Finally time to get a PC for SW developement.

[Psi]

If you're the type of developer that has a very rapid 'code / compile / test' cycle then you might want to investigate if your build environment/compiler will benefit from more cores or faster cores and where exactly the tradeoff is.
Choosing a CPU that will build faster might be worthwhile paying for and that comes down to how efficient your compiler is at multithreading. Having unused cores is kind of pointless.

[brucehoult]

I bought my first personal laptop two years ago for traveling, a Lenovo Yoga 6 (AMD 7530, 8 GB) 13.3" for $450.
I love it. Most here would probably find it inadequate.

Needs vary hugely, by orders of magnitude. It's impossible to make any recommendation without fully understanding the exact task to be performed.

My travel machine is a Lenovo E14 Gen 2 AMD with 6 core 4500U with 8 GB RAM. Your 7530u is about 40% faster. The 4500U is perfectly fine for general web browsing or working on relatively small and simple development projects — it's likely fine with anything that is your own from-scratch project.

But spend $1500+? No, thanks.

For other needs $1500 is absolutely trivial. My i9-13900 laptop cost around $1500 on discount after the next model was already out. I thought it was an absolute bargain. The alternative was an M3 Max 14 core 36 GB RAM for $3500, or 16 core 48 GB for $4000.

I remember paying $3300 for a 266 MHz single core G3 with 64 MB RAM, and thought that was good. My 32 core ThreadRipper cost $4600 to build. I got my money's worth from that too as my main computer from early 2019 to early 2024.

It's shocking to me how reluctant programmers (or their employers) are to spend money for equipment for people who they are paying $100k+/year. $1500 too much? It's ridiculous on the face of it, if spending more improves productivity by even a couple of percent.

[brucehoult]

If you're the type of developer that has a very rapid 'code / compile / test' cycle then you might want to investigate if your build environment/compiler will benefit from more cores or faster cores and where exactly the tradeoff is.
Choosing a CPU that will build faster might be worthwhile paying for and that comes down to how efficient your compiler is at multithreading. Having unused cores is kind of pointless.

You need to know your own needs.

Linux kernel builds and LLVM builds scale linearly out to at least 64 or 96 CPU cores (maybe more, that's the most I've tested with). GCC and other GNU software with sequential recursive make steps, each with an autoconf, not so much ... maximum overall speedup on a GCC build is about 9x even with a lot more cores than that (but more like 5x if you only have 8 cores).

Building C# in Visual Studio? I really have no idea how many cores it can use profitably. I'd hope all of them.

[chilternview]

What should I choose?

Something with a lot of cores.

I have a PC built 10 years ago, twin E5-2687W Xeon cpus with 24 cores each, so 48 logical processors. And 256Gb DDR4 memory (bought when memory was a lot cheaper). Since Visual Studio can make use of those cores while compiling a large-ish project with a few thousand files split into 20 dll's, it's still reasonably fast today, although my Mac M3 laptop feels faster.

[Psi]

Yep, if the compiler can peg all the core you throw at it to 100%, then that will be a better option.
It may actually be better to get much older generation chip with double the cores vs a modern chip for the same price but less cores. That may make thing faster but also cheaper if it moves you from DDR5 to DDR4. But you might have to weigh that against getting a modern chip with better performance running VMs.

Not all compilers can peg all cores, hence why it's important to know what your compiler will do with the CPU you give it.

Sometimes you are locked in to using a proprietary compiler which often sucks.

[brucehoult]

Not all compilers can peg all cores, hence why it's important to know what your compiler will do with the CPU you give it.

Sometimes you are locked in to using a proprietary compiler which often sucks.

Every compiler I've ever seen uses ONE core.

It's the build system, specifically a program such as gmake, cmake, ninja (even when those are kicked off from a fancy GUI IDE), that decides how many copies of the compiler to run at the same time. And this in turn depends on how your build scripts are written and how good your dependencies are.

[nctnico]

If you're the type of developer that has a very rapid 'code / compile / test' cycle then you might want to investigate if your build environment/compiler will benefit from more cores or faster cores and where exactly the tradeoff is.
Choosing a CPU that will build faster might be worthwhile paying for and that comes down to how efficient your compiler is at multithreading. Having unused cores is kind of pointless.

