AMD RyZen CPU Architecture for 2017

A Chinese website has roadmaps containing Pinnacle Ridge and Raven Ridge.

X3buZW5.png


TntLlXz.png


There's no sign of the rumored DDR4 + HBM2 Raven Ridge.
I think someone just thought all 1st gen Zen APUs would be Raven Ridge, the original HBM(2)-APU-leaks were about the HPC-APU which was apparently MCM to boot (not just MCM as in APU+HBM but as in Zeppelin + GPU + HBM
 
On a somewhat unrelated note, the R7 1700 seems to be really, really power efficient:

Ryzen needs a huge voltage boost to go past 3.3GHz. At 3.3GHz it's nearly twice as power efficient as it is at 3.9GHz.

At peak clock speed it probably has somewhat worse perf/W than Broadwell, and it's quite behind the single threaded performance scaling peak. But at lower performance points this changes dramatically, and based on what we've seen so far I could see the 32 core Naples MCM being very competitive with the high core Broadwell Xeons, which can't hit especially high clock speeds even under turbo. And it could be a complete blowout vs the current Xeon D lineup. Then there's laptops where it hits the same perf/W sweetspots for the most part, and will probably have access to better GPU perf and perf/W if they manage to have HBM memory at a reasonable price.

Their current R7 1800X flagship is perhaps the least interesting thing they could be doing with the uarch, or at least would be if it weren't so competitively priced against Intel's HEDT alternatives.

This is the outcome of a uarch that sacrifices peak single threaded performance for better efficiency, and consequently better density, getting comparable perf/area on an inferior process. This may well force Intel to invest in a new uarch family that makes the same tradeoff in order to strike back.
 
Yes. The 1800X enjoys a comfortable lead over the i7-7700K in non-gaming workloads, but not a huge one. Intel is about to introduce 6-core processors with Coffee Lake and I would expect those to close that gap, and perhaps even overcome Ryzen.
AM4 is basically equivalent to the S115x platform. Next year, intel will still only offer 2 cores less than AMD is offering today. And Coffeelake isn't supposed to have any major improvements over Skylake and Kabylake, isn't it? Basically just another refinement on the energy efficiency and clock speed side (the latter being immediately consumed by the bump to 6 cores), right?
There will probably be Skylake-E SKUs with 8, 10, and perhaps 12 cores as well.

So I think AMD needs 12 cores to maintain Ryzen's edge, which shouldn't be difficult on a 14nm process that is already pretty mature now, let alone next year, especially when Ryzen is apparently something like 215mm². The other option would be for Pinnacle Ridge to have at least 10% higher IPC compared to Summit Ridge, slightly higher clock speeds, and fewer issues with its interconnect and memory subsystems. This is possible, but adding more cores would be easier, and not incompatible with any of this.

Alternatively, AMD could make relatively affordable—$1000?—versions of Naples with 16 cores, essentially offering something comparable to Intel's HEDT platform, but with much more raw power.
There is actually also Snowy Owl and its plattform which should have a smaller socket than the massive 4000+ pin LGA socket of Naples, if it is not the rumored HPC-APU with an MCM made out of 2 Zeppelin dies and a GPU. I would assume that there is also an MCM with just two Zeppelin dies. This should result in a quad channel memory interface and up to 64 PCIe lanes. It would be tailored to workstations and entry level servers and would be adaptable to the HEDT market in the same way as intel does with the LGA 2011/2066 platforms. That would be the natural opponent in my opinion if AMD decides to compete in the HEDT segment.

