A 7950X is 50% faster than an M1 Ultra and it only costs 550 or so.
It also has 1/8th the memory bandwidth, a MUCH slower iGPU, and takes much more power.
Sure you can add a nice GPU, which helps for GPU workloads, but if it's from the current gen it's at least another $800 and 300 watts.
The M1 Ultra in comparison has the fastest iGPU out, competitive (depending on workload) with some of the mid range GPUs, but generally good enough for things except for gaming. The apple studio is small, quiet, sips power, and has a single large fan that under most workloads is silent.
A 3070 goes for under 500 these days and is a very capable card, certainly significantly more powerful than a PS5. It's apparently around 50% faster than the M1 Ultra's GPU at OpenCL.
The Mac hardware is a great piece of engineering, but for many workstation use cases the latest x86 beats it handily in both performance and price.
The RTX3070 is physically large, does have a nice 448GB/sec or bandwidth, but only for 8GB ram, and of course only for things that can run on a GPU. The 7950X is limited for 83GB/sec (at least without overclocking), which is 5GB/sec per core. The M1 max is 400GB/sec and the M1 Ultra is 800GB/sec, granted the CPUs can only use approximately 50% of that, but as a bonus the GPU can access all of ram, not just 8GB.
So even a MiniITX + 3070 is going to be physically large, power intensive, and hard to get silent, at least until you start getting pretty exotic and expensive with water cooling.
PyTorch and related are supporting Apple's CPU, GPU, and AMX I believe. Benchmarks vary of course. One interesting development is I've just started seeing support for Apple's neural engine and on the models benchmarks it was something along the lines of 7x as fast as the M1's GPU. Something to keep an eye on.
But if gaming framerate per $ is your goal I'd definitely get the 7950x, it's going to love the RTX3070, is cache friendly, and won't care that the 7950x has a 80GB/sec memory interface.
> it's going to take massive investment by intel or AMD to catch up in the price/power/performance category
in the context of desktop workstations is simply untrue. The post compared a Mac Studio to a Threadripper that retails for over $3000, but my point was that it's not a fair comparison: a high-end consumer part that retails for less than 20% of the price also beats the Mac Studio at many use cases, such as code compilation.
There's this weird thing I see where a lot of people think, probably due to Apple's marketing, that Apple Silicon hardware is not just "strictly better" but a gargantuan leap beyond x86. That just isn't true, now that AMD is on a comparable process node. Apple Silicon is better at some things and worse at others, which is great -- competition is good! -- and completely normal.
I'm no mac fanatic, but arguably no good competitors came out for the 2012-era macbook pro (the retina thingy, like I said, not a zealot) in terms of hardware quality. Yeah the CPUs eventually got faster, and Apple drank their koolaid with the bad keyboards and ribbon.
The laptop industry never really responded to that, and I don't really see the ARM competition from AMD/Intel... just excuses about "well they were on the next generation of process".
Like, what process jump previous to this for x86 produced such dramatic gains in efficiency and power? Oh right, certainly none in the last 20 years.
I do hope an ARM M-killer is being worked on. But the absolute silence means it is far away. If it was even a year away there would be rumblings in the downstream Windows/Microsoft market, hardware OEMs, compilers, etc.
Anyone heard anything?
Also, holy crap, $3000 for a processor? That you bought for personal use?
My company got it for me, standard issue here (working on a very compute instensive backend that we run locally)