I have been using my local PC with dGPU for about 7h/day of interactive work on average. I might use it for, say, 10 years. At that rate, I will have paid ca. €4,500 for PC and power. For an online service at, let me guess, ca. €1.5/h for a similarly fast GPU performance, I would pay ca. €38,360, that is, 8.5 times the amount. If I upgrade my PC twice during the 10 years for altogether another ca. €2,000, my total expensive will be ca. €6,500. My local PC is very much designed for such intensive, long-term use!

Quote from: RobertJasiek on November 06, 2023, 22:56:32€3/h for _1h/day = _€1,096/year
€3/h for 10h/day = €10,096/year
€3/h for 24h/day = €26,304/year

At which point would you say that upgrading a local PC is financially preferable? :)

When you don't need to rent it every day. The scenarios you are suggesting point to a more server like use case which a local machine isn't designed for.

€3/h for _1h/day = _€1,096/year
€3/h for 10h/day = €10,096/year
€3/h for 24h/day = €26,304/year

At which point would you say that upgrading a local PC is financially preferable? :)

Quote from: RobertJasiek on November 04, 2023, 19:46:48Renting GPUs is an option for a one-time limited period. E.g., A100 at €3/h seems ok. However, in the long term of intensive use, it becomes much more expensive than one's own PC.

Not necessarily. With the rate at which technology is progressing and needing the latest hardware, keeping a local PC up to date can become a costly endeavor.

Renting GPUs is an option for a one-time limited period. E.g., A100 at €3/h seems ok. However, in the long term of intensive use, it becomes much more expensive than one's own PC.

Quote from: RobertJasiek on November 04, 2023, 16:03:21...
In 5, 10 or 15 years, Apple might catch up to Nvidia notebooks for compute applications, presuming software then written for Apple libraries. Until then, Apple is what it is: a company for mainstream editing usage - video, photo, audio, maybe prosumer modeling. Even then, always check if the specific task is supported. (As one should for every system.)

Thanks for extended reply! I have found provided links rather useful. Especially efficiency estimation in "visits per second per watt". I knew that m1-3 chips lack in performance in stuff like matlab, or anything already adapted for rt/tensor cores. But, lack of efficiency compared to cuda is something new to me.

Actually, i would allow my self to note my motivation in considering macbooks. It seems to me, that offloading compute to a rented server is quite reasonable and economical solution in my case. So, i considered a laptop as a "client" machine. But, if performance gap is so big, that everything should be completly offloaded, it becomes a bit questionable strategy.

Apple's problems are not with Intel, AMD, Nvidia. No Apple has a problem with its M1 and M2 being so good that maybe nobody even cares about how good the M3 is. I think both the M1 and M2 have pretty much satisfied the bulk of Mac users no matter what they do. Apple could have spaced out the performance a bit more but their roadmap seems to be one where frequent upgrades of hardware is not needed. I can imagine most Mac users on a M series are plenty satisfied and don't even come close to even using all the performance. Now maybe Apple get's some Intel Mac holdouts to buy into the M3 but I doubt that will save Apple's 4th quarter slide in Mac sales.

Quote from: Alistair Karim on November 04, 2023, 12:48:20

Quote from: RobertJasiek on November 04, 2023, 07:07:16OpenCL, which is 1.81 times slower than Nvidia CUDA libraries and 2.95 times slower than Nvidia TensorRT libraries. Multiply this by the factor of GFXBench's supported APIs versus a speed that should be supported and measurable for OpenCL.

Can you elaborate on that one, please? Do i understand correctly that you imply that those macbooks would underperform in general compute different from videogames? Like, all sort of neural nets or some sort of texture baking?

Apple M has some hardcoded video codecs, which result in reasonable speeds for de-, encoding or sometimes both if a) the specific task uses the specific codec, b) there is enough unified memory storage for the task and c) bandwidth does not create a bottleneck.

I do not include video transcoding in "compute" tasks, which are about machine learning (which includes all sort of neural nets), 3D rendering and other number-crunching applications.

Typically, "texture" is a graphics application and might belong to 3D gaming (we will have to await tests to see how which M3 chip performs) or 3D modeling (in which case perforance depends on the used texture algorithms).

In general GPU-like dependent compute applications, Apple M underperforms very much to extremely both in speed and efficiency. I have calculated it for the application I am familiar with: Go. See:

home.snafu.de/jasiek/AI_Computer.html#mozTocId147960

home.snafu.de/jasiek/AI_Computer.html#mozTocId752162

M3 Max 128 GB unified memory will perform a bit better (still factors worse than an RXT 4070 desktop) than M1 and M2, but the other M3 chips might perform similarly as before (or worse due to the decreased bandwidth for affected applications).