You need to know your own needs.

Linux kernel builds and LLVM builds scale linearly out to at least 64 or 96 CPU cores (maybe more, that's the most I've tested with). GCC and other GNU software with sequential recursive make steps, each with an autoconf, not so much ... maximum overall speedup on a GCC build is about 9x even with a lot more cores than that (but more like 5x if you only have 8 cores).

Building C# in Visual Studio? I really have no idea how many cores it can use profitably. I'd hope all of them.

It will probably be slow. Compiling microcontroller firmware using GCC on Windows takes like 5 to 6 times longer compared to Linux with the same GCC on the same PC with the exact same make file. It is all about the speed of task switching.

[Smokey]

If you can live with a desktop, check out https://www.ibuypower.com

Overpriced, only have windows home edition.
Separate video cards adds a cost I do not require.
8gb base versions are a joke.  Units with 32gb are over 5k$.

Huh?  Their pre build machines are cheaper than if you bought all the parts separately.  And they certainly have windows pro. 
My friends and I used to always build machines from parts and we all get iBuyPower stuff now because it's significantly cheaper and has warranty and all that. 

[Psi]

Not all compilers can peg all cores, hence why it's important to know what your compiler will do with the CPU you give it.

Sometimes you are locked in to using a proprietary compiler which often sucks.

Every compiler I've ever seen uses ONE core.

It's the build system, specifically a program such as gmake, cmake, ninja (even when those are kicked off from a fancy GUI IDE), that decides how many copies of the compiler to run at the same time. And this in turn depends on how your build scripts are written and how good your dependencies are.

Yep, i was using the word loosely/incorrectly, my bad.

[brucehoult]

Huh?  Their pre build machines are cheaper than if you bought all the parts separately.  And they certainly have windows pro. 
My friends and I used to always build machines from parts and we all get iBuyPower stuff now because it's significantly cheaper and has warranty and all that.

Yeah, I used to physically build machines myself until I found places that would assemble and test parts of your choosing for you for around US$60/NZ$100.

Quite apart from not using an hour or two of my time, I build about one machine every five years while they're probably doing five a day. So they're better at it. And if some part turns out to be DOA then they have others already on hand that they can swap out to find out which part the problem is with, at zero hassle to me. I don't HAVE spare parts of the current generation, and it's a PITA and time (and latency!) sink to RMA things.

My Threadripper was built by adamant.com when I lived in the US. My i7-6700K was assembled by a vendor in Moscow. My i7-4790K was assembled by PBTech here in NZ.

In the 2000s and 2010s ascent.co.nz would actually assemble your selected parts FOR FREE, on the basis that this cost them less than tech support for people who did their own assembly. They stopped offering this around 2021, and now instead offer a 2% discount if you buy all parts for a complete computer using their online configuration tool. This at least ensures that customers order compatible parts.

[udok]

Well, here I am.  I will unfortunately need to work with VisualStudio.

VS can be a highly productive development environment if you do not fight it mentally.
Visual Studio Code is a good alternative too.
If you want you can use the VS compilers from the MSYS2 command line.
Set the INCLUDE, LIB and PATH environment variables from a bash shell and call "cl -O1 -c file.cpp"

This is what I was recommended:  (I will consider a Mac lateron when compiling for iOS and MacOS)
Ryzen 9 9955HX
64GB ram
2tb SSD
Windows 11 Pro with WSL2 and Ubuntu 24.04 LTS

This is overkill in my opinion.  I have a 32 GB Laptop and do not use more than 14 GB with Win11 and normal development work.
For developement work it is important to have at least 6 fast cores for parallel compilation.
Other factors are not very important.  I prefer Intel CPUs as the the single core performance and the DDRAM bandwidth is better
(see https://www.igorslab.de/en/jedec-ddr5-9000-intel-arrow-lake-refresh-ryzen-9000-x3d-ram-scaling/5/).

More cores do not help more because Laptops have a limited thermal budget. 
After the PL2 limit (28 seconds on my Dell) the CPU clock is reduced and the more cores the faster the thermal throttling.
I have a Dell with an i7 and with an i9.  The i7 is 5% slower but it is much quieter and stays cooler.

