet.

Again a SoC is the best possible performance and the parts for it are typically not much of the CPU at all:

350x700px-LL-ff3b9b78_Core_I7_LGA_2011_Diesm.jpeg


That is a Haswell-E die shot but the top is where the integrated parts are which is also part of a lot of other components. Even removing that would not give the cores much room to grow and again putting them on a Southbridge would make them slower overall.
 
On a CPU (no iGPU) the lrgest parts go like this:

Cache -> Cores -> Memory Controller/interlink/everything else.

Removing the other components would not only make no difference to the overall CPU performance but would actually drop performance of those parts. The reason why the memory controller was moved on die was due to performance limitations of it being on the Northbridge. Intel gained almost 4x the memory bandwidth on Nahelem compared to Yorksfield.

And no, not all will have HDMI/DP/DVI. Some don't. Either way a HDMI and DP doesn't take much space and 10Gbe is not even out as a onboard NIC yet, otherwise Intel would have one but not even they do yet.

Again a SoC is the best possible performance and the parts for it are typically not much of the CPU at all:

350x700px-LL-ff3b9b78_Core_I7_LGA_2011_Diesm.jpeg


That is a Haswell-E die shot but the top is where the integrated parts are which is also part of a lot of other components. Even removing that would not give the cores much room to grow and again putting them on a Southbridge would make them slower overall.
Click to expand...

great info here I was assuming the memory controller cashe etc would be left on die as I knew they were seen to make big improvements, but sata? regardless I do see your point you would gain roughly 10-5% more die space tops to remove this extra componets such as sata and PCIe and that I think would not be a very helpful tradeoff, but what do I know im just a guy theorizing.

AMD how about you take a hint here 10 Gbe onboard nic for your top end desktops..... I would go for that. id pay $50 more for that maybe $100 10gbe is the future!
 
The problem with 10Gbe is that AMD does not design NIC cards. Intel, however, does and are one of the leads in the specifications and implementations of new CAT standards. 10Gbe is now available in a add-on card but not integrated on a motherboard yet. Then there is the router, no router currently has 10Gbe. Some really high end firewall/switch combos do which are near the $1K mark. Then the modem. Most modems probably run a 1Gbps NIC.

It will take probably another 5 or so years before 10Gbe becomes a more common standard.
 
The advantages of SoC are cheaper to manufacture, higher performance (latency), and less power consumption (which can add to performance as well). Not only AMD will not eliminate stuff from SoC, but the overall industry tendency is to add more and more stuff to the SoC. As Papermaster disclosed at FAD 15, the whole future technology is modular with interchangeable modules in a SoC. Papermaster gave several examples of SoCs and the different elements integrated on the SoC. Their whole strategy is based in migrating away from a falling PCs market towards data-center, semicustom, and other markets. AMD has future plans to add even SSD to the SoC. But this is not exclusive to AMD. Future motherboards will be very simple boards with I/O connectors and little more.
 

two words that rule tech industry.. or so to speak: saves moniez.
the useful stuff you mentioned - have cost added for each additional port, layout etc. which forces the end price creep up. if zen-processors turn out to be disappointing (very subjectively speaking 😛) the mobos will look like waste of money.
 


given this theory/ plan to make future pc's completely SoC based which I understand the reasoning for, I have to give a lot of doubt to amd's current plan then. they are not even trying to get into embedded market that I know of and their server cpu market took a huge hit when they stoped making opterons. the only thing they have been advancing in the last years is APU and graphics departments.

zen/ future plans seem to support what your saying.

what I get from this is a plan for the future having simple motherboards with 50 or so APU SoC choices from each company and you pick the APU for the performance you want. meaning many motherboards will go from cpu gpu to perhaps dual APU setups for a crossfire setup? on die hbm could fuel as shared memory and speed us on quite well. hbm 2 supports upto 32 gb I think? modern pc's still only need 4 gb to run everyday tasks so that's rather future proof in my opinion.

I still cant imagine internal Die storage, unless internet/ datasenters are so widespread and cheap that local storage is irrelevant. I am uncomfortable with the idea of needing internet to do anything. internet is already stupid expensive for the piss poor speeds they give you. im paying something like $50 month for 4 mb/s rated ( it acutally gets .4 mb/s downloading steam games). if I have to really on that paid connection for simply playing back a song or opening a file it will be a sad day indeed.

and im looking into 10 gbe switch for my home plex server but raid hdd's only hit 350 mb/s anyway maybe that's really overkill I do have a spare dual 1 gbe intel card. can setup link sharing or what ever that's called to get 200 mb/s speed :/

home server is Kabini based :) 8 tb storage + 8 gb ram quad core all under $500
 


Why would you need a 10GBe router for a home or small office? A switch would suffice if you want that speed locally. Otherwise yes, most implementations would probably use the 4x 1GBe option instead.
 


its not if I need, but rather can I get it cheap. as all computer fans know more speed is always welcomed at a price. how do you think NVidia sold titan and titan x? I could transfer a 20 gb game file from one pc to another in 30 seconds with 10 gbe XD who could say no to that?
 


does it matter? Linux users comprise of 1.3% of all computers while windows 7,8,8.1,10 comprise of about 80%.
open closed as long as its on windows and works is all anyone really cares about. the people who actually want this are less than .1% of computer users as even most Linux users aren't going to change vulkan.

as a company struggling to get by I would hope they aren't wasting time with Linux. unless they really think steamos with vulkan would be decent sales and then closed source still would be just fine.

 

Probably the idea of making it open source is to let the community do the development for you. Even if Linux is 1.3% market share, that stil is millions of Linux users, and you only need like 10 or 20 actively developing it. By doing this, they save a lot of resources, and very few people are needed to maintain it from inside the company.

EDIT: i'm bad at formatting quotes 😛
 


Correcting a misleading FUDzilla heading matters, and Linux users are the majority of the whole market. That 1.3% is only for desktop (linux dominates the ~90% of HPC for instance) and even for desktops that 1% is incorrect because it is based in inaccurate webcounts. I already explained all this on the old thread and give realistic statistics for different OSes.
 


While he was better than who they have had for a while his biggest ideas were based on tech from DEC-Alpha.
 


That is interesting in a lot of ways. Some sadder than others.

According to the link he is leaving to pursue other stuff in his career. What that is, I have no idea. One thing is for sure, he left in the worst possible moment for AMD. Even if the design for Zen is done and well into sampling stage, he'll miss the optimize phase (Zen+) and making the ARM SoC a reality. I wonder if the team he leaves behind is as good.

Also, I'm curious to know what he is going to do now.

Cheers!
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom