Hacker News Reader

Comment by AmVess | Hacker News Reader

40 cores? This thing will be dazzling.

Adding a bunch of tiny cores to an SOC is easy. What’s much harder is writing the software to make use of them.

Nah, that's pretty easy if you do anything involving video. AV1 will eat all the cores.

It’s definitely not easy. In general it is very hard. Some applications have been optimized and for a few applications the fit is very natural (ie, where it is easy to avoid single threaded bottlenecks, or where you don’t need to pass data from core to core via the cache). But for general purpose workloads, it will take a lot of optimization.

Agree to disagree. There's a large host of applications that can benefit from parallel processing power (think, anything that can take advantage of a GPGPU).

I don't dispute that some applications are hard to parallelize. However, in my day to day work, I've found that parallel operations are often not that complicated to introduce.

The bigger issue is that CPUs are now so fast I find myself blocked by IO far more often than I'm blocked by actual processing.

For general purpose workloads maybe, but the target-audience for the Mac Pro is really for people who are using applications that are highly multi-threaded. Just look at the use cases Apple describes the existing Mac Pro for:

> From production rendering to playing hundreds of virtual instruments to simulating an iOS app on multiple devices at once, it’s exceedingly capable.

> Increased speed for Logic Pro X, MATLAB, Adobe Photoshop 2020, Autodesk Maya, Wolfram Mathematica Build Time, Final Cut Pro X

> Tasks like animating 3D film assets, compositing 8K scenes and building lifelike gaming environments.

This is the target market - and pretty much everything listed here tends to be multi-threaded. Making multi threaded applications aren't easy, but for high performance tasks all these use-cases have to be multi-threaded where possible or they are just too slow.

It depends on the type of video processing and whether you can handle adding latency.

True enough. Memory is also often a major constraint.

Absolutely.

I think that people grossly underestimate the amount of processing that can be done on video. Reasonably sized filters, something like deinterlacing, high quality transforms. Swamping a modern CPU in terms of memory access or processing isn't too hard.

Sure, it's likely why AMD 64 core Threadrippers are overtaking the pro video industry at about 1/3 the cost of Apple's current top of the line Mac Pro. 8k/12k uncompressed RAW video is very demanding. It will be interesting to see how this 40 core apple cpu competes with the 64 core. Maybe they'll get a bit of market share back, but it better outperform by more than a few percentage points especially if they will cost more, which is likely.

Not quite that simple: Using AOM AV1, a 12-core Ryzen 5900X beats a 128-core dual EPYC 7763 (same Zen3 architecture) in every encoding benchmark, according to https://openbenchmarking.org/

You have to be smart about how you process the video.

Yes, if you just use AOM directly you'll have a hard time saturating the core with work.

However, if you split your video up into scenes and start an AOM instance per scene, it becomes trivial to outperform the 12 core ryzen with a 128 core machine. The main bottleneck becomes memory bandwidth.

This is what the av1an project is doing https://github.com/master-of-zen/Av1an

All that said, you need videos long enough for there to be a benefit here. The Ryzen will still win if you are talking about making 10 second av1 gifs.

Apple's cores are anything but tiny. The M1, a 4+4 processor, uses 16B transistors, approximately equal to a NVIDIA RTX 3070.

Rust makes writing multithreaded code stupid-easy. The compiler itself will saturate every core of your machine. Rust development on a threadripper is such a joy for this reason, and even the M1 is quite pleasant, so I'm looking forward to these higher core counts.

Reply on news.ycombinator.com