This is probably the dumbest article I have ever read. The fanboism is strong with this "techtuber".

Quote from: Donny on March 31, 2021, 06:44:31
Too bad probably none of this is true in real world lmao, I know small mobile level cores have made jumps (intel actually makes a good move with it's new mobile) but in real world it's impossible that 15w Trump's a full fledged fucking GPU sorry. Physics literally dictates it's impossible to transfer less power into more power at a certain scale. Efficiency only works up to a certain point, so sure maybe it does more work at a certain small scale level, but clearly it fails at a level of power these weak levels can't match. This is borderline clickbait for fucking apple fanboys hahahahaha. Nobody with brains reads and believes this bllshit at a real world level that almost every human is fucking used to. You can't give me a GPU that kicks a** at 10watts and tell me it's the best when the power level literally dictates it's never gonna size up in the real world.

At least you don't have to wait long to see how wrong I am. ;) And I did say... this is all speculation. :) Have a great day!

Too bad probably none of this is true in real world lmao, I know small mobile level cores have made jumps (intel actually makes a good move with it's new mobile) but in real world it's impossible that 15w Trump's a full fledged fucking GPU sorry. Physics literally dictates it's impossible to transfer less power into more power at a certain scale. Efficiency only works up to a certain point, so sure maybe it does more work at a certain small scale level, but clearly it fails at a level of power these weak levels can't match. This is borderline clickbait for fucking apple fanboys hahahahaha. Nobody with brains reads and believes this bllshit at a real world level that almost every human is fucking used to. You can't give me a GPU that kicks a** at 10watts and tell me it's the best when the power level literally dictates it's never gonna size up in the real world.

I like multiplication. And turtles. ;D

The average 3990X runs at 4.1GHz when running Geekbench 5.
The average 3970X runs at 4.2GHz when running Geekbench 5.
And you think it is a mainly a thermal issue not a scaling issue? The 3990X jumps up over to nearly 35K points when run with a different scheduler on Linux.

Quote from: Gawamatsu on March 30, 2021, 11:14:56
Is it good for mining? Miners should go for it and leave PC gaming alone, for gamers...

Apparently the hash rate on Etherium is 2mhs while using 17 watts. The RX580 does about 30 mhs at 73 watts , so not even good for that. I question how powerful the M1 GPU is with that score.

Quote from: Ayoh on March 30, 2021, 13:05:57
Why stop there, why not predict the performance of a 128 core MX CPU and a 25,600 core Apple GPU - it will totally destroy the competition

What about a 256 core MZ CPU? wow it will annihilate the competition.

Now it just leaves the trivial matter of manufacturing the chips and we are done

I decided to stop at just the stuff that Bloomberg has already reported that they have in testing right now, Just seemed sensible.

Quote from: Mate on March 30, 2021, 13:26:43
It looks like many of those tech tubers forget that chip manufacturing cost rises exponentially with die size. M1 is already big, manufacturing M1Z as one chip like now would increase costs way above every CPU ever created for laptop. Ok, they can go for chiplets but there biggest obstacle are AMD patents - as they are only company that is doing it on mass scale.

Completely agree with you here, which is why in previous videos we've been talking about either using multiple SOCs, or more likely more modular systems like the M1 is now, the CPU, GPU and Unified Memory are on Apple's own custom fabric layer, so they may be able to just use more of those modules. We already know they're testing up to the 32 core version on TSMC's 5nm process. But yes, of course as you scale the yields drop.

Quote from: _MT_ on March 30, 2021, 19:17:44

Quote from: Johannes S on March 30, 2021, 18:03:45
I'd like to point out that the wonderful multithread scaling Geekbench 5 exhibits:

AMD Ryzen Threadripper 3990X 2.9 GHz (64 cores) scores 24954
AMD Ryzen Threadripper 3970X 3.7 GHz (32 cores) scores 22445

As I wrote, you could expect to see something like this when a processor hits power limit. Processor essentially turns energy into computations. Once you hit power limit, you can only generate more points by increasing efficiency. It's not necessarily a scaling problem.

