-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
AMD Ryzen Thread: Affordable Core Act
- Thread starter ·feist·
- Start date
It kind of makes sense and was expected though.
It's more likely for a single core to be defective in a ccx rather than a half/whole ccx itself to be defective.
I think the assumption was that these would be single-CCX designs, not dual-CCX with an entire CCX disabled.
I guess it must be more cost effective for them to just have one CPU design and keep binning it lower.
Yup, bring on the 5 and 7 cores
brain_stew
Member
Welp, there goes a lot of my hype for the R5 range.
A single CCX unlocked R5 with SMT enabled that can clock to 4ghz for $170 sounded like a brilliant budget gaming CPU.
I worry about how inconsistent game performance could be in a modern game that utilises 8 threads which is going to require constant communication between CCX units.
Yeah, I also thought they'd be a single CCX.
2+2 probably makes it a much less enticing options for games (and any parallel applications which need low-latency synchronization).
Coulomb_Barrier
Member
Lol you took the bait.
Yes of course, but reading the posts here, apparently they're all going to be junk anyway
Coulomb_Barrier
Member
Wait a minute, some of you (Paragon was one) was saying that the Windows scheduler is not the problem. And now you are also saying the CPU Complex is an issue. So which is it - if the way Windows allocates the cores (during gaming mainly) was optimal, then there wouldn't be such a penalty with the CCX design and communication between cores, so the CPU Complex design wouldn't be inherently bad.
These problems can be mitigated if the way Windows recognises Ryzen CPUs is addressed via the Windows scheduler. Windows already has an optimised design for Ryzen's CPU Complex design in its NUMA support (in this case treating a 1700 as 2 4-core CPUs) as opposed to Ryzen being treated as an SMP CPU.
I've seen examples where games are running perfectly on a dual-Intel Xeon set-up.
That said, AMD will need to work with software devs to optimize for Ryzen and that again will bring up performance.
hanspampel
Banned
Raven ridge (the apu 4 core zen + gpu) will be h2 2017 for mobile
Raven ridge for desktops will be 2018
These will be for the more energy efficient APUs like Ryzen Mobile.
PC Perspective have a video which goes into this in-depth: https://www.youtube.com/watch?v=6laL-_hiAK0
What people had been suggesting is that Windows was:
a) Unaware of AMD's SMT implementation and treating all cores as physical cores
b) Thinking that each core had 16MB of L3 cache each.
PC Perspective's testing has shown that Windows is aware of the difference, and the author of the tool people were using to check the available cache has said that it's an issue with their application. I believe it may even have been an outdated version people were using.
If you start an application with 8 threads, it will send them to all the physical cores first, before putting work on the SMT cores.
That's how things are supposed to work.
In the majority of cases, you would have reduced performance if you put all 8 threads of an application on a single CCX (four physical cores, eight threads) since there is less computational ability there.
It's possible that some workloads would prefer to be on a single CCX, but unlikely.
It's not the scheduler's job to prevent applications from accessing all 8 CPU cores.
Doing so would greatly harm the performance in most applications that can use >4 threads.
The CCX structure is a fact of life for the Ryzen CPUs.
It's not a bug that the OS has to "fix".
What you do want to happen is, if an application only uses four threads, you want the system to try and keep them all on a single CCX of an R7 CPU instead of spreading them across both CCXes.
The fewer cores you have per CCX, the more likely it is that an application is going to have threads spanning across both CCXes.
So with a 6-core CPU you only get 3 threads before they spill over to the other CCX, and with a 4-core CPU you only get 2 threads per CCX.
Therefore the lower core-count Ryzen CPUs are going to be more affected by the performance hit of cross-CCX communication than the 8 core R7 CPUs.
I'm sure there is some performance to be gained by developers optimizing specifically for Ryzen processors, trying to keep related threads on the same CCX where possible instead of allowing them to be assigned at random, but that is per-application - not something the OS can do.
And the issue with Windows 7 vs Windows 10 appears to be that Windows 7 was not parking cores, and would (sometimes?) not be taking advantage of SMT, only assigning 8 threads.
So in applications where disabling SMT would improve performance, Windows 7 outperformed Windows 10 with SMT enabled. You could probably achieve the same results by setting the affinity of that application to only use physical cores, instead of completely disabling SMT.
It's pure speculation on my part, but I would guess that, in applications where disabling SMT improves performance, the issue is that the extra threads being spawned are causing more communication between CCXes to occur, which is why performance is reduced.
Setting Windows 10 to the high performance power preset disables core parking, putting it on-par with Windows 7.
With one or two rare exceptions, performance is better overall on Windows 10 in the high performance power preset with SMT enabled than not.
That said, you're still getting a 6-core 12-thread CPU for less than the cost of a 4c/4t CPU from Intel, and a 4-core 8-thread CPU for less than the cost of a 2c/4t CPU from Intel with the R5 CPUs.
