AMD RyZen CPU Architecture for 2017

[3dilettante]

So you claim 4 cores is final solution, the end of progress on the desktop.
Were stuck with that for the rest of times.

I do not see where I made that prediction.

If you run an 8 core against the workloads that so far have a quad core as their sweet spot, the 8 core is going to have certain drawbacks or fail to reach full utilization. This is the state of things now and for some time in the future, so it will impact what this larger chip will provide as a benefit over a smaller, cooler, and faster-clocked quad, and that will affect what it can be priced as.

In the case that there are workloads that can do better, the 8-core will in reality be fighting against an 8-core with SMT, with 16 threads and very low additional area or power cost.

 

[Gubbi]

So you claim 4 cores is final solution, the end of progress on the desktop.
Were stuck with that for the rest of times.

In a way it is the end of desktop.

The vast majority of the PC market is in laptops, ultrabooks and tablets/convertibles (2 in 1s). CPU SKUs for this part of the market have a fairly low core count to curb power consumption. Desktop is an every diminishing fraction of the total PC market.

IMO, there is a fair chance Intel will ditch desktop SKUs altogether and just sell rebranded Xeons as desktop SKUs.

Cheers

 

[Voxilla]

In a way it is the end of desktop.

IMO, there is a fair chance Intel will ditch desktop SKUs altogether and just sell rebranded Xeons as desktop SKUs.

Cheers

I'd rather have rebranded Xeons on the desktop than rebranded mobile parts.
But what exactly is Xeon ? Does it mean many cores and AVX-512 ?
Apparently we can now also expect a Xeon mobile, so things get pretty confusing.

 

[Voxilla]

If you run an 8 core against the workloads that so far have a quad core as their sweet spot, the 8 core is going to have certain drawbacks or fail to reach full utilization.

You may appreciate the ability of software to adapt to the number of cores at hand. Especially the divide-and-conquer algorithms and image analysis and synthesis (including rendering and physics), and deep neural networks (a strong newcomer) all lend them naturally for this. In fact this software flexibility applies to a huge number of computational intensive problems.

In practice such software spawns a number of threads, proportional to the number of cores, and thus automatically makes use of all available resources. Be it a quad core, 8 cores, with or without SMT, it all doesn’t matter. Higher core count always helps, even if that means clock frequency needs to be reduced during utilization of all cores.

So I wouldn’t necessarily use ‘software to be apparently tied to a particular core count’, as an argument to justify a certain amount of cores is enough.

 

[fellix]

IMO, there is a fair chance Intel will ditch desktop SKUs altogether and just sell rebranded Xeons as desktop SKUs.

More likely to extend the Xeon brand to the desktops, since the traditional desktop/DIY form-factor is definitely evolving as a subset of the workstation class. All of the Core SKUs should be repositioned strictly to the mobile and set-top boxes (embedded) designs and probably leave some models only for the big desktop OEMs. This will simplify marketing and product segmentation a bit.

 

[3dilettante]

You may appreciate the ability of software to adapt to the number of cores at hand. Especially the divide-and-conquer algorithms and image analysis and synthesis (including rendering and physics), and deep neural networks (a strong newcomer) all lend them naturally for this. In fact this software flexibility applies to a huge number of computational intensive problems.

There are chips with SMT that are more suited to that kind of workload than a desktop quad-core i7. And the buyers that could appreciate that can appreciate an 8-core with SMT.

So I wouldn’t necessarily use ‘software to be apparently tied to a particular core count’, as an argument to justify a certain amount of cores is enough.

Not what was claimed.

 

[Voxilla]

There are chips with SMT that are more suited to that kind of workload than a desktop quad-core i7. And the buyers that could appreciate that can appreciate an 8-core with SMT.

They can also appreciate a discrete GPU for that workload, and they do.
Which is often a better investment than your 8 core with SMT.

Not what was claimed.

Thanks for the clarification.

 

[3dilettante]

They can also appreciate a descrete GPU for that workload, and they do.
Which is often a better investment than your 8 core with SMT.

So if I am to try my hand at creating some kind of strawman of your position, people should buy an 8-core priced like an SMT-quad core because they would be better buying a GPU instead?

 

[Voxilla]

So if I am to try my hand at creating some kind of strawman of your position, people should buy an 8-core priced like an SMT-quad core because they would be better buying a GPU instead?

I don't quite understand your advanced terminology, ' ie strawman' or what you try to imply.
Probably other readers here are no native English too.