Agreed - The 20 core Apple Silicon processor assuming the power draw is roughly in line with what we have now will have a TDP around 100W vs the Threadripper's 280W.

Quote from: Johannes S on March 30, 2021, 18:03:45
I'd like to point out that the wonderful multithread scaling Geekbench 5 exhibits:

AMD Ryzen Threadripper 3990X 2.9 GHz (64 cores) scores 24954
AMD Ryzen Threadripper 3970X 3.7 GHz (32 cores) scores 22445

As I wrote, you could expect to see something like this when a processor hits power limit. Processor essentially turns energy into computations. Once you hit power limit, you can only generate more points by increasing efficiency. It's not necessarily a scaling problem.

Quote from: Johannes S on March 30, 2021, 18:03:45
If I naively extrapolate the 5950X's performance of the Zen 3 core scaled to 64 cores, like this YouTuber did, then a hypothetical 64 core 5990X would score 68200! Of course we all know this scaling is not the case.

It gets even better with 5800X as a base: 83752.

I'd like to point out that the wonderful multithread scaling Geekbench 5 exhibits:

AMD Ryzen Threadripper 3990X 2.9 GHz (64 cores) scores 24954
AMD Ryzen Threadripper 3970X 3.7 GHz (32 cores) scores 22445

If I naively extrapolate the 5950X's performance of the Zen 3 core scaled to 64 cores, like this YouTuber did, then a hypothetical 64 core 5990X would score 68200! Of course we all know this scaling is not the case.

It looks like many of those tech tubers forget that chip manufacturing cost rises exponentially with die size. M1 is already big, manufacturing M1Z as one chip like now would increase costs way above every CPU ever created for laptop. Ok, they can go for chiplets but there biggest obstacle are AMD patents - as they are only company that is doing it on mass scale.

Why stop there, why not predict the performance of a 128 core MX CPU and a 25,600 core Apple GPU - it will totally destroy the competition

What about a 256 core MZ CPU? wow it will annihilate the competition.

Now it just leaves the trivial matter of manufacturing the chips and we are done

I'm not sure how seriously I can take an OpenCL benchmark where A100 scores less than 3080 and Titan RTX scores less than 2080 Ti. That would need some explaining.

Also, when you look at the Threadripper family, 3990X isn't much better than 3970X and that in turn isn't much better than 3960X. The gap between 3960X and 3950X is larger and more proportional (you've got 50 % more cores for 40 % higher score). Which puts into question how well the benchmark scales above about 24 cores. However, when I look at 3900X and 3800X, scaling is not that great. 3900X has 50 % more cores, but scores only 33 % more than 3800X. And 3950X has 33 % more cores, but scores only 20 % more than 3900X. In other words, 3960X has three times the number of cores (+200 %), but scores only 123 % more than 3800X. I could imagine seeing something like this when you hit power limit (the improvements would actually be in efficiency resulting from lower frequency). But then I would expect a bigger jump between 3960X and 3950X. Which would bring into question what exactly is the power budget. Even the 3700X isn't far behind (I think it's losing something like 5 % on 3800X).

I think at this point, I would find it difficult to predict how well the CPU benchmark scales. Even when I look at single-core vs. multi-core, the scaling isn't what I would expect. I'm assuming single-core test runs actually a single thread, bypassing SMT. Even at low core counts where power budget is less of a factor. With the exception of M1. M1 displays much better scaling than I would expect from seeing the other results. It's closer to 8 core processors than 4 core processors. That doesn't match up to Cinebench, for example. Which would suggest it's not a question of power. If you want to have a laugh, according to Geekbench, M1 has higher multi-core score than 5800H. Even 4800U should wipe the floor with it as long as the benchmark scales well. It does not compute in my brain.