… I just have a hard time imagining hardware 6x the power struggling to run games that were made on similar yet 7 year old architecture.
It's about
latency. Take your idea of
speed and break it up into two parts.
Throughput and
latency.
Throughput: a measure of how long it takes an action to finish once it starts
Latency: how fast till something starts in the first place.
On Switch, lets say you have a 10MB shader. When the game needs it, they ship it to the GPU, then GPU starts executing. Switch memory bandwidth is 25GB/s, so it takes about 0.4ms for the shader to start. That's latency - and it's
blocking latency, meaning the rest of rendering has to stop until the shader execution can begin.
On Drake, for a native game, it works the same, only with much faster memory bandwidth, of 102GB/s. so that 0.4ms becomes 0.1ms. Big win.
But for back compat, there are intermediate steps. The game tries to ship the shader to the GPU, but it won't run so the CPU captures it. The CPU then needs to read the shader. That's 0.1ms, minimum. Then the CPU needs to generate a new, Drake compatible shader, and write it into memory. That write takes another 0.1ms. Then the GPU can start reading it, another 0.1ms operation, before it starts executing.
All those memory read/write operations take 0.3ms total. The original had 0.4ms. That leaves 1/10,000th of a second for the CPU to generate a new shader from the old shader, before stutters happen. Shaders often take
dozens of milliseconds to compile. And remember, a game might need
multiple shaders per frame.
Look at any modern, shader driven PC game, you will see stutter, or literally an
hour of precompiling shaders before a game starts, on PCs with CPUs 10 times more powerful than the one in the NG. Look at Yuzu emulating a new Switch game on a monster PC with a cold shader cache.
Yes, there are
huge advantages to the architecture being similar! That's the only reason this is even possible to pull off. But the magnitude of the problem is high.
