| ▲ | plqbfbv a day ago | |
> There also seems to have been a major shift in strategy from AMD / Intel / Apple / NVidia due to this, as I was used to CPU / GPU / chipset upgrades coming out more regularly with double-digit percentage improvements each year, but that seems to be basically over for the consumer space now that everyone is just trying to build as much AI hardware as fast as possible for enterprise markets. I believe all of them realized nobody really needs more than 16c32t on a desktop, and architectural changes nowadays can only go so far to bump performance. The biggest changes across generations in the last ~5 years were the introduction of specialized instructions for wider vectors (AVX512/AVX2, and even then you need specialized workloads to use them), bigger L3 caches and/or efficiency cores to lower power draw. The double-digits gains recently were in memory bandwidth, frequency and locality, because you can make a 5GHz CPU and make it go 10% faster than the previous generation, but you also need to supply the extra data for the CPU to make useful work so that it's not data-starved. > I also used to rebuild my desktop every 1-2 years and upgrade laptops every 2-3 years, but I'm currently a bit past that and have no plans to refresh anything in the near future. I'll probably wait another CPU and GPU generation or two to even consider it, especially with prices being this ridiculous. I built my desktop in 2019 (right before the COVID madness, luckily) and I was also used to 1-3 year cycles. I upgraded the CPU and GPU since then, but the performance is still that of a workstation, even though perhaps a little bit behind in pure single-core benchmarks. I plan to keep this until either CPU, RAM or MB fall out (hopefully another 5y at least), and then I'll make a huge leap, instead of making 4x smaller ones which deliver little uplift anyway. | ||
| ▲ | atiedebee 13 hours ago | parent [-] | |
The biggest changes across generations in the last ~5 years were the introduction of specialized instructions for wider vectors (AVX512/AVX2, and even then you need specialized workloads to use them), [...] AVX2 is actually 13 years (!) old already (first appearing in Intel's Haswell architecture). AVX512 is also nearly 10 years old right now, but not many CPUs actually support it (none of the devices I own support it) | ||