A few years ago I proposed a "golden rule" as to when to upgrade your PC: only when you can get 2x the performance for the same price you paid last time. This rule has been perfectly applicable to CPUs. However, the pricing evolution of dedicated GPUs has been so appalling in recent years, the rule is pretty much inapplicable to them. Let's take a look.
The Nvidia GTX 1080 Ti (released in Mar/2017) was the flagship consumer GPU on the market of its time. Its MSRP was about $700, and it sported a PassMark performance score of ~18.5K. (I'm using those scores for GPU performance, so neither average FPS for specific games, nor general gigaflop/compute scores; just those scores from that benchmark.)
Only by Oct/2022 a consumer GPU model appeared on the market capable of exceeding 2x that score of the 1080 Ti. This was the Nvidia RTX 4090, which can achieve ~38.8K.
So it took industry ~5.5. years, but consumers can finally get more than 2x the PassMark score of a 1080 Ti. Here's the catch though: the MSRP of the RTX 4090 was $1600, also more than 2x the MSRP of the 1080 Ti.
So after 5.5 years, if you wanted a flagship GPU with 2x the performance score, you could finally get one, but you also have to pay more than 2x the original cost of the older one. Not to mention, it is also almost twice as power-hungry -> higher electricity costs per year (see GTX 1080 Ti vs. RTX 4090.)
This is hugely different from the pricing/performance evolution of CPUs. If you want a CPU with 2x or more the performance score of a five+ year old CPU, you can easily get one in fact paying even less than the original cost of the older one, and in some cases also getting one that is less power-hungry. An example among many: Intel i7-7700K (Q4/2016) vs. i5-12500 (Q1/2022).
How terrible is the "coupled-doubling" of performance and pricing of flagship GPUs? Well, quite literally: exponentially terrible.
Let's project the trend into the future, doubling both performance and pricing every 5.5 years. If we do that ten times, we get from Oct/2022 to Sep/2077. What performance should we get from a flagship consumer GPU in Sep/2077 then? And at what cost?
That GPU would offer more than 2^10 times (that's 1024x) the performance score of the current flagship RTX 4090 (nice!), but also would cost more than 2^10 times (1024x) the MSRP of the 4090, which was $1600. $1600 times 1024 gives us... cough... more than $1.6 million. I ought to repeat that in bold and uppercase: MORE THAN $1.6 MILLION!!!
And that's without making any adjustment for inflation between 2022 and 2077. Believe it or not, given the current trend, that's what a flagship consumer discrete GPU ought to cost in 50+ years. Even if three orders of magnitud more powerful, also three orders of magnitude more expensive.
Thanks, but no, thanks.
Basically, with the current pricing trend, it's impossible to get double the performance score in a new GPU at the same original price of a 5+ year-old flagship GPU. If sticking to my "golden rule," I might effectively never again buy a dedicated GPU.
Besides ignoring/breaking my own rule, there is another possible way out: SoCs like the Apple M2 Max. Mighty powerful CPU + GPU in a single chip. Similar to the integrated APUs from AMD used in some gaming consoles/handhelds, and laptops, just way more powerful, and yet more power efficient. Seriously hoping the price+performance evolution of SoCs/APUs will be much more promising than what we have seen recently from discrete GPUs.
