Hacker News Reader

Comment by cma | Hacker News Reader

If you compile large c++ codebases they are a godsend. A slide showed 5950x was only 9% better at a compiler benchmark than the 3950x though. A deeper pipeline based IPC increase may suffer with missed branch prediction in the type of parsing and optimizing workloads of compilers and not get the full benefit.

View on HN

Yeah I noticed that comparison and was surprised at that given the amazing single thread benchmark performance improvement. We really have to see the benchmark.

One might end up in a situation where for developers the older generation cpu being cheaper and more available might make more sense.

I wish somebody really did some suitable benchmark suite for developers.

You don't want older generation CPU. You won't clean/build your project after every change. You'll use some kind of incremental build which will use few cores. Your IDE will use few cores. And significantly faster single-thread performance from new generation will be extremely helpful.

Do C++ developers prefer desktop computers at work? I think a lot of them use laptops.

We are a company that only does development in C++ and everyone has a desktop workstation. Laptops are simply too slow unless absolutely huge and besides my minimum requirement in a workstation nowadays is 64GB of ram with 128GB preferred. Hard to find a laptop that can support that much.

Honest question, why no build farm?

We have a build farm. This is just for running the editor.

Why does it need that much?

Because the data it loads is about 40GB in size? If I then want to actually run locally built client, then it very easily hits 128GB. Games development :P

My MacBook has 64, don’t think 128 is an option.

You can always get roughly 2-3x the CPU performance of a laptop in a desktop (2x core count, higher clock due to much bigger heatsink). It makes sense to use a desktop for the compile times.

The most performance you can get in a mobile CPU right now is an 8 core ryzen. On a desktop you can get the 3990x at 64 cores. That's at least 8x the performance.

Can't fit a 32 core threadripper in a laptop no matter how hard you try. Afaik most use workstation desktops if they're working on large codebases with long compile times (think chrome, Firefox, unreal engine, etc. with compile times often in the hours on slow computers).

I usually find myself developing C++ code directly on servers, via SSH. In that setup, a laptop + external monitor is my preference, because of portability.

If I ever get to do local-machine development again, the choice of laptop vs. desktop really depends on how build times compare, and which platform has Intel CPUs with the required ISA extensions.

Which ISA extensions are useful for compilation?

None really, compilation is considered a branch/integer workload and has been that way for decades. Its a lot of tree traversals, and short string manipulation. One of the bigger boosts in recently memory were the "fast strings (rep mov*)" implementations about 10 years ago.

Single core compilation is actually a fairly good metric for general purpose performance. Of course it scales almost linearly with core count too.

Regarding the speed of compilation, I'm not sure which if any recent ISA extensions matter.

What matters to me is the ISA extensions that are needed by the software I'm building, typically DL or HPC code. Life is easier when I can build and test the software on my local machine, without having to involve some fancy new server.

Many of us who do use laptops do the builds elsewhere.

depends on the project. it takes me about 20 minutes to do a full build on an overclocked 9700k. if I were forced to use a laptop, I would probably quit.

What about spinning up a 24core vm and mount your code with sshfs for those compiles? Edit: corrected ashes to sshfs

this has been discussed internally. for a variety of reasons, testing the software on a remote machine is kinda painful, so we would still have the overhead of copying the executable back to the developer's machine. the more straightforward approach is just to buy fast desktops for all the devs. a desktop with 32GB ram, a top of the line consumer cpu, and a middling gpu is roughly the same price as a good ultrabook anyway.

also, c++ compilation doesn't scale perfectly with more cores. linking is still single threaded, AFAIK.

lld will parallelize your link step. Major improvement over ld. Barely any change to the rest of your build process.

Have you tried it with a Rust project? Is there a real gain? Really curious and willing to try.

sshfs adds far to much latency to large compiles which tend to open a ton of little files, read a KB or two, and close them. If your on a local LAN its doable on NFS or other low latency network sharing, but for everything going over the internet ssh is going to be far to slow. In those cases its better to just git pull the changes and maintain a local copy on the compile target.

What do you mean? The code is on the build server. You only need to push your human-timescale changes.

Is DistCC (or related tech) not a thing anymore?

Do have to wonder if 5950X is that much better than 5900X in GCC?

It should be: all the way up to the 64 core threadripper scales pretty linearly for compiling. It probably won't be worth it over whatever the 3950X price drops to though, if per-core performance is really only 9% better in compiling workloads.

I mean more in price to performance. Will +200 dollars be worth the lower base clock and increased core count?

11% lower base clock, 33% more cores. The 5950X also has 100% more L3 cache.

Reply on news.ycombinator.com