Select a laptop with at least 45 Watt and with two 80 mm fans in an aluminium housing.
I have very good experience with the Dell Precision 15" series (5530 up to 5570).
The successor is the Dell Pro Plus 16 which should be ok too, but it does not have the centered touchpad.
If you use fast DirectX or 3D CAD you may need a good Nvidia graphic card too.

Would I really save on a mini PC with comparable specs?

No. The mini PC has the same thermal problems. 

If you do not need the mobility of a Laptop buy a normal PC.   
They have better performance especially if you are compiling larger projects.
No thermal throttling and they do not use space on your deskop and they are cheaper.

[SteveThackery]

I will unfortunately need to work with VisualStudio.

Making my project a .net and using modern coding agents means VisualStudio.
Visualstudio means I must use Win11 pro somewhere.

If you allow yourself, you might - shock!, horror! - actually like it. Visual Studio is an immensely capable application, fine-tuned and polished over three decades. As for W11, if you decide in advance that you'll hate it, then of course you will. Approached with an open mind, you'll find it to be perfectly tolerable. After all, a billion users get their work done on it.

[macboy]

I just put together a computer mainly for SW Dev. It is a dual socket Intel LGA2011-3 with DDR3, which is highly unusual since that socket is normally DDR4. I went for the unusual DDR3 config because I had a drawer full of it from the "before times", and can't justify the cost of new DDR4 for what is frankly a toy. .

Only a subset of the E5-26xx V3/V4 CPUs work with DDR3. Fortunately, the cream of the crop is one of them, E5-2696v4. That is an OEM part (not found in Intel Ark) which is identical to the flagship E5-2699v4. It is a 22 core 44 thread CPU with 55 MB cache, 3.7 GHz max, and 2.7 GHz all cores active. So the dual socket setup has 44 cores / 88 threads, and 8 channel memory giving about 110 GB/s bandwidth. Not stellar but not too shabby. Of course these XEON CPUs cost many thousands each  originally but now maybe $60 each. Chinese random brand motherboard around $100

I can compile the Linux kernel (with Debian x64 config) in about 90 seconds. Single threaded performance is frankly rather meh, but with 88 threads, it adds up to something. The 7zip benchmark score is over 1.5x as high as my much newer, much more costly, Core Ultra 7. I was also playing with Qwen 3.6 35b which runs at about 24 Token/s output rate entirely on CPU using ik-llama. Again, not stellar but given the cost of the hardware, really respectable. Due to the low-ish single core performance and high thread count, this system needs workloads that can make use of many threads. Mine do, so I'm happy.

[brucehoult]

E5-2696v4

150W TDP. Owww. And you're running two of them. 300W.

Qwen 3.6 35b which runs at about 24 Token/s

That's probably near maximum power too. So likely 12W/tok/s.

My little RISC-V SBC only gets 4.5-5 tok/s on that model (Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf), but it's also only using 14W doing it, also entirely on CPU (the eight 2.0 GHz dual-issue in-order A100 cores).

[macboy]

E5-2696v4

150W TDP. Owww. And you're running two of them. 300W.

Qwen 3.6 35b which runs at about 24 Token/s

That's probably near maximum power too. So likely 12W/tok/s.

My little RISC-V SBC only gets 4.5-5 tok/s on that model (Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf), but it's also only using 14W doing it, also entirely on CPU (the eight 2.0 GHz dual-issue in-order A100 cores).

Yes 300 W TDP, which doesn't necessarily mean they use that much power in reality. But in this case they do, especially when using SIMD-type instructions.

Token rate is bound by memory bandwidth. Your token rate is exactly as expected for a single channel of decent DDR4. I found that when running AI, I can limit the engine to far fewer CPU cores and/or turn down CPU clock speed, both of which significantly reduce power consumption without reducing token rate. So then about 6 W/Tok/s, "merely" double your power consumption. While 4.5-5 Tok/s is impressive at 14 W, that token rate feels unbearably slow in an interactive session. When you get over roughly 20 Tok/s, it meets or exceeds most people's reading speed, so it feels much more natural to use. My main use for this computer is definitely not AI. It was just a fun experiment to see what very cheap old hardware could do with a decidedly modern workload which normally requires hideously expensive modern hardware. It is roughly on par with a mid range GPU+CPU, for both power and token rate.