So let me do a quick lookup of that term, seems to be:
strawman - a person used as a cover for some questionable activity
That could be like somebody trying to promote here certain products for a certain company.
Funny as I would associate that term more with you, but who am I ?

Maybe you mean
straw man - a weak or sham argument set up to be easily refuted
In that case I would say: people would buy or would have bought an 8-core priced like an SMT-quad core if that would be available. Speaking for myself that resulted in me buying nothing.
With the saved money, the next thing I'll buy is a GPU, with HMB2.
I also would have bought, a quad-SMT, if there was AVX-512, mainly to write new software.
But as the market for that now has been severely reduced, my interest has diminished.

 

[Gubbi]

More likely to extend the Xeon brand to the desktops, since the traditional desktop/DIY form-factor is definitely evolving as a subset of the workstation class.

Whily technically Xeon SKUs, I'd imagine they'll be binned for higher clocking amd higher TDP . I also expect it to have multi socket capability disabled, unlocked for overclocking, relaxed constraints on RAM clocking (ie faster memory possible), shorter validation period (and thus shorter time-to-market). Launching these as Xeons SKUs would both confuse and possibly taint the Xeon brand.

I see three desktop types:
1. Workstations, served by Xeons
2. Enthusiast, served by Xeon derived SKUs with their own brand name described above.
3. Regular PCs (office/home/all-in-ones) served by mobile SKUs with relaxed TDP constraints

Cheers

 

[pjbliverpool]

Whily technically Xeon SKUs, I'd imagine they'll be binned for higher clocking amd higher TDP . I also expect it to have multi socket capability disabled, unlocked for overclocking, relaxed constraints on RAM clocking (ie faster memory possible), shorter validation period (and thus shorter time-to-market). Launching these as Xeons SKUs would both confuse and possibly taint the Xeon brand.

I see three desktop types:
1. Workstations, served by Xeons
2. Enthusiast, served by Xeon derived SKUs with their own brand name described above.
3. Regular PCs (office/home/all-in-ones) served by mobile SKUs with relaxed TDP constraints

Cheers

Isn't this already the case? The "E" range are already derived from Xeons and you could argue that everything else IS a mobile SKU with relaxed TDP.

The problem I see with this approach is that the IGP is going to be a requirement in enthusiast level CPU's moving forwards due to 1. The L4 offering enhanced performance and 2. Multi adapter making use of the IGP for gaming.

 

[fellix]

Intel already did this with LGA1366 back in the day. The platform was essentially rebranded Xeon class for the desktops due to the delays of the mainstream LGA1156.

 

[Gubbi]

Intel already did this with LGA1366 back in the day. The platform was essentially rebranded Xeon class for the desktops due to the delays of the mainstream LGA1156.

And very successful despite the insane power consumption; My main PC at home still use a I7-920 which would throttle with the default HSF with just one core active (and burn 180W when all cores are maxed).

Cheers

 

[Blazkowicz]

Conversely Intel takes what's entirely consumer chips and sells them as Xeon, it's the Xeon E3 range.

 

[Voxilla]

On the one hand I see it as probably easier for AMDs limited engineering to make a fast, 'simple' non-SMT core.
If you're already going to have 8 real cores does 16 logical cores really help in many non-HPC type situations?

I completely agree with you, thumbs up for AMD

 

[hoom]

So I guess its probably been hashed out in the AMD Execution thread or something but relevant to Zen:
Jim Keller left AMD http://wccftech.com/legendary-cpu-architect-jim-keller-left-amd-pursue-opportunities-x86-zen-track/
And they might be getting it made at TSMC instead of GlobalFoundries http://wccftech.com/amd-contracts-tsmc-produce-zen-16nm-woes-14nm-process-troubles-globalfoundries/

Late 2016 is the current ETA which is kinda sad, I was more hoping for early/mid because my current CPU is feeling its age.

 

[Deleted member 13524]

I hope that at least the part where they have the option to use TSMC in case GF fails their roadmap is true.

Zen's success is way too critical to AMD for them to be dependent on a single manufacturer.

 

[swaaye]

Isn't it a major undertaking to move a chip to another fabrication process?

 

[Alexko]

Yeah, but moving from GloFo to Samsung shouldn't involve too much work. I wonder why TSMC is popping up in rumors instead.

 

[tunafish]

The GloFo/Samsung processes are supposed to be identical, and the same design should work in both. They still would have to do the masks separately for each, simply because you just don't ship finished masks that far.

I wonder how the WSA allows AMD to order chips in the situation where Samsung has idle capacity and GloFo can't get good yields/is producing at max rate.