Windows 10 has pretty much been confirmed as the reason why Ryzen is slower for gaming due to a problem in it's scheduler. Tested on Windows 7 games are performing much better in DX11.
Microsoft suck Intel cock so don't expect a patch until Intel rain on AMD's parade.
forums.anandtech.com
forums.anandtech.com
Other urls found in this thread:
hardware.fr
support.amd.com
servethehome.com
youtube.com
servethehome.com
phoronix.com
phoronix.com
phoronix.com
phoronix.com
phoronix.com
phoronix.com
twitter.com
it's one of many software problems.
Warhammer
Windows 10 - 1080 Ultra DX11:
49.39fps (Min) 57.16fps (Min)
Windows 7 - 1080 Ultra DX11:
62.33fps (Min) 62.00fps (Min)
Retina Display doesn't have this problem
It's like ryzen can even tell which operating systems are good
Pretty interesting if it's true.
Would be nice if some official source like Anand could do some proper testing on this.
>Windows 10
>CCX load balancing
>SMT
>BIOS/fw problems, especially ASUS
AMD has a lot of work ahead of it.
Jesus Christ, that's a 15% min frame loss.
W10 must be shuffling threads unnecessarily, leading to cache evictions.
Fucking MS.
>reddit face on a minecraft body
please RMA yourself
I have win10, it isn't bad, I prefer linux (and will switch back once my games backlog is done) but it's not worse than any other version of windows
DELET THIS! AMD FUCKED UP THEIR LAUNCH! IT HAS NOTHING TO DO WITH WINDOWS, OR MOTHERBOARD MANUFACTURERS, OR GAME DEVELOPERS! REEEEEEEEEE
Isn't pic related the best Windows version for gaming?
It gives the best performance.
Is can see how people would prefer the Windows 7 interface, but that's not really important if you only use it to start Steam or games directly.
-BB-BBBUT RYZEN SIN'T SUPPOSED TO WORK IN WINDOWS 7
M-MASAKA!
T-THERE ARE NO DRIVERS
Someone needs to test Ryzen on 8.1
It also doesn't have Ryzen.
>currently installing win10 for the first time
Should I have stuck with 7? Not sure how this'll go. I can always roll back to 7 if it's shit.
No, the reason Ryzen is slower is for gaming is because of the flawed memory subsystem.
Here is the best test on ryzen I read and they have 2 pages about it and explaining why it performs less in gaming tasks.
hardware.fr
It's in French, just use google translate.
>stuck with 7
no. She is getting old now.
>No
Yes
>windows 10
into the trash.
AMD just sucked dick at developing their Windows 10 driver.
Maybe most of their developers are still using Windows 7.
Windows 10 is probably the cause of both the CCX and SMT issue.
For fucks sakes W10 reads Ryzen to have 170MB cache!
Behaviour on windows 10 is just a consequence of this memory subsystem design. This is a very interesting read.
No, it's the double memory controller design.
ONE PERSON with 2113mhz ram. This doesn't confirm anything i will wait for tests by more trustable people especially reviewers.
How easy people forget, X99's launch was filled with bugs and errata, and suddenly everyone's surprised AMD's new platform is buggy on the software side?
TLDR translation:
L1 bandwidth higher on i7, doesn't explain results.
L2 higher on Ryzen, intel lower latency.
L3 ryzen latency issues, and bandwith to a lesser degree
L3 can have a different latency in function to which core inside a CCX
behavior of this type on the first 4mb and strong increase afterwards
after 8mb, on access of 12mb in a combined L3 cache of 16mb, latency is close than access higher than the 8mb cache of l3
in this case figure the L3 of the second CCX isn't used, Ryzen behaving as if it only had 8mb L3
the other CCX isn't solicited, its cache not filling
W10 likes throwing thread processes from one core to another
the bandwidth between the two CCX is only 22gbs
Bandwidth is weaker than the L3 inside the CCX (at least 175gb/s) and esp. lower than the memory bandwidth
It is a possibility MS patch windows so it has a more adapted behaviour on ryzen platform, or through a driver. But the design can't be patched and they'll have to do a hardware revision to "fix" it.
Why is it doing so good in threaded applications against a 6900k and sometimes a 6950X then? The OS doesn't throw threads over the CCX willy-nilly?
>I-It's because of windows!
Like pottery.
Probably because those applications spawn less threads than a game.
