Retiring Almost Everything
My fleet of desktops and servers is at the end of the line.
The arrow of time points just one direction and this week I’ve discovered the reason I’m probably going to retire every desktop and server I own.
Advanced Vector Extensions are a feature that first arrived thirteen years ago in first and second generation Intel Sandy Bridge E5 Xeon processors, which are the type in my systems. This is a SIMD feature set, which allows the application of a single instruction to multiple registers. Processors have a few dozen registers that hold what they are working on in the moment and in this case they are 128 bits wide.
AVX2 was first available in the third generation E5 Xeons, which doubled the register width to 256 bits, and which also ushered in the first use of DDR4 memory. This is the minimum needed to run LM Studio and now and I believe this requirement is going to exist for all of the current AI tools.
Every processor AND every bit of ram I owned in January of this year is end of life as far as my purposes.
The sole exception is the 2019 vintage Dell Precision 7730 that just arrived. The E-2176M processor in it supports AVX2. The 256 bit registers in the AVX2 standard are starting to give way to AVX-512, which is available in HP Z6 workstations. So I can get a Z440 and have AVX2, or some sort of Z6, and really future proof myself.
But it pains me to think of recycling half a terabyte of DDR3 memory …
GPUs As Well:
I’ve never been a PC gamer so the I seldom think of the Nvidia GTX1060 that drives my monitors. This is a 2016 part with first generation tensor cores. The RTX 30X0 have third generation and the 40X0 have fourth gen cores. There were second generation cores but they were only found in datacenter products.
The cores are a secondary issue for AI, onboard memory is the top factor - there’s a correlation between model size and result quality. Neither the 6GB GTX1060 in my desktop nor the 8GB P4200 in the new laptop are sufficient. A 16GB RTX4060 is around $500 while the a 24GB RTX4090 is an eye popping $2,000.
And what if I create something here and want to share it with the world? A server that can take a GPU will be 2U tall and that doubles my hosting cost. If I want to covert all three rackmounts and the RTX 4090 is what’s needed it will be around $9,000 by the time it’s done. YIKES.
Conclusion:
My first computer had an eight bit processor that almost reached one megahertz; I’ve always had to squeeze every last bit of juice out of whatever equipment I own. Being a bit compulsive about system tuning, I’ll dig deep into device specifications as I’m optimizing things. It looks like 2024 is going to involve changing processors, memory, storage, and graphics capabilities. At least I’ll be able to keep my monitors.
All kidding aside, the equipment I have was six years old when I got it and I’ve milked another six years out of it. Excessive frugality can turn to foolishness, so I’m not going to cling to things. I guess it’s time for some changes.
