New AMD Zen 3 and Ryzen processors

[Cardoso]

Why not wait yet one more year for Zen 5

He has a point though: if you need to buy everything (motherboard, ddr4 and cpu) waiting might have sense because next year you could buy:
1) motherboards with a new socket which will support Zen 4, 5, ... while AM4 is EOL with Zen 3
2) new DDR5 memory. When you decide to upgrade your Zen 3 build, the DDR4 will be of no use

So far I've never re-used a motherboard or RAM for a new build. My current desktop is a 6-core i3930K with 64GB DDR3 from 2012. That's what you get when you keep waiting for the next new thing to come out

There's a lot of wisdom in wating. Today people (specially young people) want instant gratification.
I had a core i7 920 from 2009, only in the beginning of this year I upgraded to a Ryzen 3900x. I'm really glad I saved a lot of money during these almost stagnagted 11 years of not so great cpu upgrades. Since the 920 was a quad core and I wanted at least 8 core for the next upgrade, I think the long 11 years wait paid off.
About Zen 4 I think I got the release year wrong, it looks zen 4 will be 2022. In 2021 there probably should be a minor zen 3 refresh still on 7nm.

[brianr]

I also have an old i920 that I was able to comfortably overclock to 4GHz (from 2.6GHz).
Over the years I added more RAM, SSD, faster GPU (GTX770 to 1070 to 2060), etc.
It was fine for most things.
My even older Q6600 (also easily OC'd, but so old no popcnt) would get the "hand me down" components from the i920.

However, lacking some more modern instructions limited my ability to upgrade TensorFlow past v1.5.
Given the rapidly changing machine learning libraries and library inter-dependencies, and often lack of backwards compatibility, this became a significant problem when training NNs after a while.
Also, I think there is better instruction level parallelism with the newer CPUs.

A year ago I upgraded to a used 1920x (12 core), although it was already a generation+ old and the next gen AMDs were expected fairly soon.
As my primary usage is for net training, I chose to spend more on an even faster GPU and got a 2080ti (although the Nvida 30x0 series was also talked about). Additional costs were a better power supply and more SSD and HDD.

I have been very pleased overall.
My biggest issue has been learning to cope with Windows 10, which is incessantly intrusive and frustrating, although I have the Pro version and have it "locked down" as much as I can. By comparison, it was pretty painless to move from Ubuntu 16 to 18.

[mvanthoor]

There's a lot of wisdom in waiting.

That can indeed be true. My previous computer lasted for 8 years.

The reason why I'm now wanting a new computer is because I got into chess programming, and the first version of my engine is nearing completion. After testing it, it needs to be retested again and again, after each new function I add. With current-day CPU's, these tests can be done at 1 minute + 5 seconds or something like that (20 years ago that would have been out of the question), but they still take a lot of time. Faster CPU's allow short(er) time controls without making the engine play like crap, and more cores allow more concurrent games.

Ás I'm not convinced with regard to AMD's BMI2/PEXT/PDEP and AVX2 implementations, which I possibly will be using somewhere down the line, I'm not going for an AMD CPU. I don't want to start using a function and then seeing my engine get slower and thus not be able to test it. If at all possible, I'd like at least a 16-core Intel CPU with enough RAM to give each engine a hash of 1024 MB. (32 GB of hash tables + OS stuff and other programs; that'll come down to 64 GB of RAM, up from my current 32 GB.) A computer like that will probably last at least 6-8 years.

The longer you wait, the more you'll probably get for the same money.

[Joost Buijs]

There's a lot of wisdom in waiting.

That can indeed be true. My previous computer lasted for 8 years.

The reason why I'm now wanting a new computer is because I got into chess programming, and the first version of my engine is nearing completion. After testing it, it needs to be retested again and again, after each new function I add. With current-day CPU's, these tests can be done at 1 minute + 5 seconds or something like that (20 years ago that would have been out of the question), but they still take a lot of time. Faster CPU's allow short(er) time controls without making the engine play like crap, and more cores allow more concurrent games.

Ás I'm not convinced with regard to AMD's BMI2/PEXT/PDEP and AVX2 implementations, which I possibly will be using somewhere down the line, I'm not going for an AMD CPU. I don't want to start using a function and then seeing my engine get slower and thus not be able to test it. If at all possible, I'd like at least a 16-core Intel CPU with enough RAM to give each engine a hash of 1024 MB. (32 GB of hash tables + OS stuff and other programs; that'll come down to 64 GB of RAM, up from my current 32 GB.) A computer like that will probably last at least 6-8 years.

The longer you wait, the more you'll probably get for the same money.

Of course this is true, but then you will wait forever. As a rule of thumb I buy new hardware every three years or so, I can afford it and I can't think of any reason why I shouldn't. Especially if you want to do something with ML or other types of AI, having decent hardware will help you a lot.

[mvanthoor]

Of course this is true, but then you will wait forever. As a rule of thumb I buy new hardware every three years or so, I can afford it and I can't think of any reason why I shouldn't. Especially if you want to do something with ML or other types of AI, having decent hardware will help you a lot.

True enough; I can afford a new computer every three years easily, but I don't want to. I hate building and installing new computers since I started to work full time. I don't have the time or incentive anymore to tinker with the hardware and/or drivers. Therefore I build the most powerful computer which I think can still be called "just short of completely idiotic" with regard to cost, and then stick with it for 6 years at least.

My current computer was quite expensive because of the graphics card, which I wanted at the time for The Witcher 3. Now I play even LESS games than I did back then, so I even plan to put my current 4 year old GTX 1070 into the next computer, as it is still more than fast enough to run every game I have. The money saved on an expensive graphics card will go to the CPU, RAM, and mainboard. (I basically never replace those; I only add memory, if necessary. I leave half the slots free when building a new computer.) If I ever get into neural networks running on the GPU (only if I at least train the network myself, and better yet, write the code), I'll see what graphics cards will be available at that time. That will be at least the successor to the current RTX 3000 series.