Even with the potential performance hit from fewer cores per CCX, Intel's CPUs are looking less and less attractive every day.
I can see these comparing a lot more favorably to the i3-7350K and i5-7600K than the R7 CPUs did to the 7700K.
I don't know that I want to wait until the end of the year to upgrade or build a new PC, and even though gaming is the impetus for this upgrade, I'm still leaning towards a 1700 instead of a 7700K because I do other work on the machine that will benefit from >4 cores. I'd like to have a new system by the time Prey is out, and Intel probably won't have anything new until much later in the year. I definitely don't want to buy into X99 now.
So it is actually super easy to simulate how a 4core and 6core Zen will perform, you can just disable x cores on the 1700/1800x
Question is if it will also be 50/50 disabled on the 4 core because that would be horrible.
or is it just random what is disabled an you can be lucky and get one where the 4 cores is on the same CCX.
Question is if it will also be 50/50 disabled on the 4 core because that would be horrible.
or is it just random what is disabled an you can be lucky and get one where the 4 cores is on the same CCX.
If you have N threads to schedule, then if the graph describing low-latency communication between them can not be partitioned into two sub-graphs of exactly N/2 vertices each which are not connected by any low-latency communication then the scheduler can not make any mapping decision which will not be hampered by communication latency to some degree. With current games, this scenario gets more likely with smaller values of N, so a 2+2 setup would, even in an ideal schedule, be more problematic in this regard than a 4+4 setup.
dr_rus
Member
I think that we'll have such SKUs eventually. No reason not to considering that whatever failing 8 core chips they'll have will most likely be suitable for a 3+3 hex config and with a quad core it's more interesting to produce a cheaper smaller 4 core chip in the long run.
But yeah, for the initial batch this is a downer.
If the Windows scheduler were sane this wouldn't happen: https://www.youtube.com/watch?v=JbryPYcnscA
Whenever performance can be improved by manually setting CPU affinity the scheduler isn't doing its job.
Whenever performance can be improved by manually setting CPU affinity the scheduler isn't doing its job.
Coulomb_Barrier
Member
Yes I understand that, 2+2 CPU complex doesn't sound great but with the hex core 1600X (6C/12T) the bonus is it has 8MB of L3 cache per 3 cores instead of 4 which would be a small perf gain per core to offset latency penalties somewhat.
Well exactly, even if performance stays as is, the perf per dollar on these Ryzens is going to be king. The R5's could very well offer 85% of the performance of a 7700K but for almost half the cost.
metareferential
Member
For a new build sure; I'd like to see how they perform against past i5/i7 (meaning that incentive to upgrade might be somewhat limited giving the price of a top kaby Lake i5 today).
I'm tired of waiting xD
hanspampel
Banned
and the Ryzen 3 Desktop models
+ likely future Ryzen 5 4 core Desktop models, when the yields get better
i'm not talking about that.
just what other CPUs will be single CCX and not dual CCX with deactivated cores
to my understanding, Ryzen 3 will be single CCX and Ryzen 5 4 core, too (in the future)
but right now, there seems to be just one die, they produce.
the single CCX die is not yet ready for shipping
The Ryzen 3 models won't have SMT though (which appears to be more efficient than Intel's HT so far whenever it works as intended).
Ideally that should never happen, but splitting a 2-thread application over two CCXes is of minor consequence since that is easily fixed by manually setting the affinity or using a tool like Process Lasso to manage things automatically if the application's total threads can fit on a single CCX.
Hopefully Windows will be updated to be better about that, but it shouldn't be a frequent occurrence.
It highlights the underlying issue with Ryzen's dual-CCX architecture though.
If your application uses more threads than a CCX has cores (4 for R7, 3 for R5 1600/X, 2 for R5 1500X) then the work is going to be split across CCXes and suffer a performance penalty.
Until now, it has not really mattered how games or other applications order their workload.
With Ryzen, the order that threads are assigned in suddenly matters, and for the best performance developers have to specifically code their applications to keep similar tasks grouped on a single CCX where possible.
The fewer cores per CCX you have, the more likely it is that you will run into applications with work being split across them.
I want to know what Intel's response is going to be, it seems that Ryzen and the 10nm delay put them between a rock and a hard place this year.
If they don't modify the prices with coffelake, they have to resign some sales to ryzen.
If they do, what happens to the pentium/celeron, i3, i5, product lines? Many products in a small price range. They maintain marketshare but lose revenue.
I think they will put all their marketing muscle on promoting their high clocks and IPC and see if it works this year waiting for the arrival of the 10nm process. But coffelake would need to be a real improvement over kabylake ofc..........
If they don't modify the prices with coffelake, they have to resign some sales to ryzen.
