This isn’t really possible anymore (at least unaided by LLM). In the era of fixed function pipelines, it was possible to do something new under the hood.
The current standards of Vulkan and DirectX have exposed a lot of inner workings to extract performance which makes it hard to make a large shift. While this makes it easy to use the HW, it also makes it hard for the devs to have a single algorithm for diff HWs, ie, each new feature will need in investment by them.
For any novel functionality, devs will not like to work on features only on a small set of HW, unless they are aided by HW vendors. This creates a chicken egg situation: new features don’t get prioritized up by devs which makes it hard to market the features which makes it a small market which makes it likely not to be a priority by devs.
In this era where code is cheap, perhaps we’ll see more fast moving devs and HW both
Maintain the same algorithm optimised for each new generation. The algo needs to tailored to suit the game but adjusting per HW which was a daunting task (and why PC ports were sub optimal on a console comparable HW)
AI needs much simpler HW (since the SW is bespoke) and that’s why we see so many startups tackling this. I’m sure as ML becomes a bigger part of gaming (via coop matrix and LinAlg), we might see analog entering the gaming HW.
Gaming HW has a lot more cruft added on by standards and legacy SW. The number of gaming HW startups can be counted on one hand. QC and Apple have taken generations to get their HW ready. Even Intel needed 2 generations for their dGPU despite having a long line of iGPU
FPGA are very inefficient as compared to ASIC and need dedicated programming. It feels further than analog if we’re talking about a gaming HW using them
This isn’t really possible anymore (at least unaided by LLM). In the era of fixed function pipelines, it was possible to do something new under the hood.
The current standards of Vulkan and DirectX have exposed a lot of inner workings to extract performance which makes it hard to make a large shift. While this makes it easy to use the HW, it also makes it hard for the devs to have a single algorithm for diff HWs, ie, each new feature will need in investment by them.
For any novel functionality, devs will not like to work on features only on a small set of HW, unless they are aided by HW vendors. This creates a chicken egg situation: new features don’t get prioritized up by devs which makes it hard to market the features which makes it a small market which makes it likely not to be a priority by devs.
In this era where code is cheap, perhaps we’ll see more fast moving devs and HW both
What’s an LLM going to do to help, repeat their own documentation back at them?
Maintain the same algorithm optimised for each new generation. The algo needs to tailored to suit the game but adjusting per HW which was a daunting task (and why PC ports were sub optimal on a console comparable HW)
You’re not considering FPGAs, or the analog domain. For example:
AI needs much simpler HW (since the SW is bespoke) and that’s why we see so many startups tackling this. I’m sure as ML becomes a bigger part of gaming (via coop matrix and LinAlg), we might see analog entering the gaming HW.
Gaming HW has a lot more cruft added on by standards and legacy SW. The number of gaming HW startups can be counted on one hand. QC and Apple have taken generations to get their HW ready. Even Intel needed 2 generations for their dGPU despite having a long line of iGPU
FPGA are very inefficient as compared to ASIC and need dedicated programming. It feels further than analog if we’re talking about a gaming HW using them