AMD's Lisa Su responds to disappointing Ryzen gaming. "We're optimizing now"

[dasa]

People need to compare results to similar CPUs, not CPUs with half the cores and half the threads.

people should compare with everything to make a choice based on what best suits there needs
for those with mixed workloads ryzen looks good but for those that just want to game... like much cheaper cpu it will do the job but other cpu for a similar price can do it better
yes amd will get better but only time will tell how much

guru3d results are largely gpu bottlnecked and have some real odd results
the division
fx8370 82fps
1800x 74fps
farcry primal
7600k 108fps
4790k 96fps
6700k 92fps

[Ket]

I'm not saying ditch Quad results, just move them to their own table Mixing 6 and 8 core CPUs in with quads just misrepresents them IMO.

[Metroid]

I'm still waiting for the overclocking with smt off, any attempts? I mean back few years ago, I clocked my 17 920 from 4.2 max oc smt on, to 4.6 with smt off same voltage, wonder about ryzen....

[Chrono Detector]

At at Ryzen isn't a massive disappointment like Bulldozer was, at least it wasn't power hungry and it wins in half of the benchmarks against the i7 6900k which is rather impressive, though the gaming benchmarks at 1080p were disappointing but at higher resolutions this wasn't a problem.

[Sparky]

I'm still pretty sure my next upgrade - whenever that is - will by Ryzen, at this point. I do a mixed bag of stuff with my PC, and Ryzen will still be a boost from my old 2600k, and it will be also something new to play with, plus hopping camps is a good way to maintain a little bit of (in)sanity

[NEOAethyr]

Sorry for this bit of out of place spam :\ , I was doing some math to figure out if ram bandwidth would fit into the interconnect.
I kept thinking about it last night...

2x ddr4 3200 = 50gb/s ~

64 / 8 = 8 bytes
3200 x 8 = 25,600 mb/s...
x 1024 = 26,214,400 kb/s
x 1024 = 26,843,545,600 bytes per sec
/ 1024 = 26,214,400 kb/s
/ 1024 = 25,600 mb/s
/ 1024 = 25 gb/s ...

x 2 = 50gb/s for dual chan

3200 x 8 / 1024 = 25 gb/s max theorectical
2666 x 8 = 21,328 / 1024 = 20.828125 gb/s x 2 = 41.65625 gb/s
2400 x 8 / 1024 = 18.75 gb/s x 2 = 37.5 gb/s dc

The frabic interconnect thing is supposed 22gb/s
Bytes or bits I don't know for sure, nor do I know bitrate or if it's bidirectional or multi, or singe direction...

I would assume it's the max offical ram div, 2666?, either that or it's 2400.
The frabric supposedly run at a fixed ratio of the ramd div in some way, I've read half of the ram div.

For example:

Ram:
100 x 21.33 = 2133
100 x 24 = 2400
100 x 32 = 3200

Interconnect:
100 x 10.665 = 1,066.5 mhz
100 x 12 = 1200 mhz
100 x 16 = 1600mhz


1200 x 8 = 9,600 / 1024 = 9.375 gb/s

2400 mhz ram is:
18.75 gb/s single chan
37.5 gb/s dual chan

That is assuming each mem controller ?, ie ccx ? is 64bit..., 8 bytes wide.
If it's 16 bytes wide, 128bit per ccx?, does this make sense btw?, will the apu share l3 and mem controller sorta what seem's is going on with ryzen?
If 128bit then in theory it should match up with the ram theoretical speed.
And then there is no actual bandwidth issue at all...

However if it is 64bit per ccx like I think it is.
And I take whowever's word for it that said the rate was 22gb/s a sec.
Which I don't know the math for, it could be 11gb/s a direction for all I know.
22gb/s a sec is just fine for a single channel, but dual chan it can't handle.
In that sense what you're getting for bench results is the total sum of both l3 cache's.
With the giant latency that likely incurs because of constant cache invalidation and such haven't to check with the ram over and over to see if it's even in the cache.
My theory would be, if this is truly the case, that you need to double up whatever is the stock ram div.
To get that 44gb/s (lets call it 44k, ie 44,000 mb/s).
In reality you're more or less looking for at least 30k to compete with intel, 30-40k max.
I haven't done that math but I believe it would be around ruffly between 3800~4266 (probably to compare with intel at 3200, this is the speed I believe latencies will be almost normal).

So in order for it to compete, I think..., with intel mainstream, you either have to get the 3600 ram div working and push your luck from there.
Or use the 3200 ram div at 125mhz bclk ratio, and push for 4ghz ram, which probably would be the way to go.

Which means, it's probably better off to get hynix then it is samsung, if hynix plays well.
And obvously since apperently the mem conroller is kinda weak as it is, only 2x single sidded sticks, 8gb max each.
Since apperently 16gb sticks don't even post.

I think if this is the flaw, and you can get the rams speed up there to force the l3 cache to speed up, and fixing the issue compared to intel mainstream (not as good as x79/99 of course).
The crazy numbers in bandwidth you'll see will be inflated, take it as something that's better then intel in that case .
And that's if the l3 isn't a weak spot on clocking... ^^
And... the timings are in check, at a max of say around 20t at 4000mhz (could be one timing higher, the rest lower, or whatever sorta thing)
Hmm, that inflated mem bandwidth # could possibly counter slack timings too...