A lot of issues seem to be fixed with a monolithic core complex but I think AMD would lose its fantastic power efficiency for low threaded workloads under certain conditions so I have no idea if that's the best bet.
>b-b-b--bbut people said Windows 10 is good, that there is no issues, ever!
It takes a special kind of retard to use windows 10 a literal downgrade with a lot of issues and privacy invasion, but hey goys will buy up the marketing shit, even when they made it free and tried desperately to shove it down everybodys throats people didn't stop to consider why
seems no drivers is better than microsoft drivers that actively fuck up shit :^)
...
more and last TLDR:
>Why is it doing so good in threaded applications
choice to separate integer and floating point which is specific to zen architecture, intel mixes all units on ports (independent lines) without separation.
scheduler dedicated to floating point has 96 entries for a total of 180 entries per core (against 97 for skylake), in practice it's a double scheduler
units of execution 2 FMA 128 bits parallel per cycle, Haswell advantage of two FMA 256 bits per cycle; eventual point where zen arch can be limited.
In summary re mem controller on W10
communication between CCX has a cost, which isn't negligeable depending on the applications
For video endoding for example the effect is close to nil
For others like games, ex. BF1 20% less perf
the memory latency on ryzen is higher than normal possibly because of the design of two independent controllers both linked to the CCX via the data fabric
The only link between the CCX to the data fabric passes all communications, RAM, L3 sync and PCIe. Bandwidth is relatively weak considering need and largely inferior to L3 bandwidth
Technical limitations, the evident solution is to patch the Windows scheduler to limit thread movement from a CCX to another.
Another solution is the arrival of "Game Mode" which limits thread movement
Honestly why are we still letting OSes handle thread scheduling instead of the CPU directly.
haven't seen much about posted about it and I don't really care much since I'm on linux, but ryzen is supported on win7
>support.amd.com
>Supports Windows 10/7 (64-bit)
>reddit face
the irony, Sup Forums doesn't seem to remember his former self
>windows 10 is garbage
We've known that since windows 8
Nothing wrong with Windows 10, it's a great, fast and lean OS
>implying suits that buy multi-socket racks cooled by 200kW+ of cooling care about Windows 10 or gaymer performance
boy these shillbots are really shit aren't they...
i see no issue here
only dumb cunts use win10, that cpu is ready to roll out
Serves AMD right for only supporting Windows 10 and not Windows 7.
LINUX PERFORMENCE BENCHMARKS WHEN
free driver plox
servethehome.com
There's Phoronix test too,
And this is on the buggy ASUS with an old BIOS, Gigabyte boards can get up to 20% jumps in certain tests.
>these guys have a fucking XeonPhi and a swath of Dual Xeon setups and can run AWS whenever they feel like it
I hate not being rich.
youtube.com
Same guys booting Wangblows and testing Cinnebench under a 256 thread Knight's Landing.
It's a shame most review sites don't test these things, they seem much more interesting than sysmark/passmark and other synthetic nonsense.
>ubuntu 1404
>1404
RHEL, CentOS and SLES don't run bleeding edge, nobody will fucking run non-LTS on a production server.
>she
servethehome.com
Here are the 4.10 results.
Still, nobody will run this on a new kernel, maybe just with backported ISA support
They should do this again because AMD pushed a SMT fix yesterday.
phoronix.com
When will Windows fix it?
The 1700 would be sick value for a custom kernel, it can cTDP down to 35W on 1.9GHz on all cores, imagine running your dnssec/pfsense/NAS on these babies?
>ASUS Prime B350-Plus
>B350
>B
Triggered
Not a good idea looking forward, eventually most new games will run on only on APIs that W7 will not support (or will but at a loss of features) requiring you to upgrade anyway.
Games that support both DX11 and DX12 have shown much better performance at DX12, which probably means this issue is likely specific to DX11 on W10.
For me W10 had major usability and ui improvements and minor compatibility loss on certain exotic hardware that I occasionally use (esata port multiplier has power saving issues for example)
That is, until I wanted to play some JA2v1.13 - turns out neither W10 nor W8 can run it for more than a few hours until it freezes, and even that's only using compatibility tricks in registry and wined3d dlls to run it at all.
>Games that support both DX11 and DX12 have shown much better performance at DX12
Literally the opposite of truth.
Check quantum break for example of this.
> (or will but at a loss of features) requiring you to upgrade anyway.
