Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Oh no.. Nvidia RTX 30 series performances
- Thread starter Danny82
- Start date
- Sort by reaction score
I'd put it like this: it is definitely a fast GPU and I am sure that it makes a lot of sense for gamers, but I find it completely uninteresting. They simply traded more power for more performance. The 3080 does deliver more performance per watt, but nothing to get exited about. And most importantly, the biggest difference seems to be in 4K gaming, which can be entirely explained by the crazy amount of RAM bandwidth these GPUs have.
What does this mean in practice? That Ampere will most likely disappoint when you put it into a restricted scenario such as a laptop. At 50W and regular GDDR6 I would expect the performance improvements to be modest. That's why I am really looking forward to the 3060 reviews...
Where the Ampere cards excel however are simpler compute workloads. High memory bandwidth and ability to run two operations per clock will really make things run fast.
I have seen a few reviews talk about how these cards at 1080p are completely CPU limited. Jaytwocents on YT exclaimed that the 1440p results are as fast as the 1080p results for the 2080ti.
I will be interested in seeing how big of a frame(tile) buffer apple will include with their GPU, especially for notebooks where the native resolution is near 2160p.
As an aside do you know how many tiles it takes to draw a frame with Apples current GPU’s in say the iPad Pro?
Tile size on Apple GPUs can be one of 32 x 32, 32 x 16 or 16 x 16 pixels. I'd assume that each GPU core (or maybe a slice of a core) renders to a different tile, possibly multiple tiles in parallel to hide memory latency.
So the amount of tile buffer memory is not that important per se, it only limits the amount of data you can have associated with the tile while rendering to it. Apple GPUs currently limit it to 32KB, which allows you to store 32 bytes of data per pixel for the 32 x 32 tile or 128 bytes of data if you choose the 16 x 16 tile size. This is probably enough for most practical applications.
To speed things up, they need to a) increase the number of GPU cores b) improve the memory bandwidth per GPU core and c) improve the GPU core ability to work on multiple tiles in parallel. The c) is the only part that would benefit for larger per-core memory.
That is interesting, it makes the math seem mind boggling. If I understand it right, a tile can be up to 1024 pixels in size, meaning you would need something like 486 tiles per GPU core to render 1 frame on an iPad Pro 11.