To the point -->
What I'm saying is I "think" these chips may fair better againt intel in gaming if you clock the crap out of the memory.
The ccx scheduling issue then becomes a non issue (mostly...).
Though your ram may get an extra beating compared to intel equ (from what I gather with cache invalidation).


Also, scheduling updates will probably happens around the time the apu's are ready to be released, that's when some oem will likely attempt to fix it up, my guess is 3 months +.
It could happen faster though.
There's also black plate mounting issues with some boards and kits causing no post, asus doesn't have iommu working I guess either.
I read somewhere they are working on getting 16gig sticks to boot.

[chew*]

http://www.xtremesystems.org/forums/...=1#post5254761

I think this explains alot. AMD is being rather conservative with subs we have no access to.

Plenty of performance to be extracted still.

Fixed speed gains

Smt off gains

Memory bandwidth gains

All adds up.. Enthusiast lvl. Probably be equal.

Noob mainstream running stock? Amd needs to solve those issues.

[MattiasNYC]

Do you guys with experience and more brain capacity than me have any opinions on the following:

One particularly interesting recent discovery revealed a bug that’s significantly detrimental to Ryzen’s performance in Microsoft’s Windows 10 scheduler. The scheduler does not appropriately recognize Ryzen’s cache size and cannot distinguish physical cores from SMT threads. This in turn is causing it to often incorrectly schedule tasks in the much slower — approximately 4 times slower — SMT threads rather than primary physical core threads.

?

That was from wccftech here, and it's further discussed on reddit and anandtech.....

?

[chew*]

Its rather simple a real core is faster than a "logical core" ( thread ). The OS is when the actual cores are available not using them and instead picking a thread.

Threads are preety good but not as good as an actual core performance wise.

In the area they are complaining the most "gaming".....which uses about 4 cores in most its not using actual cores most likely because its not being told to or does not know how to differentiate between what is a core and what is a thread.

Win7 reportedly is not having this issue.

[MattiasNYC]

Right. That's what I thought.

It'd be interesting to see if an adjustment on the software side actually yields pretty significant increased performance.

I went down the "rabbit hole" and I found among other things the following:

In 3D testing I did recently on Windows 10, the title which illustrated the biggest SMT regression was Total War: Warhammer.

All of these were recorded at 3.5GHz, 2133MHz MEMCLK with R9 Nano:

Windows 10 - 1080 Ultra DX11:

8C/16T - 49.39fps (Min), 72.36fps (Avg)
8C/8T - 57.16fps (Min), 72.46fps (Avg)

Windows 7 - 1080 Ultra DX11:

8C/16T - 62.33fps (Min), 78.18fps (Avg)
8C/8T - 62.00fps (Min), 73.22fps (Avg)

https://forums.anandtech.com/threads...#post-38775732

[chew*]

I'm not ready to jump off windows 7 myself.

Yes that plus just disabling SMT like I would anyway while gaming. No need for 16.

I like the fact and totally had full intention of using chip like this.

Gaming 8 core

work station 8c/16t

email internet 4c

The ability to do that in one socket one one board with 1 system at the touch of a button is priceless.

Also I posted up some info in the AMD section regarding memory which proves more performance is on the table and untapped due to current agesa and or bios.

Have a feeling its a little of both.

[MattiasNYC]

Agree.

On my end, doing content creation (audio) three's been one issue that reared its head that is a potential show-stopper for part of our industry. Turns out that under certain types of loads we get lower performance until the buffer in the audio-specific hardware/software increases. And since it's audio there are quite a few people who don't want a higher latency, meaning to them this is currently not a high-performance chip compared to the 7700k for example.

So, it'd again be interesting if this issue somehow penalizes the CPU where for FPS and audio latency there's a clear issue, and if a "simple" OS update (and as you say also bios updates etc) fixes this to a great degree. I'm certainly hoping that'll be the case.

[chew*]

Might be related to the info posted in my review thread as well as something mentioned on hwbot. Cache related.

[NEOAethyr]

@MattiasNYC
Have you seen this video?:
https://www.youtube.com/watch?v=Pj57X_WDojc

If you are looking to upgrade to this platform 24/7 and are doing audio, you may wanna hold off a month or so.
I saw one guys video with a b350 where most of his bios options did nothing.
Linux is broken a little (acpi and iommu).

And a windows update some time or other should help with this sort of thing (scheduler, the cpu info needs plugged in and probably some fixes).

[MattiasNYC]

Thanks for the suggestion. Yes, I saw that video yesterday or whenever it was.

I actually do have a bit of a problem with that video, because I tend to be pretty pragmatic about my purchases. The way the data is presented in that video isn't all that great in my opinion, and I think the conclusion is questionable. I'm too tired to explain the difference right now but if anyone cares I can. It's fairly significant actually (for audio guys).

I can certainly hold off for a month I think. And my situation is such that if I need a workstation soon I can just push the cost to my client one way or another.