Both, intel and AMD offerings can be grouped in 3 different categories (intel may have more than a single socket for one category):
(i) 2 channel memory interface:
AMD Ryzen up to 8 cores
now Core i# with up to 4 cores, next year up to 6 cores with Coffeelake
(ii) 4 channel memory interface:
possibly Snowy Owl with up to 16 cores
now: HEDT Core i7 (Broadwell-E) up to 8/10 cores, Xeon up to 24 cores (Broadwell-EP/EX)
later this year: HEDT core i7 (Skylake-X) up to 10/12? cores, Xeon (Skylake-W) no idea
(iii) 8/6 channel memory interface
Naples with up to 32 cores and 8 channels
intel now: only Xeon Phi , later: Skylake-EP/EX on LGA 3647 with 6 memory channels and up to 28/32 cores
 
Last edited:
One thing I don't understand is why AMD artificially cap the 4 core CPUs to 3.7 instead of letting them get to 4Ghz like their bigger brothers. That will hurt performance out of the box and maybe make the chip look worse in reviews that only measure that.
 
The real champion is the 1600. You are getting a 6800K, in both single and multi threat for half the price.

Next year will be interesting since its Intel process change so no IPC gain just lower TDP and more core at the top, maybe SMT in the middle to compete. But Ryzen will have an arq. update and if AMD manage to improve 10% of it...

Intel will have to changes its philosophy because if they don't they will have to wait 5 years(!!!) to be able to place an arq. update design to counter ryzen.
 
Intel will have to changes its philosophy because if they don't they will have to wait 5 years(!!!) to be able to place an arq. update design to counter ryzen.
Why 5 years?
AFAIK Icelake is a brand new architecture for 10nm to arrive around 2019?
 
We have confirmation from AMD that there are no silly games going to be played with Ryzen 5. The six-core parts will be a strict 3+3 combination, while the four-core parts will use 2+2. This will be true across all CPUs, ensuring a consistent performance throughout.
http://www.anandtech.com/show/11202/amd-announces-ryzen-5-april-11th

This means that there will be two active CCX in the four core parts. Split L3 cache remains (one per 2 cores). Total L3 size is doubled (compared to single CCX), but split cache problems remain.
 
http://www.anandtech.com/show/11202/amd-announces-ryzen-5-april-11th

This means that there will be two active CCX in the four core parts. Split L3 cache remains (one per 2 cores). Total L3 size is doubled (compared to single CCX), but split cache problems remain.
It's clear you don't mix configurations within the same model. But they could choose different ones for different models. First candidate could be the 1400, which comes with only 8 MB L3. And Ryzen 3 could also feature different configs.
 
Tom's Harware's brand new gaming review:
http://www.tomshardware.com/reviews/amd-ryzen-vs-intel-kaby-lake-gaming,4977.html

TLDR
Gaming: All Ryzen models at top (big margin) in Deus Ex (1080p & 1440p). All Ryzen models slightly ahead of Intel in Shadow of Mordor at 1440p. Intel wins all other games, some with huge margin.
Synthetic: Ryzen wins 3DMark physics tests and is very close to 6900K in DX12 draw call test (both have 33% lead over 7700K).
Conclusion: Intel is better in most games. i7 7700K is currently the best pure gaming CPU.
 
Last edited:
It will be interesting to see what happens to AoS when they release their Ryzen patch, as the 6900K is well ahead of the 7700K, which is well ahead of the 7600K, which is ahead of the R7's. With the R7's being much more closely bunched.
 
Because Intel bumps core count every year?
Fix for some low hanging performance bottlenecks & a clock bump seems like a reasonable expectation to me.

Not every year, but this year! And who knows about next year, now that there's real competition?

AM4 is basically equivalent to the S115x platform. Next year, intel will still only offer 2 cores less than AMD is offering today. And Coffeelake isn't supposed to have any major improvements over Skylake and Kabylake, isn't it? Basically just another refinement on the energy efficiency and clock speed side (the latter being immediately consumed by the bump to 6 cores), right?
There is actually also Snowy Owl and its plattform which should have a smaller socket than the massive 4000+ pin LGA socket of Naples, if it is not the rumored HPC-APU with an MCM made out of 2 Zeppelin dies and a GPU. I would assume that there is also an MCM with just two Zeppelin dies. This should result in a quad channel memory interface and up to 64 PCIe lanes. It would be tailored to workstations and entry level servers and would be adaptable to the HEDT market in the same way as intel does with the LGA 2011/2066 platforms. That would be the natural opponent in my opinion if AMD decides to compete in the HEDT segment.