Except vulkan doesn't require you to use windows 10. But nice try microshill.
>not treating all technology as 'she'
>being this much of a faggot
iPhone 7 Plus doesn't have this problem.
>Gaming
Back to your containment board
A few days ago was looking at DX11 issues on W10 and it was suggested that DX11 is emulated unlike on W7 which has a straightforward driver.
They followed through with microsoft SDK installs to force native DX11 without emulation overheads.
I did some further searching and saw more commentators on DX11 emulation annoyance with W10.
Considering that most of these will be used in ITX setups with X300 chipsets then I don't see the problem testing it with a lowend chipset on a cheap mobo, these things at stock don't need 15 chokes.
>tfw got into the hype
>it's broken-by-design shit once again
There are already some Ryzen Linux benchmarks including the gaming thing with the use of Linux kernel 4.10:
phoronix.com
phoronix.com
phoronix.com
phoronix.com
Though in gaming Ryzen has no advantage over an i7 7700k. Newer kernel doesn't improve much:
phoronix.com
>hurr durr muh DX12 just wait
>MUH vulkan
WE GOT| RUST GUYZ
We just can't have nice things, can we?
>using any windows version past 7
I'm not even an AMD user, but you must be retarded to use 8 or 10.
>windows 10 is shit
more news at 11
pajeets and gaymers BTFO
>win10 has """smart""" multithreading
>contraproductive
>win7 just werks
Maybe this is also the reason why some older Core2 computers run like dogshit with win10
as first approximation you can assume you'll need as much power in cooling as you have in equipment.
and even 10 kW/rack isn't really achievable unless in some specific conditions.
if you assume redundancy and 2 full rack PDUs on 32A/230V (which would be more or less the same for 63A/110V, i'm not really familiar with murrican electrical installations) sockets you get 7.3 W/rack max.
for comparision full 47U rack in telco r&d environment rarely takes more than 3kW, but all of the equipment is on -48V.
>But the design can't be patched and they'll have to do a hardware revision to "fix" it.
What you're saying is that Windows doesn't utilize it properly, so Instead of having windows change to utilize it, we should have AMD change?
No, what I mean is that it's a design that shows its limits in this windows 10 situation.
Of course MS and AMD can work so windows adapts its behaviour for ryzen. But the design in itself is flawed, ideally it shouldn't be bothered by how windows 10 handles threads. It's a design flaw.
Wow... Lol.
Which is why Linux and Earlier windows visions handled it just fine?
It's not a windows 10 problem, it's a design flaw?
>ideally it shouldn't be bothered by how windows 10 handles threads. It's a design flaw.
I won't lie, I'm probably reading this incorrectly
but are you calling it a design flaw because it should have been designed around windows 10/take it into account? Or it's just an innate design failure?
Has Jim Keller failed everyone once again?
No, it's because both CCX have to talk through the memory controller for accessing the other CCX cache, which induces latency. AMD could have designed it so it doesn't add latency. Performance on windows 10 is a consequence of this, not a cause.
But AMD and MS will work it out I suppose by decreasing thread roaming, but they'll never be able to fix cache latency because it is flawed by the design itself. The problem will always be there, MS will just make it so there are less thread swaps, which is the condition where latency occur (because of flowed design).
So you're saying there's room to reduce thread swaps?
Wouldn't having excessive thread swaps be a design flaw for Windows?
Thanks for talking with someone whose not really up to speed on the subject.
Of course, thread swap, or just make it so thread are only swapped inside a CCX, not between both CCX.
>Wouldn't having excessive thread swaps be a design flaw for Windows?
Intel CPUs aren't victim of this behavior. MS and AMD will have to do what's basically a fix to circumvent this flaw. In practice it should be globally fixed and it will be transparent for the user.
Intel BabyKake BTFO
>Wouldn't having excessive thread swaps be a design flaw for Windows?
it does it for intel as well, it works fine for intel since it's implemented for intel, for what purpose anyone's guess
linux doesn't swap threads all the time and works fine
>it works fine for intel since hasn't as much cache latency
Fixed.
it happens when software makes you dump x10 times more cache than you actually have
Random web forums truly the authority on hardware and OS issues.
This is fucking sad, it's getting to ancient aliens levels of retarded conspiracy and "research".
May've those guys can figure out who shot JFK or unlock the power of Yakub and his genetic grafting techniques next.