So...does anyone else remember that wonderful time in the past where hardware was constantly moving forward, and Moore's law was easy to harness as a cheap way to have shortcomings disappear beneath a pile of raw computational horsepower?
Does anybody else remember when everything started to slow down? Yes, the Intel shenanigans with 4 core chips for a decade was entirely because they didn't have AMD to compete against, but once that was sorted we got back to generational improvements not of 30-50% (read: peak in the 90's and early 00's), but of 7-13%. Anybody else remember when the concept of a generational refresh was something absolutely silly, because nobody would buy another $150 card if it basically only showed minor improvements. Yeah, I am old enough to remember all of that and team Blue, Red, and Green all claiming that we were reaching near peak hardware performance with generational improvements basically being decreased to minor changes.
So what are all three of the teams doing? Well, those old enough should recognize this as a netburst moment. What I mean by that is that instead of pursuing an ever less possible improvement in performance you can write entirely new rules. Just like Musk claims that Tesla should be viewed as a software company...(snicker)...if AMD, Intel, and Nvidia all become software purveyors they don't have to live and die on improvements in tangible things. They can claim something like a 30% increase in performance in their latest release, and in reality just have code optimization cover the entirety of their flaws to get that magical performance boost.
Let me offer a parallel. Warframe is a game made by Digital Extremes. Through decades of spaghetti code, at one point they were supporting all the way back to 32 bit windows xp. Kinda nuts, but one day they decided they needed modern instruction sets...and stopped supporting basically everything before Sandybridge and Bulldozer. This allowed them to cull the code base, and increase performance...but realistically it was improving the performance to a piece of software not bogged down with a decades old restriction set.
What truly frightens me though is the future of FPGAs and mismarketing. Let me state that Nvidia has specific processing for ray tracing....right? You've also got AI accelerators. Then you can design a bunch of other unique things...and claim how special you are. If you want to do this on a budget you get a big old FPGA and keep half a dozen images for accelerators...flashing them as needed to make the ultimate processor and co-processor combination. This then gets down to literally being software companies running relatively commodity hardware...at which point instead of companies trying to improve performance we'll be back to who can stick the biggest FPGA on, or who has the one ASIC image to flash to your shiny new FPGA. I don't even want to read all of the AI buzz word garbage...because it by definition is a fad. If you believe that anybody is going to sell you a consumer viable priced system to effectively allow reasonably quick local instances of LLMs then you didn't read that recent article where China was buying H20's from Nvidia for $12,000 each.
and ensure high quality UEFI.