Both, intel and AMD offerings can be grouped in 3 different categories (intel may have more than a single socket for one category):
(i) 2 channel memory interface:
AMD Ryzen up to 8 cores
now Core i# with up to 4 cores, next year up to 6 cores with Coffeelake
(ii) 4 channel memory interface:
possibly Snowy Owl with up to 16 cores
now: HEDT Core i7 (Broadwell-E) up to 8/10 cores, Xeon up to 24 cores (Broadwell-EP/EX)
later this year: HEDT core i7 (Skylake-X) up to 10/12? cores, Xeon (Skylake-W) no idea
(iii) 8/6 channel memory interface
Naples with up to 32 cores and 8 channels
intel now: only Xeon Phi , later: Skylake-EP/EX on LGA 3647 with 6 memory channels and up to 28/32 cores
Ah! I did not know about Snowy Owl. That does seem like a sound way to go. By the time Coffee Lake is out and the dust has settled, top Ryzen 7 SKUs will probably sell for somewhere around $350, leaving a nice $500–$1000 space for Snowy Owl, which could be called Ryzen 9, or Ryzen X or whatever. A couple of 12- and a couple of 16-core SKUs would fill that space quite nicely.

I still think that core count ought to go up pretty quickly, if not with Pinnacle Ridge, then with the next generation. You'd have 12 core CPUs for the mainstream platform—with faster DDR4? just more cache? a third channel? HBM?—and up to 24 cores on the HEDT platform. On 10nm—assuming that would be ready by 2019, which sounds reasonable—you'd get a perfectly manageable die size.
 
Tom's Harware's brand new gaming review:
http://www.tomshardware.com/reviews/amd-ryzen-vs-intel-kaby-lake-gaming,4977.html

TLDR
Gaming: All Ryzen models at top (big margin) in Deus Ex (1080p & 1440p). All Ryzen models slightly ahead of Intel in Shadow of Mordor at 1440p. Intel wins all other games, some with huge margin.
Synthetic: Ryzen wins 3DMark physics tests and is very close to 6900K in DX12 draw call test (both have 33% lead over 7700K).
Conclusion: Intel is better in most games. i7 7700K is currently the best pure gaming CPU.

Deus Ex is very interesting. I can see how a multi-threaded game might like Ryzen, but the latter is way ahead of even the i7-6900K, which is curious. There must be some microarchitectural feature in Ryzen that Deus Ex really likes.
 
Deus Ex is very interesting. I can see how a multi-threaded game might like Ryzen, but the latter is way ahead of even the i7-6900K, which is curious. There must be some microarchitectural feature in Ryzen that Deus Ex really likes.
Very interesting indeed. Even Ryzen 1700 is 12% faster than 6900K. Can't be explained by clock rate advantage. Ryzen has 2x larger L2 caches, but Intel's L3 is bigger and faster than Ryzen's, so I doubt it matters much.
 
Why 5 years?
AFAIK Icelake is a brand new architecture for 10nm to arrive around 2019?
That are rumors, Intels philosophy is process -> Arq -> Optimization. SK is Opt. So cannon lake is process(same arq, different process) and coffe lake is Arq and the next will be Opt. So Intel will need to wait until Arq time again and doing to math that would be 5 years from now.

If we take into consideration the time that takes to make a CPU its pretty clear that CL and CL are already type out and everything is decided, Intel could tweak here and there but can't do much to them. So Until Intel magically(or prophetically?) wanted to make a huge jump with cofee lake years ago they will need to change the philosophy and launch a new Arq before is "schedule" right now.
 
Why do you mention quad-core APUs? Ryzen goes up to $500.
That'd be because I failed to refer back to original quote & thought the core count referred specifically to the APU :oops:
But if AMD needs higher core counts next year bear in mind Ryzen/CCX architecture is already intended to scale to 16 & 32 core chips even if not (currently) planned for desktop.
 
Back
Top