[brucehoult]

I can limit the engine to far fewer CPU cores and/or turn down CPU clock speed, both of which significantly reduce power consumption without reducing token rate.

Oh yes I forgot about that. Adding "-t 4" (instead of the default increased it from 4.6 to 5.3 tok/s on the same query.

I must be a slow reader, but I find 6-7 tok/s from a slightly smaller model quite comfortable on general queries such as "Tell me about Dolly Parton". Of course it's slow for slinging large amounts of code around.

[onesystem]

Look into MSI or other gaming laptops, in used condition, like 1-2 years old. Although you get them used, their specs will fall within what you are looking for, will still run latest windows and will handle all of the requirements pretty well. Unless you are dead set on getting the brand new equipment.

[SiliconWizard]

So things have become so crazy that now even professionals are resorting to buying second-hand gear? Sad days.

[nctnico]

So things have become so crazy that now even professionals are resorting to buying second-hand gear? Sad days.

Performance increase had slowed down a lot. The laptop I bought 12 years ago became too slow last year so I finally bought a new one. And my decade old daily driver PC might be up next. But it is still quick enough for compiling stuff. The age is a stronger reason (reliability) to replace than the performance. But right now is not the right time to buy a PC. RAM prices are huge. The laptop I bought is from Dell and typically I buy Dell computers with the smallest memory and upgrade myself. But this time round buying the laptop with the amount of memory I need was cheaper than buying memory seperately.

Back in the old days buying a new CPU + motherboard every year gained you like 2 to 3 times the speed.

[SiliconWizard]

Oh, certainly it's less worth the upgrade these days, so the matter is more like if you really need to buy/build a new computer because you need one extra or your old gear has failed. This is when you start crying...
This it the path I took on my main workstation, a couple years ago when I decided to max it out (buy RAM to the max supported, buy new SSDs, and change the CPU to the highest supported on the motherboard, which had become cheaper) rather than buy a brand new machine that would have cost a lot more for modest performance increase compared to what I got.

But that was before the great inflation.

These days, even doing that may become out of budget for many.

[brucehoult]

So things have become so crazy that now even professionals are resorting to buying second-hand gear? Sad days.

Oh gosh, I've always done that! In both computers and cars and iPhones there are some people who simply MUST always have the latest model, even when it's only 2% better than last year's model. I take advantage of them.

I bought my "Late 2011" quad core i7 17" MacBook Pro used in early 2014 and then travelled the world with it almost continuously until 2020. I did augment it with a pocketable NUC with i8-8650U in 2018 but the 2011 remained my laptop.  Similarly I bought a used PowerMac 8500 120 MHz I think a year old (so 1996), and then used it for a long time, into the 2000s, only changing the CPU from 604 to 604e to G3 to G4 until I bought a year old dual core 2.0 GHz G5 PowerMac in 2004.

The rule is to buy the first model of a generational leap, with a large performance increase, when the 2nd gen with a 5% improvement comes out.

So for example buying a used M1 when the very little improved M2 came out would make sense. Though in that case I couldn't contain myself and got a 16GB M1 Mac Mini the instant they came out in late 2020. (and it's still my main computer for interactive use, web browsing etc, almost six years later with zero plans to change it)

Another thing I do is buy last year's model, brand new and unused, but at a discount. I got my first dedicated Linux machine that way ... a Pentium Pro 200 at a huge discount when the Pentium II was already out. Apparently the Pentium II ran legacy 16 bit code much better, but I was running only 32 bit code...  And my current workhorse laptop, an i9-13900 model from 2023, I bought at a large discount in early 2024 when the i9-14900 machines (which were only a 200 MHz speed bump on a 5.4 MHz single core speed ... 4%) were already available. Oh, and my 2019 model Honda CRF250L Rally motorcycle, bought at a large discount (NZ$7250 vs originally $12000) when the 2020 model was already next to it in the shop.

Mildly used or new old stock ... both good buying when any changes since that model are purely incremental.

[broly]

E5-2696v4

150W TDP. Owww. And you're running two of them. 300W.