Note the relevance of libraries for proprietary GPU cores (such as Nvidia CUDA and Tensor cores). The same GPU (or, say, a neural component on an Apple chip) is a few times faster with good (e.g. Nvidia) libraries and applications using them. The TDP, die size used for such cores and their numbers give a rough first hint of what to expect from the hardware. Even then, development of such cores takes several to many years. Nvidia has long experience while AMD and Intel fail in this respect. Do not expect Apple to catch up two decades of Nvidia's hardware development quickly.

Furthermore, from now on, expect having to always buy the top Apple M chip to have at least the dream of running behind Nvidia GPUs for compute applications.

In 5, 10 or 15 years, Apple might catch up to Nvidia notebooks for compute applications, presuming software then written for Apple libraries. Until then, Apple is what it is: a company for mainstream editing usage - video, photo, audio, maybe prosumer modeling. Even then, always check if the specific task is supported. (As one should for every system.)

Again the insinuations of the authors. An integrated Apple chip with a low TDP and memory bandwidth 3 times worse than the 4080 cannot be faster. We deliberately chose a couple of synthetic tests, but in games and real 3D applications it was a complete failure, as it should be.

To the shame of Apple, M3 chips have RAM bandwidth (which serves as VRAM) WORSE than in the M2 generation - it was 200 (Pro) and 400GB/s (Max), now it is 150 and 300. With such "progress" and such prices, Apple has nothing left to catch on the market. Sales will begin to fall quickly.

Quote from: RobertJasiek on November 04, 2023, 07:07:16According to its webpage, "GFXBench supports all the industry-standard and vendor-specific APIs including OpenGL, OpenGL ES, Vulkan, Metal, DirectX/Direct3D and DX12."

This is a lie. It claims to support all industry-standard and vendor-specific APIs but it does not even support OpenCL, which is 1.81 times slower than Nvidia CUDA libraries and 2.95 times slower than Nvidia TensorRT libraries. Multiply this by the factor of GFXBench's supported APIs versus a speed that should be supported and measurable for OpenCL.

In other words, GFXBench does not measure compute applications but only measures 3D graphics APIs. Therefore, a general statement about speed relative to 4080 Laptop is invalid - GFXBench results are only valid for 3D graphics applications. Previously, Geekbench results have been mentioned - another heavily prejudiced benchmark and efficiency result.

Do not desinform by Apple propaganda but inform us about the limits of these benchmarks!

Of course it's a GFX benchmark, and it supports all of GFX APIs. If you want compute benchmark, run compute benchmark. No lies and no prejudice here.

Quote from: RobertJasiek on November 04, 2023, 07:07:16This is a lie. It claims to support all industry-standard and vendor-specific APIs but it does not even support OpenCL, which is 1.81 times slower than Nvidia CUDA libraries and 2.95 times slower than Nvidia TensorRT libraries. Multiply this by the factor of GFXBench's supported APIs versus a speed that should be supported and measurable for OpenCL.

Hi. Can you elaborate on that one, please? Do i understand correctly that you imply that those macbooks would underperform in general compute different from videogames? Like, all sort of neural nets or some sort of texture baking?

Let's also not forget that, despite Nvidia's shameful naming scheme, laptops about twice cheaper than an M3 Max MacBook are going to wipe the floor performance-wise with those same MacBooks.

Let's not forget the 4080 laptop GPU is nothing more than a 4070 desktop GPU that has been power limited, it's not related to the actual 4080 at all. Nvidia's naming is shameful.

According to its webpage, "GFXBench supports all the industry-standard and vendor-specific APIs including OpenGL, OpenGL ES, Vulkan, Metal, DirectX/Direct3D and DX12."

This is a lie. It claims to support all industry-standard and vendor-specific APIs but it does not even support OpenCL, which is 1.81 times slower than Nvidia CUDA libraries and 2.95 times slower than Nvidia TensorRT libraries. Multiply this by the factor of GFXBench's supported APIs versus a speed that should be supported and measurable for OpenCL.

In other words, GFXBench does not measure compute applications but only measures 3D graphics APIs. Therefore, a general statement about speed relative to 4080 Laptop is invalid - GFXBench results are only valid for 3D graphics applications. Previously, Geekbench results have been mentioned - another heavily prejudiced benchmark and efficiency result.

Do not desinform by Apple propaganda but inform us about the limits of these benchmarks!