Thursday, December 15, 2016

Does Radeon Chill Improve Input Lag?

AnandTech summarizes AMD's rationale in lowering frame rates during periods of inactivity as follows:
During periods of low user interaction or little action on screen the CPU and GPU power limit is reduced, causing the hardware to slow down ... one side-effect of this means that fewer frames are queued in the buffer, which AMD claims it results in a quicker response time from frame generation to frame output.
Reducing CPU power means that any ramp up from the slower state will have to be very fast in order to ensure no latency from increased CPU time. Absent hardware based frequency shifting, which is the case for pre-Skylake Intel CPUs, this can take up to 100ms although the time is typically closer to 20ms. So that's at least on or two frames depending of input lag depending on refresh rate. And aren't the frame queues around 3 at most? Or is that OpenGL only? Or was that VSync and triple buffering?

Time for real world testing.

TechReport did just that and noticed frame times with Chill enabled were occasionally faster than with Chill disabled.
That result does seem to mesh with AMD's claim that Chill can improve responsiveness by keeping more of the GPU available for times when fast rendering in response to user input is needed.
But that might just be part of the story. The real test will require the mice + led switch and high frame rate capture analysis. Paging systema.

Tuesday, December 13, 2016

Ryzen ... and that 6700K 4.5 test

Just watched the AMD Zen stream and of course the most telling information was what was not revealed. There was no word on price, frequencies, or single threaded performance. Not revealing the price is understandable but no information on upper frequencies or single threaded performance is unfortunate.

Intel's foundries are better so it's hard to imagine Ryzen matching the frequencies Intel is able to achieve. However, the 95W TDP for their 8 core part is a good sign. Then again, if that is 95W of AMD TDP, then I'm not so sure. If the scaling is anything like the FX series, then 4.7GHz base is going to be 220W which would be pretty good considering these are 8 full cores with better IPC. Even if it's higher, either way it's watercooling territory.

Single threaded performance is almost certainly worse or else AMD would have been demonstrating a lot more benchmarks apart from Blender where IPC appears about the same. Worst case? Bulldozer scenario where AMD cherry picked a few benchmarks to show parity or superiority to Intel's CPUs but underperformed in the real world.

I don't think this is the case with Zen. The Blender and x264 tests (Handbrake and Twitch) were a really good showing not only over AMDs previous micro-architectures but also versus Intel's latest. Maybe the x264 improvements are just the result of better AVX support and Blender is a one-off? Please no.

But the most controversial moment in the stream was the demo of the 6700K @ 4.5GHz versus Ryzen.

Most comments thought there was some rigging involved since the 6700K stream/gameplay looked like a slideshow. I don't think so. The streaming settings were almost certainly at a higher quality x264 preset or resolution/fps. Each step in x264 quality/resolution/fps requires a large increase in CPU power so it is easy to bog down any given setup this way.*

No one with a 4.5GHz 6700K is streaming a CPU intensive game using 1080p60 @ medium whereas that's probably doable with the 8-core 6900K or Ryzen. So yeah the demo was rigged in the sense that streamers are not going to use settings where gameplay and stream quality are trash but it isn't rigged in the sense that the 6900K or Ryzen are going to be able to offer better quality streams even against the best mainstream Intel CPU.

But that 6700K is still going to offer better gaming framerates when not using x264. And you get the best of both worlds with a dual PC streaming setup anyway. Ryzen would be a great candidate for a streaming PC though.

Anyway, Zen seems to be as or slightly better than expected but not what a lot of us were hoping for. Its success will depend on release prices and process improvements to improve frequencies. AMD needs to offer Ryzen at a much lower price than the 6900K to compensate for its likely worse frequency and somewhat lower likely IPC across most programs. None of that $900 FX9590 nonsense.

* Stream quality increases can be done granularly with fps but less so with resolution and x264 quality.