prismatics

I'm somewhat surprised why Apple didn't take the AMD Threadripper devices. 

Key factors:
PCIe 4.0
64+ PCIe Lanes each supporting Infinity Fabric Link to AMD GPUs (effectively reducing latency and allowing for perfect workload scalability using PCIe4.0 SerDes)
(soon) up to 64 cores and 8 memory channels
extremely cheap price for a workstation class CPU (32 core for 2k$)

Hell they could even go full house with two EPYC 7742 64-core processors (for a total of 128 cores that have combined 256 PCIe 4.0 lanes (or an equivalent of 530% bandwidth compared to the intel part) that cost retail even combined less than what Intel asks for its 28-core part while consuming about a third of energy per unit of performance when compared against the Intel device in question. Imagine performing tasks with such a setup, the render is just a click away, almost real-time!

The Intel platform looks kindof End-of-Life by now. It's specs look years old. The socket does not support PCIe4.0 speeds, it doesn't supply enough memory channels for acceptable bandwidth scalability going above 32 core. I'm sure they considered Threadripper, there must be very special contracts that prevent Apple from using the better product. I really hope Apple doesn't design themselves inside a corner with this.

Mac Pro as a platform currently kinda looks obsolete for seriously professional computing looking forward considering the Threadripper parts are proven to be a big time saver by the time it ships. It doesn't seem to be a viable investment so I'll be waiting for their next Mac Pro and meanwhile continue using the Threadripper workstation I have built last year that is still faster.

pujones1 said:

I’m thankful for the freedoms that we here. They really get taken for granted sometimes. 

This could be a Cyber Security nightmare. Wonder what Apple will do. Pull out or stay on? They have to obey the laws of the countries they operate in. Yikes!!

Don’t worry, they will obey just fine. That’s no biggie, just include some free apps from the App Store that can be deleted and that’s it.

Soli said:

I thought they did that for use when you went through customs?

That’s China, not Russia.

cpsro said:

prismatics said:

What is this special kind of fake information? What do we have L1 I/D, L2, L3 caches, prefetch, TLBs and branch predictors for? Do you honestly believe that 2-way SMT as well as out of order execution and a very high cache hit rate (which by the way makes 4-way SMT unsuitable for consumer application) contribute to the CPU idling in a way where it is "mildly warm"?

Sorry, random pattern memory access _never_ happens except if you have hash tables, Directed Acyclic Graphs or something but you are not using a CPU for that kind of problem, you are using GPUs or, if you have a very special application, Content Addressable Memory (or CAM for short, which is almost never used except for core network routers where you have some billion packets per second to be switched).

You can't see dark matter. Is it fake?
You write random memory access _never_ happens and then proceed to describe instances where it can. (Congratulations!) People don't always use GPUs for the techniques you describe either.

As stated previously, the throttling does not occur on a lowly Mac mini.

This is not dark magic, you have performance counter registers for that.

Please kindly educate yourself about the topic you write about.

cpsro said:

The new thermal design has done nothing to avoid the throttling that occurs when memory larger than the L3 cache is beaten on in a random pattern, as opposed to sequentially the way common benchmarks access it. Multiple recent generations of Apple laptops do this and the latest model is no different. The CPU power consumed in such situations is a small fraction of the TDP, the CPU is driven well below the advertised sustained speed for all cores, and the processor temperature is only mildly warm.

What is this special kind of fake information? What do we have L1 I/D, L2, L3 caches, prefetch, TLBs and branch predictors for? Do you honestly believe that 2-way SMT as well as out of order execution and a very high cache hit rate (which by the way makes 4-way SMT unsuitable for consumer application) contribute to the CPU idling in a way where it is "mildly warm"?

Sorry, random pattern memory access _never_ happens except if you have hash tables, Directed Acyclic Graphs or something but you are not using a CPU for that kind of problem, you are using GPUs or, if you have a very special application, Content Addressable Memory (or CAM for short, which is almost never used except for core network routers where you have some billion packets per second to be switched).