Qwen 3.6 35b which runs at about 24 Token/s

That's probably near maximum power too. So likely 12W/tok/s.

My little RISC-V SBC only gets 4.5-5 tok/s on that model (Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf), but it's also only using 14W doing it, also entirely on CPU (the eight 2.0 GHz dual-issue in-order A100 cores).

Yes 300 W TDP, which doesn't necessarily mean they use that much power in reality. But in this case they do, especially when using SIMD-type instructions.

Token rate is bound by memory bandwidth. Your token rate is exactly as expected for a single channel of decent DDR4. I found that when running AI, I can limit the engine to far fewer CPU cores and/or turn down CPU clock speed, both of which significantly reduce power consumption without reducing token rate. So then about 6 W/Tok/s, "merely" double your power consumption. While 4.5-5 Tok/s is impressive at 14 W, that token rate feels unbearably slow in an interactive session. When you get over roughly 20 Tok/s, it meets or exceeds most people's reading speed, so it feels much more natural to use. My main use for this computer is definitely not AI. It was just a fun experiment to see what very cheap old hardware could do with a decidedly modern workload which normally requires hideously expensive modern hardware. It is roughly on par with a mid range GPU+CPU, for both power and token rate.

no opinion on the bandwagon AI flavour-of-the-day, but i did appreciate the discussion around dual socket motherboards.

after enough computations, i successfully scorched my mac pro 5.1's cpu board/northbridge by performing my computation that relies on large shared memory.

it got so bad that what used to be quick took forever because the data from each "half" had to sync before returning the result. that's when i looked at the board and saw heat patterns around specific pins.

i got lucky and scored an 18 core imacpro for 600 loserbux that had a crack in the screen (not noticeable during usage though). added 256GB to it and it has tided me over for the past two years. can't complain.

but in general i approve of your approach. i would have done something similar if i needed the cores. right now i'm waiting for the new intel stuff and i'll likely be moving to windows from APPUL.

[Siwastaja]

The days of overkill configuration to run a huge IDE thinking that magically turns into good code are fortunately over.

You can develop your mp3 player with 16GB of RAM just fine with any non-broken IDE/code editor. 32GB RAM machine is a total beast. That was cheap 1.5 years ago which is why everyone took it, not because it was a necessity. Now the question for sensible people is between 16 vs 32GB, not between 32 vs 64 anymore. 64GB is for really specialized work with special workloads that require memory. Your mp3 player development isn't one.

If you use AI to vibe code, that reduces the requirement even further. Claude Code runs with some hundreds of MB of memory fine. I do a lot of development by running it on tmux, with any screen resolution. Sometimes even from a freaking Android phone with some SSH app.

[BrianHG]

Well, It looks like I also need a laptop for offsite demos.
Preferably one with Headphone out, HDMI, and Lan port.  Of course, some USB ports as well...

It only needs to last me ~18 months at most, but may still be used after that period.

I dont care abut screen quality or battery life, but the ability at add a second flash drive for a separate Linux boot would be appreciated.

As stated above, 32gb is only to have Win and WSL going with 2 screens and a browser with multiple tabs & VS running.  I'll be using incremental builds under a minute each a lot, so I am not expecting to bake the CPU.  I dont mind waiting for the occasional full-re-builds.

I'm going to order at the end of this week, so now is the time to narrow my selection.

[brucehoult]

Once again I'll say I'm very happy with my 16" Lenovo Legion Pro 5i 2023 model (i9-13900 CPU) with 32 GB RAM.

You can still buy the i9-14900 version on lenovo.com, but of course there are also Core Ultra versions now. Or the 16 core AMD.

It has all the ports you mention, both USB-As and USB-Cs (including PD and DP). There's an empty socket for a 2nd SSD (or you can fill it when you order).

Mine lasted 6 hours on light use on the original Windows 11 before I wiped it or now 5 hours on Ubuntu. I never need more then 2-3.

When at home I use it as my main work beast, screen closed and ssh -X / nomachine from my Mac Mini which I do all my keyboarding on. It's faster in every way than the 32 core 2019 Threadripper 2990wx I still have but now seldom turn on.

And it cost me $1500 as the previous year's model on lenovo.com, which I thought was a great deal.