• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

Ryzen: Strictly technical

Page 47 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

beginner99

Diamond Member
Jun 2, 2009
4,625
1,012
136
I wonder reading about all these issues why AMD did not create 2 dies. A true 8-core and a 4-core. The 8-core for R5, R7 and server parts and the Quad core for APUs and R3. It's not like the will offer server parts on a 4-core level granularity.
 

beginner99

Diamond Member
Jun 2, 2009
4,625
1,012
136
Raven Ridge will be a native quad-core APU. One CCX.

Actually, I wonder if Ryzen R3 will simply be Raven Ridge.
Agree. R3 will probably be like the recent Athlons. RR with disabled iGPU. Makes sense to harvest chips with broken iGPU.

And yes I knew RR would be a quad. That's why I don't get why they went with a quad-core module for Zeppelin. Why not just make a native 8-core? The will make 2 versions only anyway: 8-Core and 32-core naples. I mean HEDT + server and consumer APU both should have enough volume to offer separate "dies" or say designs as in native quad and native octo-core.

I don't know much about cost but from a project management and complexity point of view is it really cheaper to go with an universal 4-core module and create all the connecting stuff plus compromises (server vs client use) vs 2 separate designs?
 

Kromaatikse

Member
Mar 4, 2017
83
169
56
In hardware design, closing the timing constraints is a great deal of the effort involved. Each CCX is an island in which those timing constraints, at core clock speed, can be resolved without affecting the entire die - which is why Infinity Fabric runs in a different clock domain. Extending the design to an 8-core CCX would have involved redoing the timing constraints across at least the L3 cache, which could have involved serious compromises overall.
 

nix_zero

Junior Member
Mar 19, 2017
12
5
51
I wonder reading about all these issues why AMD did not create 2 dies. A true 8-core and a 4-core. The 8-core for R5, R7 and server parts and the Quad core for APUs and R3. It's not like the will offer server parts on a 4-core level granularity.
because this way they make an unified module that can be reused on everything, from the semi-custom console cpu to naples without requiring any r&d, like arm does for its core modules.
modularity looks to be one of the key "zen" arcchitectural design - you have a solid core module, an interconnect fabric that can scale up to 256 lanes, a separated memory controller module that can be switched for other kind (gddr5, hbm2 etc.), an i/o module and so on, then you just play mix and match like all those "arm cpu builders" do with arm Ip modules.
 
Last edited:
  • Like
Reactions: lightmanek

Dresdenboy

Golden Member
Jul 28, 2003
1,730
554
136
citavia.blog.de
In hardware design, closing the timing constraints is a great deal of the effort involved. Each CCX is an island in which those timing constraints, at core clock speed, can be resolved without affecting the entire die - which is why Infinity Fabric runs in a different clock domain. Extending the design to an 8-core CCX would have involved redoing the timing constraints across at least the L3 cache, which could have involved serious compromises overall.
This. With some monolithic 8C CCX the L3 latencies would be higher for every access, other latencies maybe too. With a XBar in a CCX there might be higher cycle time and power consumption (longer avg wire distances). Yields might be affected as well.
This would hit every software's performance.

I'd say they did a divide'n'conquer approach.
 

SpaceBeer

Senior member
Apr 2, 2016
305
98
101
I suppose Rave Ridge will have only one CCX, since GPU part (up to 11 Vega NCUs) will take most of APUs space. But I still think R3 will be R5 without SMT. Since R3 is announced for H2, and RR for H1/2018
 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
Agree. R3 will probably be like the recent Athlons. RR with disabled iGPU. Makes sense to harvest chips with broken iGPU.

And yes I knew RR would be a quad. That's why I don't get why they went with a quad-core module for Zeppelin. Why not just make a native 8-core? The will make 2 versions only anyway: 8-Core and 32-core naples. I mean HEDT + server and consumer APU both should have enough volume to offer separate "dies" or say designs as in native quad and native octo-core.

I don't know much about cost but from a project management and complexity point of view is it really cheaper to go with an universal 4-core module and create all the connecting stuff plus compromises (server vs client use) vs 2 separate designs?
Makes sense to waste even more die space, than with Zeppelin?
 
  • Like
Reactions: CatMerc and Drazick

SpaceBeer

Senior member
Apr 2, 2016
305
98
101
Maybe there will be no low-end CPUs without GPU (like Athlon X4). It's more likely they will leave 4-5 CUs and have A8/A9 SKUs (or whatever new names are) and sell it for $100 or more. I also doubt R3 will cost less than $100

What interests me is will they have really low-end CPUs/APUs like Celerons or A4? Especially since it was stated AM4 will also replace AM1 platform. But 1 CCX + 4 CUs will still be too big die to sell it for $30 - $40, right? So is it possible they have "small" CCX which would be one half of existing one? 2c/2t and 2c/4t + 2CUs is more than enough for basic home/office tasks
 
Last edited:

Ancalagon44

Diamond Member
Feb 17, 2010
3,275
202
106
Maybe there will be no low-end CPUs without GPU (like Athlon X4). It's more likely they will leave 4-5 CUs and have A8/A9 SKUs (or whatever new names are) and sell it for $100 or more. I also doubt R3 will cost less than $100

What interests me is will they have really low-end CPUs/APUs like Celerons or A4? Especially since it was stated AM4 will also replace AM1 platform. But 1 CCX + 4 CUs will still be too big die to sell it for $30 - $40, right? So is it possible they have "small" CCX which would be one half of existing one? 2c/2t and 2c/4t + 2CUs is more than enough for basic home/office tasks
Yeah, I don't really see why they would bother to sell a Raven Ridge with a disabled GPU. Who would buy it? If someone is looking to buy a budget CPU, there is a good chance that they would want an integrated GPU with it.

