Four Years After Launch, AMD Kaveri Sees Huge Performance Boost On Linux

[haplo602]

No one is born a programmer. It's a multi-year time investment

Unfortunately some are born idiots and never get rid of it ...

[nanonyme]

Because all of us are born programmers and can contribute to the code ?? Or what are you trying to say here ?

No one is born a programmer. It's a multi-year time investment

[Namenlos]

Didn't it need a new firmware release with updated headers or something like that?

I checked, you are right, they need firmware "which supports the full 40bit addressing". I'll ping bridgman as it may have been forgotten after 6 months. But I am still in good hopes now, as AMD fixed UVD in the A780g chipsets too after some time.

[haplo602]

Thanks, your waiting is helping a lot!

Because all of us are born programmers and can contribute to the code ?? Or what are you trying to say here ?

[xpue]

A10-7850k paired with a GTX 1060

Thats some massive CPU bottleneck.

[dungeon]

I am aware, and I completely agree. It's just that in practice, I don't see that kind of difference.

Try openarena phoronix-test-suite benchmark, that config is memory bandwidth bound... there should be 3x maybe plus diff

Playing Dirt Rally as a quick example got me around 30-35 fps on medium settings with the APU (2 Gb shared) and 50-55 with the RX 560. Maybe it is due to the fact that I'm playing on a ridiculous 1366x768. I will try soon on FHD with my projector though, where the difference should be bigger.

Dirt Rally is CPU bound which is worse case scenario, so expected not even twice diff Try env var mesa_glthread=true if that is smart enough to change something there.

Real world GL is severely CPU bound (even in pletora of different ways), but real GPU differences could only be seen once you are GPU bound instead

So, in many cases, you should probably be seeing a fair bit more than 3x. I think getting the RX 560 was a good call.

It is better to say - only in ideal, non-CPU bound cases.

GNM, Mantle, Vulkan, D3D12, Metal... APIs are not invented for nothing. These lower level APIs tries to avoid or at least to diminish CPU boundware as much as possible.

[Mez']

Let's look at the specs, shall we?

Model	Generation	Lithography	Shaders	Clock	fp32 GFLOPS	Mem Bandwidth
A10-7860K	GCN 1.1	28 nm	512	757	775	34
RX 560	GCN 1.4	14 nm	1024	1175	2406	112

So, on the face of it, one would expect a ratio in the ballpark of 3x. However, the APU has the unique disadvantage of having to share memory bandwidth between the CPU and GPU cores. Another factor not shown in the above chart is that, over the generations, GCN has accumulated numerous improvements in such areas as geometry processing (tessellation, caching, early primitive discard), texture compression and caching, etc. The only thing the APU has going for it is lower-latency communication between the CPU and GPU cores. That might help select few GPU-compute scenarios, but probably little else.

So, in many cases, you should probably be seeing a fair bit more than 3x. I think getting the RX 560 was a good call.

I am aware, and I completely agree. It's just that in practice, I don't see that kind of difference.
Playing Dirt Rally as a quick example got me around 30-35 fps on medium settings with the APU (2 Gb shared) and 50-55 with the RX 560. Maybe it is due to the fact that I'm playing on a ridiculous 1366x768. I will try soon on FHD with my projector though, where the difference should be bigger.

When I have a bit of time, I will compare both again, this shouldn't take too much time as it is only a couple of grub boot options and bios options to adapt for the switch.

[dungeon]

So, on the face of it, one would expect a ratio in the ballpark of 3x.

That is in clear GPU bound cases, which could be expected to be seen in proper GPU benchmarks only... in real world apps and drivers are often bound by something else like a CPU so there might be less difference.

But yeah, it is safe to say that it should be at least 2 times faster if not more.

Theoretical GFLOPS shader peak shows 3 times difference, but in reality it really depend what happens - what app is doing, what driver is doing, it is hard in general to say which boundware per app will comes first.

[coder]

Supposedly, I agree.
In practice, that was around the difference between both when I added the RX560. With these improvements, the gap should be quite close now.

Let's look at the specs, shall we?

Model	Generation	Lithography	Shaders	Clock	fp32 GFLOPS	Mem Bandwidth
A10-7860K	GCN 1.1	28 nm	512	757	775	34
RX 560	GCN 1.4	14 nm	1024	1175	2406	112

So, on the face of it, one would expect a ratio in the ballpark of 3x. However, the APU has the unique disadvantage of having to share memory bandwidth between the CPU and GPU cores. Another factor not shown in the above chart is that, over the generations, GCN has accumulated numerous improvements in such areas as geometry processing (tessellation, caching, early primitive discard), texture compression and caching, etc. The only thing the APU has going for it is lower-latency communication between the CPU and GPU cores. That might help select few GPU-compute scenarios, but probably little else.

So, in many cases, you should probably be seeing a fair bit more than 3x. I think getting the RX 560 was a good call.

[dungeon]

VS ?

versus of course, not vertex shader

Nope, there's nothing Linux-specific there. If you check the release notes for the corresponding Windows drivers the supported hardware list is also dGPU-only. Every so often a workstation APU SKU will be produced, and in those cases the business folks ask us to provide support in the workstation drivers.

Just curious, other than the natural human desire to pour sand in gears...

I don't think natural human desire is to pour sand in gears as even gears are not natural, just man maded thing So we put air, because it have advantages thus is more practical. Air in gears is so practical since that pushed us to make and maintain special asphalt roads all around globe but these roads killed a lot of people, more then in all World Wars together

Same desire to be faster just like on roads, exist in computing - to be faster is main desire of both Until is stops being so practical and unsafe to be so fast, then same desire could be satisfied on Formula 1 races or so

How 'sand in gears' became 'pistons in wheels' and I got to Formula 1 and victims of WWWs and thse on streets and so on... no one knows, but both are the same and means the same 'absurd' but even that absurd is not far from reality

But this isn't any absurd, as speaking of practicality If someone who bought these PRO APUs expect workstation app to work on Linux too, but you provided driver that does not work how you expect these to be happy? People are happy when there is no much PITA with drivers, preferably to not even notice how driver exists and things just works

, what made you say this was a Linux-only decision ?

Your artifical VS differitation mentioned before, meanwhile advertising on several places for these PRO APUs talk about CAD too If quoting launch text only was not enough, here are couple pictures

My understanding is that Radeon Pro covers a slightly broader range of target users than FirePro did, with more focus on things like 3D rendering.

To me that is the same, just more current chips and of course different name

AFAIK, Radeon Pro is brand name for professional class GPU since year 2016. Before that was named Fire Pro, further before Fire GL, etc...

Radeon Pro drivers might only be name for drivers aimed for these professional class GPUs produced since year 2016.