Back when DirectX 11 was new and exciting, the Unified Shader Model meant that a card could render a scene with a "heavier geometry workload, but lighter pixel workloads", or "a lighter geometry workload but heavier pixel workloads". It was flexible. Plus, the DirectX 11 spec had hull shaders, domain shaders, geometry shaders and compute shader. A mixed precision would have limited the flexibility of developers.
Mixed precision does not limit flexibility in any way, in fact, it gives the developer more flexibility. It’s another tool in the programmers arsenal. Modern GPUs support full precision floating point, half precision floating point and integer Dara types natively, with different performance characteristics. Fir example new Nvidia and AMD cards can execute single and half precision operations simultaneously on the same shader core.
As to the other point you mention, that’s a complex one. Various shaders have run in the same hardware fir a while now, but it’s only recently that GPUs are able to schedule tasks efficiently. Apple GPUs are fully asynchronous, they have dedicated hardware task controllers that try to use all the available cores efficiently. Older desktop GPUs rely on driver to do task scheduling.
Interesting! Unfortunately that article is biased, b/c it's written by a company that uses TBDR instead of IMR, and thus only talks about the strengths of TBDR, not its weaknesses. Here's what appears to be a more balanced comparison, on anandtech. It's very old (2011), but still worth reading b/c of its clarity.
www.anandtech.com
Essentially, the anandtech article agrees with the Imagination article that TBDR can be more efficient, since it eliminates overdraw (drawing objects that are hidden behind other objects and thus can't be seen anyways). But it mentions that TBDR can have problems with more complex geometries. Since you don't see very complex geometries in mobile games, it seems Apple's approach back then was optimized for that market (where efficiency matters).
As you say, that article is old. Ten years ago mobile GPUs were often front-end limited. I don’t think this is the case anymore for modern Apple stuff. Asynchronous shaders take care if situations were you happen to have a lot of geometry. If you have too many primitives, the bin lists can spil, causing overdraw, but that’s Stil more efficient than immediate rendering. And even if one argues that binning (sorting polygons into tiles) demands a lot of work - well, current Nvidia and AMD GPUs do it as well, so it must be worth it for them.
I would like to see some geometry throughout benchmarks, I don’t think that anyone published data on that.
So this is conceptually interesting: Just as the AS Mac CPU's will be qualitatively different from those of essentially everyone else in the PC space by having an ARM ISA instead of an x86 ISA, the AS Mac GPU's will be qualitatively different from everyone else's by using TBDR instead of IMR (again, referring to the PC space). Though the difference in ISA is a macroarchitecture difference, while that between TBDR and IMR is, IIUC, one of microarchitecture (of course, AS CPUs will also have extensively customized microarchitecture as well).
I wouldn’t call it a microarchitecture difference, it’s an algorithmic difference. I’m quite exited to get TBDR on desktop, especially one with programmable GPU cache as offered by Apple. It allows one to utilize the hardware in much more efficient ways.
That isn't a meaningful comparison. It's unlikely that they could match a recent model gpu that would be considered anything but low end today. That hasn't been Apple's target in the past though. If they beat Intel's igpu numbers by 50% or more on the latest generation, then that would probably get significant interest. If they compare to an older chip that they happened to use in whatever machine, then it's more disingenuous marketing.
A two year old Apple tablet GPU already outperforms any integrated graphics solution currently offered by Intel or AMD and is head to head with last gen Nvidia 50watt laptop GPUs. With more GPU cores, faster RAM, newer architecture and smaller process AS has the potential to offer very competitive performance. I expect 5300M pro class in the 13” laptops at 30W combined SoC TDP.
[automerge]1595576803[/automerge]
I would say the current ARM in the Mac mini is not that fast GPU (metal) wise, but there you have a comparison. Given the speed improvements from my dated iMac late 2015 (Oct 2015) -> iMac 2017 -> iMac 2019 -> rumored iMac 2020 paired with GPU we have to see whose horse is faster. Whatever it is it has to be able to show a 4K HDR video via T2 chip.
Again, who cares. You are comparing your iMac to an iPad. This is not representative of the upcoming Apple desktop hardware.
