Year in Review 2018: GPUs

anandtech.com

23 points·egnehots··7 comments

Why quad data rate? The fundamental advantage of double data rate is that the clock signal has the same maximum frequency - and the same signal integriry requirements - as the data lines, whereas in single data rate it runs at twice the frequency making it the limiting factor. I can't see much obvious advantage to going above that though.

The quad data rate trick is reading on differentiated and level gated clock. Actual clock signal is still same rate. Works when clock rates are high as the transition time gets close to clock pulse time - the clock is no longer considered square but exponential trapezoid.

It is 2x bandwidth compared to just using clock edges, but requires nice bypassing to get nice clock transitions and is obviously quite a bit higher transistor count for stabilization. It is similar to clock doubling but much simpler and thus faster.

Of course you get some latency benefit.

Nothing of relevance happened.

The famous last word of any self confident predictor of tech trend.

Yeah, still waiting for a ray-tracing quantum-differential-tensor-kanren GPU...

Are any of these cards supported by FLOSS drivers? If not, then they're essentially useless for anyone that wants a stable system.

To answer my own question, looks like the mentioned AMD RX580 has a non-Free EULA: https://www.amd.com/en/support/gpu-pro-eula

They can keep it.

Reply on news.ycombinator.com