If they do, what happens to the pentium/celeron, i3, i5, product lines? Many products in a small price range. They maintain marketshare but lose revenue.
I think they will put all their marketing muscle on promoting their high clocks and IPC and see if it works this year waiting for the arrival of the 10nm process. But coffelake would need to be a real improvement over kabylake ofc..........
Beautiful Ninja
Member
Well the major rumor surrounding Coffee Lake now is that it's bringing 6-cores to the mainstream. If prices remain as they are now, it'll be a 6c/12t Core i7 competing with the 8c/16t Ryzen, but with Intel's superior clock speed and IPC, the difference in productivity would shrink considerably, while likely increasing the lead on game performance even further.
Intel's felt the need for more cores/threads coming for a while, we saw this with the Pentium gaining Hyper Threading this generation with Kaby Lake, 2c/2t wasn't cutting the mustard anymore. If the upcoming Coffee Lake SKU's end up being a 6c/12t i7, 6c/6c or 4c/8t i5, 4c/8t or 4c/4t i3 and 2c/4t Pentium, that would compete nicely across Ryzen in all types of workloads.
As of right now, the i3 is stuck in a peculiar place where it basically has no reason to exist, the Pentium is half the price and offers most of the performance in anything outside of things like applications that use AVX.
Problems is, an (hypothetical 4c/4t) i3 not having HT in this day and age feels cheap with Ryzen having it for 169. Same with a 6c/6t. The lack of HT in some lines was artificial and it looks like it's coming back to bite Intel in the ass.
Well, only time will tell what is ahead.
Well, only time will tell what is ahead.
OK RyzenGAF, I need a hand (again)
AMD's advice on RAM is as follows:
Add to this the fact that RAM + Ryzen currently means lower clock rates than that anyway, and I'm confused on what to do regarding a motherboard and memory?
And is there any reason why AMD suggest the 16GB G.Skill (2x8) DDR4-3200 specifically? Why G.Skill?
To top it off, do I really need 3200Mhz RAM? I'm gonna be working in 3D and game dev applications with some video editing on the side.
Every other aspect of my build has come together nicely, but I'm honestly stumped on the Motherboard and RAM!
-EDIT- Might as well share the product page for the Asrock board. How does this look? The bad reviews for the Asus boards so far have put me off theirs, which begs another question - even if Asus' shitty UEFI updates are a problem now, will buying the board and coping till a fix work out? Will there be fixes?: http://www.asrock.com/mb/AMD/AB350 Pro4/#Memory
AMD's advice on RAM is as follows:
"The AMD Ryzen™ processor does not offer memory dividers for DDR4-3000 or DDR4-3400. Users shooting for higher memory clocks should aim for 3200 or 3500 MT/s."
However, Asrock's AB350 Pro4 (the mobo I as looking at) Qualified Vendors List doesn't mention the 3200Mhz G-Skill Ripjaw V RAM AMD recommends in their Ryzen build featured on their blog - nor any DDR4-3200 aside from one Patriot product. Add to this the fact that RAM + Ryzen currently means lower clock rates than that anyway, and I'm confused on what to do regarding a motherboard and memory?
And is there any reason why AMD suggest the 16GB G.Skill (2x8) DDR4-3200 specifically? Why G.Skill?
To top it off, do I really need 3200Mhz RAM? I'm gonna be working in 3D and game dev applications with some video editing on the side.
Every other aspect of my build has come together nicely, but I'm honestly stumped on the Motherboard and RAM!
-EDIT- Might as well share the product page for the Asrock board. How does this look? The bad reviews for the Asus boards so far have put me off theirs, which begs another question - even if Asus' shitty UEFI updates are a problem now, will buying the board and coping till a fix work out? Will there be fixes?: http://www.asrock.com/mb/AMD/AB350 Pro4/#Memory
Übermatik;232207049 said:
Seems like Ryzen's data fabric loves high speed ram. From one reddit comment: "Running 4GHz and 3200-14-14-14-34 here. Pretty much all my gameplay is GPU bound on this 480, even at low settings. 50% faster data fabric vs stock is no joke." seems like everyone that is hitting 3200mhz is using G Skill Tridents atm. If it's not on the QVL, it just means it hasn't been tested on that setup. So there is some risk, but it should work... I'd think.
Personally, wondering if we'll eventually get support for ram speeds over that even. Seems like the faster the ram the better this architecture performs.
As a 480 owner these are good impressions, thanks!
Surely there's no difference between manufacturer though? Whether it's G.Skill or Corsair or whoever, providing the specs are the same?
Anyway, I'll ask around for some motherboard impressions too.
Übermatik;232210799 said:
Yeah it seems regardless of manufacturer you want the underlying memory to be Samsung B-dies
From Techpowerup:
Great news - that Corsair model is exactly the one I was looking at, except in white. Thanks kotodama!
