AMD's Display Core difficulties

Posted Dec 14, 2016 16:35 UTC (Wed) by pizza (subscriber, #46)
In reply to: AMD's Display Core difficulties by Felix
Parent article: AMD's Display Core difficulties

> What is certainly interesting is to look why numbers are vastly different and why certain regressions happen, bisecting (or at least reporting) bugs with certain drivers. But Michael doesn't do any of this. Usually he does not provide anything like a "big picture" which factors might contribute to the results he observed, he does not provide any technical interpretation

Personally, I find the Phoronix benchmarks to be very superficial. In many cases that is indeed all that is called for (eg more FPS in games == better) but when there is more than one axis for a given situation they tend to not be of much use in the real world.

For network performance, the details matter quit a bit (Does FreeBSD even have a firewall by default?) What's the CPU load? Does other I/O in the system get starved out? What about handling a thousand connections going in both directions versus a single connection in one? To the same recipient or different recipients? What about using raw packet sockets, or packet filtering or QoS disciplines?

What I'm getting at -- there are *many* variables here, and mindlessly focusing on single-stream throughput is what lead to the "bufferbloat" problem that makes real-world users suffer.

(This is also why reason the server industry uses big, unweidly benchmarks that simulate real-world applications undergoing real-world loads, rather than microbenchmarks for individual components..)