As for if they can sell it for $40 - that I'm not sure, sounds doubtful. I don't think they would want to. The AM1 had a heyday for a while, but these days, I think there is a lot of competition in that segment from things like Gigabyte Brix, raspberry pi, android boxes, etc etc. The CPU that Intel puts in their NUC will definitely not sell for $40!
 

SpaceBeer

Senior member
Apr 2, 2016
305
98
101
For NUC like PCs, they will for sure use the same Raven Ridge APUs they use for laptops (1CCX + >4 GPU CUs). But there has to (should) be replacement for Bristol/Stoney Ridge models (2 CPU + 2-4 GPU CUs) in $50 range. So is it possible they have "1/2-CCX" design?
 

Kromaatikse

Member
Mar 4, 2017
83
169
56
There is a 2-core Zen design in the pipeline somewhere, principally aimed at inexpensive laptops. I assume this would be a low-end APU, and would thus replace Stoney Ridge. It'll be a relatively late addition to the first-generation Zen lineup, though.
 

Atari2600

Golden Member
Nov 22, 2016
1,220
1,360
106
Would it pay to make a specific 2C mask?

Would the profit from a specific line of 2C+GPU pay its way relative to the loss of harvesting 4C+GPU?

I think I argued it would going from 8C to 4C as they already have most of the logic in the CCX. But the profit margin at 4C would be far above that at 2C.
 

Kromaatikse

Member
Mar 4, 2017
83
169
56
From a timing-closure point of view, reducing from 4C to 2C would be a heck of a lot easier than expanding from 4C to 8C. They probably wouldn't be looking to enhance any latencies in the process, just maintain the status quo.

I would also speculate that they'd reduce the PCIe lane count (similar to current APUs), use a single-channel IMC, drop all of the MCM support features, and use a considerably smaller iGPU than Raven Ridge to match.

This would result in not only a much smaller and cheaper die, but also a significantly lower system cost due to the fewer board traces and sockets required to complete the system. This therefore addresses a much lower-cost and higher-volume segment of the market than either Summit Ridge or Raven Ridge can, even with harvesting. That segment would otherwise be left to Stoney Ridge (using obsolete Excavator cores and Fiji-derived iGPU on equally obsolete 28nm process), or even to Intel (heaven forbid).

Of course, by reducing production costs, a decent profit margin can still be maintained that, with high volume sales, contributes materially to the bottom line.
 
  • Like
Reactions: looncraz

SpaceBeer

Senior member
Apr 2, 2016
305
98
101
Estimated PC sales is around 250 million units per year (40% desktop, 60% laptop). Systems under $400 take about 30%. I know margins are low, but if AMD can take 20-30% of those 30%, it means they could sell 15-25 million of those low-end units per year. Though I don't know if that is enough to cover development and production costs
 

Abwx

Diamond Member
Apr 2, 2011
9,113
889
126
Would the profit from a specific line of 2C+GPU pay its way relative to the loss of harvesting 4C+GPU?
If it wasnt they would have used Carrizo/Bristol Ridge harvested dies instead of designing a specific one for Stoney Ridge.

The latter is cost efficient because not only the core count is lower but the GPU is also substancially smaller with 192 SPs instead of 512.
 

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
There is a 2-core Zen design in the pipeline somewhere, principally aimed at inexpensive laptops. I assume this would be a low-end APU, and would thus replace Stoney Ridge. It'll be a relatively late addition to the first-generation Zen lineup, though.
Stoney Ridge is still on the lowest end of roadmap. If there will be dual core Zen APUs, they'll probably be Raven Ridge salvages, and it won't be the dirt cheap tier you hypothesize.
 

Ajay

Diamond Member
Jan 8, 2001
7,423
2,602
136
This. With some monolithic 8C CCX the L3 latencies would be higher for every access, other latencies maybe too. With a XBar in a CCX there might be higher cycle time and power consumption (longer avg wire distances). Yields might be affected as well.
This would hit every software's performance.

I'd say they did a divide'n'conquer approach.
I suspect there was a development cost and TTM factor as well.
 

CrazyElf

Member
May 28, 2013
88
21
81
Thinking about your power consumption figures, do you think that the 32 core could run 180W at around 3.2 or 3.3 GHz? Ryzen seems to be uber efficient at the optimal clocks.

Also, check this out:
https://videocardz.com/67594/rumor-amd-x390-and-x399-chipsets-diagrams-leaked

Cross posted from OCN.





Considering we have these and claims from Canard PC, I'd say the odds of this X399 platform being legit are very high. There seem to be 2 platforms, X399 and X390. The X399 seems to be the 2P version.
 
Last edited:

imported_jjj

Senior member
Feb 14, 2009
660
430
136
Ryzen 3 is scheduled for H2 this year.
Raven Ridge desktop for next year so no R3 is not RR based.

That X399 seems overkill in consumer.
It's 2x32cores so costs get a bit crazy.
 

beginner99

Diamond Member
Jun 2, 2009
4,625
1,012
136
Ryzen 3 is scheduled for H2 this year.
Raven Ridge desktop for next year so no R3 is not RR based.
True. But raven ridge mobile is H2 and R3 could be those mobile parts with defunct iGPU. Maybe.Why are they waiting with Raven Ridge desktop for next year if it is ready earlier?
 
Status
Not open for further replies.

ASK THE COMMUNITY