No high end laptop manufacturer is going to dump Intel for Ryzen.
Nostremitus
Member
In a 2x2 configuration, would the active cores have access to the full cache allocation for each module?
I have CL14 2x16 32GB G.Skill Trident 3200MHz. Unlike the AM4 motherboards which have an external clock generator for CPU and RAM speed tweaking, my X370 does not. It refuses to post when I set the 3200MHz XMP profile, instead defaulting to 2133MHz at loose CL15 timings.
I briefly ran it at the speed, but have since gone to 2667MHz CL14, which makes for noticeably more responsive performance.
All board manufacturers will be getting an update from AMD this May to support higher than 3200MHz DDR4 w/o needing refclk adjustments.
Note, the motherboard makers can opt to only include this update for select models in their lineup. Either way, shame this won't be available before the Ryzen 5 launch. It would have been nice to see it included with reviews since most sites will not re-bench later on.
The highlights from AMD's "Tips for Building a Better AMD Ryzen System" posted here: http://www.neogaf.com/forum/showpost.php?p=232080327&postcount=1810
Under the "Memory Matters" section:
Some months ago, it was speculated the 6-cores would have access to the full cache, while the memory for 4-core parts would be halved.
Seems that isn't the case:
From HardwareCanucks' "AMD RYZEN 5 - Explained! (1600X, 1600, 1500X & 1400!)" video.
Nostremitus
Member
·feist·;232234041 said:
Well that sucks, would have been cool if the unused cache mirrored the data on the other module for faster access in the event it was needed instead of having to pull across the fabric at the moment it's needed. But that would probably open a whole new can of worms.
Or, if it could have functioned as supplementary cache for the logical cores.
Oh cool, so the 1500x will have access to the full L3?
Anandtech reports that the R5-1500X will have 16MB L3 cache while the R5-1400 will have 8MB cache.
Ryzen has a limited range of multipliers available for memory, and the higher multipliers currently seem to be unstable/unreliable right now. (possibly improving via UEFI updates)Übermatik;232207049 said:
This means that you need to increase the base clock and use it with a lower multiplier to reach high speeds. Most reviewers only seemed to hit 2933MT/s without increasing the base clock.
The Ryzen platform requires an external clock generator to change the base clock (BCLK) - so the only boards which currently support high speed RAM are the ASUS Crosshair VI, ASRock X370 Taichi, and Gigabyte X370 Gaming K7.
People have suggested that the MSI X370 XPower Gaming Titanium will also have this, but I haven't seen anything on their website about it. Nice looking board though.
There's a video which goes in-depth about this here: https://www.youtube.com/watch?v=frw9HRwqODk
It seems to matter more for gaming than anything else.Übermatik;232207049 said:
Typically RAM only brings minor improvements compared to how much you pay extra for really fast RAM (and a motherboard to support it).
Coulomb_Barrier
Member
What a silly thing to say.
First of all, you have no idea what high end laptop makers are going to do. Secondly, at lower frequency Ryzen is a really efficient and powerful arch. There are also rumours Apple are going to do exactly that - dump Intel for Ryzen in their new Macbook Pros.
I think the win for AMD in the portable segment will be Ryzen based APUs. Plenty of CPU grunt and a far better IGP than Intel can offer gives buyers a much more balanced product that can do everyday things as well as a bit of light gaming.
Well I suppose it has been six years now, so I wouldn't put it past them, but I would be surprised if Apple are going to drop Thunderbolt support so soon.
TatteredHat
Member
I think I'll probably end up going for a 1600X, but what's this about high frequency memory issues? Is that still going to be relevant by the time R5 releases? I haven't really been keeping up with Ryzen news.
Currently higher frequency memory has to be down clocked a lot of times as per Feist's post a couple posts above.
From Feist:
If you're buying in April, it would probably be best to get minimum 3200Mhz Samsung B-dies even if you can't use that speed till May.
jfoul
Member
Has this been posted?
TECHPOWERUP - AMD Ryzen 7 1800X Review
Once the motherboards & BIOS mature, I think I'm going to build a 6/12 core work desktop.
TECHPOWERUP - AMD Ryzen 7 1800X Review
Once the motherboards & BIOS mature, I think I'm going to build a 6/12 core work desktop.
TatteredHat
Member
Nice power consumption graphs there:
An 8 core more power efficient than a 4 core. They're gonna have good laptop chips it seems.
Though, I have to wonder why it uses more energy with SMT off.
JohnnyFootball
GerAlt-Right. Ciriously.
Very excited that the 1600X is coming next month. This should be an excellent CPU for gamers on a budget.
franklyshankly
Member
Is it expected to be better than the 7600K for gaming? The 1600X is technically more expensive so it will be interesting how it matches up.