Raven Ridge With The Ryzen 5 2400G On Mesa 18.2 + Linux 4.17 Is Finally Stable

[dungeon]

Not so fast, cowboy. The chart seems wrong.

No, it is not wrong, i expected exactly this range in openarena pbench DDR4 is the same/similar, eDRAM is faster than average DDR4 (eDRAM behave somewhat like DDR4@3600Mhz but with lower timings too or to say like normal at 4000Mhz) and DDR3 is of course slower.
Remove one memory module, you will see about half speed. Put slowest dGPU just with GDDR5 and that will beat all of them. and so on...

Whole purpose of having that chunk of eDRAM is to be less bandwidth bound.

So it is normal to be slower there, only maybe DCC engine might help there on Ryzen APUs to be in range of Iris Pro's eDRAM, as that at least theoretically can save bandwidth by up to 30%.

[Space Beer]

... but since it does not support ECC I've sold the Ryzen 5 2400G APU and I'm waiting for the PRO version that according to the motherboard manufacturers will support ECC.

bridgman any idea on when the PRO version might appear?

Raven Ridge has ECC support. It's the motherboard that might lack that

https://forums.anandtech.com/threads/amd-ama-starts-at-12pm-est-on-wednesday-may-9.2545291/page-3#post-39418197

[starshipeleven]

Why exactly is ECC for an APU so important to you, to the point that you'd sell your 2400G? It's kind of the equivalent of putting premium gas in a cheap economy car.

I'd like to point out that this APU runs more or less the same as the non-APU Ryzen 5 1600 and is still a quadcore with hyperthreading like an i7, it's not a "cheap economy car", but squarely into midrange CPUs especially for multithreaded workloads. I only need some kind of integrated graphics, currently I'm perfectly fine with a HD4000 (Ivy Bridge from Intel) don't need that much graphics power as the APU can get me as I have a separate gaming system with a dedicated GPU, but it sure is nice.

That said it should be obvious. I want best possible data integrity, same reason I use btrfs (with safe and proven features, on the distro that is most likely to do it right, OpenSUSE, and I never had issues because of that).

Also, since RAM cost a ton anyway, there is no real reason to not get the best, the price difference is insignificant. Having ECC on consumer hardware is nice because I can actually find out pretty easily if the damn thing works, by just OCing the RAM until I start getting errors in the logs. Good luck trying to "validate ECC" on server hardware, and I don't give a shit about their certifications, I don't need a scapegoat I can blame if something goes wrong, I want the thing to actually work.

I was fully expecting to lose some money, but due to ebay being full of crazy, I lost only 30 euros as I resold it for a pretty decent price. The rest of the system I built is now running with a Ryzen 5 1600x and the rx570, so whatever, I did have a backup plan.

I could have done it far more cheaply with Intel, as they still offer lowish-end CPUs (i3, Pentiums at least) with ECC support, not to mention the 2-3 gen older used Xeons, but I prefer AMD because of all the small things. It is newer, it has a UEFI option to disable access to the PSP (which may or may not work, but is still not a total PITA like the ME_cleaner that still does not guarantee much more than that but requires hardware flashing to work at all, and I'm sick and tired of SPI flashing bullshit issues), isn't affected by Meltdown, but most importantly is a small minority of the overall CPU market, so the PSP is a far less likely target for malware than Intel's ME that is found in more than 80% of all hardware in the wild.

I would really like to get my hands on some program that can actually test the communication with the PSP though, so I can see if the UEFI option is actually working or not.

[schmidtbag]

Hey Micheal.... a request...
I know that just about everyone is killing for those Desktop APU, but would there be a different between them and laptop APU's? Would be be able to test any Ryzen laptops?
I am concerned if only because I am shopping around for a new laptop and Ryzen is sounding really good for a linux laptop right now. Unless the APU's have issues.

If you can find an APU laptop with dual-channel memory, I'm confident that would make for a decent buy. I'm not sure if current ones are Vega based, but they have good performance regardless. The tricky part is most of them are configured for single-channel memory. Considering how both Ryzen and Vega are starved for bandwidth, you should only consider dual-channel setups.

... but since it does not support ECC I've sold the Ryzen 5 2400G APU and I'm waiting for the PRO version that according to the motherboard manufacturers will support ECC.

Why exactly is ECC for an APU so important to you, to the point that you'd sell your 2400G? It's kind of the equivalent of putting premium gas in a cheap economy car.

[bridgman]

bridgman any idea on when the PRO version might appear?

I see Aug 29 mentioned a lot - but some sites are saying that is "retail availability" and others are saying "more information". I'll see if I can find out which one is correct - the first sounds more believable to me.

[starshipeleven]

... but since it does not support ECC I've sold the Ryzen 5 2400G APU and I'm waiting for the PRO version that according to the motherboard manufacturers will support ECC.

bridgman any idea on when the PRO version might appear?

[Michael]

Now that shader caches have become more common, is it worth revisiting the idea of throwing away the results of the first run in order to get a more representative performance number ?

PTS already has that ability to always toss off the first run or an arbitrary run position, but I don't think the shader cache is causing any significant difference. Since right now after the third run it's dynamically increasing the run count if the standard deviation is about 3.5%, but I rarely see that happening... Only with some very CPU limited scenarios do I ever see it then go 4, 5, or 6 runs to lower that std dev below 3.5%.

As far as the frame times go, the last run is used for matters of not wanting to get the shader caching or so from the first run, etc. And in that sense of frame-time likely doesn't make sense trying to average each frames time.

[bridgman]

Now that shader caches have become more common, is it worth revisiting the idea of throwing away the results of the first run in order to get a more representative performance number ?

[marek]

The frame-time data is based upon the last run of the program while the chart FPS data is based upon the average.

OK. It looks like there were some crazy stalls slowing everything down. I don't have an explanation for that, but I think the shader compiler performance is the first thing to investigate.

[Michael]

The chart seems wrong. If you look at the Total Frame Time chart, it shows that average frame time = 5.62, so: 1000/5.62 = 177 FPS.

Ryzen 5 2400G is the only tested chip where the FPS chart and the Total Frame Time chart don't match. Weird.

The frame-time data is based upon the last run of the program while the chart FPS data is based upon the average.