From: "Jarle Leirpoll" <AE-List@media-motion.tv>
To: "'After Effects Mail List'" <AE-List@media-motion.tv>
Subject: RE: [AE] Uservoice feedback website - Top post - Full Program.
Date: Sun, 21 Oct 2018 13:13:14 +0200
Message-ID: <004301d4692f$0fa3f610$2eebe230$@online.no>
MIME-Version: 1.0
Content-Type: text/plain;
	charset="us-ascii"
Content-Transfer-Encoding: 7bit
Thread-Index: AdRpKxL+kFAnx1jBQqe1b7A7DiXasQ==
Content-Language: no


>Stephen van Vuuren:
> There has been no statement from Adobe that it will move to GPU as the
primary and sole rendering engine.

I never said so, and I don't even think Adobe wants to do that. Some tasks
do run faster on a CPU, and those would benefit from being multi-threaded.
As stated before, some effects _are_ written for multi-threading. But they
do keep adding GPU accelerated effects for a reason. 

>> When everything is done on the GPU, frames can be kept in the GPU buffer,
and AE will get a lot faster.
>Do you have any evidence to back this up? 

Evidence? Not really. But since the processors are separate units and don't
share the memory, the result of the calculations of one needs to be
transferred to the other. That takes time, and I don't think any proof is
needed for that. 

Plus, the pipeline for this transfer is not the fastest. The bandwidth
between different CPU cores, and between CPU cores and RAM is much higher
than the bandwidth between CPU and GPU. Bandwidth between GPU cores and GPU
RAM is also very high. So every time you need to pass data from the CPU to
the GPU and back it slows down the process.

Say you want AE to add four effects to a layer:

The first effect is GPU accelerated, so a frame is sent to the GPU. The GPU
calculates a new frame. 

The second effect is not accelerated, and needs to be run on the CPU. So the
frame from the GPU must be transferred to the CPU, where the effect is
calculated, and outputs a new frame.

The third effect is GPU accelerated, so the frame is transferred again to
the GPU, and the GPU outputs a new frame.

The fourth effect is not accelerated, so again we need to transfer the
frame, do the calculations, and output a final frame.

Compare this to running all four effects on the GPU. One transfer to the
GPU, and the GPU runs all four effects (at a higher speed than the CPU could
ever hope for), stores every output in its own RAM, to which it has a
super-fast pipeline, and outputs the final frame. 

If you don't think this is beneficial, or don't believe this is how the CPU
and GPU work together, that's OK with me. 

I happen to think faster is better, and that this is why it's faster. I also
happen to think that this is why every video post processing software on the
planet seems to rely more and more on the GPU.