SoylentNews is people
SoylentNews is powered by your submissions, so send in your scoop. Only 15 submissions in the queue.
SoylentNews
On Youtube I watched a Mac user who had bought an iMac last year. It was maxed out with 40 GB of RAM costing him about $4000. He watched in disbelief how his hyper expensive iMac was being demolished by his new M1 Mac Mini, which he had paid a measly $700 for.
In real world test after test, the M1 Macs are not merely inching past top of the line Intel Macs, they are destroying them. In disbelief people have started asking how on earth this is possible?
If you are one of those people, you have come to the right place. Here I plan to break it down into digestible pieces exactly what it is that Apple has done with the M1.
Related:
What Does RISC and CISC Mean in 2020?
This discussion has been archived.
No new comments can be posted.
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
cashtration, n.:
The act of buying a house, which renders the subject financially
impotent for an indefinite period.
(Score: 0) by Anonymous Coward on Thursday December 03 2020, @11:18AM
x87 stack architecture was the most efficient way to implement it as a coprocessor without gobs of additional glue logic. Today, it doesn't matter: The CPU detects the common pattern of push/pop and optimizes it away. None of those "memory" operations for x87 ever hit actual memory today under typical circumstances.
"Unified" memory is just marketing blather for shared memory. It's true that some low-end CPUs have to copy data between CPU and GPU. The Raspberry Pi used to work like this, for example. It makes sense: The CPU used in the Pi was intended for TV set-top boxes where all the real work happened in the GPU and you just needed a small CPU on the side to do things like talk network protocols. You were probably streaming video over the network and the chip would copy the data directly from the network into the GPU part of memory, so it didn't matter what happened in the CPU's memory area. Now that it's used in a mini-PC, I think the latest drivers have eliminated or at least are working toward eliminating this problem. Architectures designed for serious graphics performance using shared memory - such as game consoles - certainly never had this sort of limitation.
The real problem with shared memory that kills performance isn't copying textures from one part of RAM to another. It's that GPU and CPU have conflicting requirements. The GPU needs RAM that's very fast and optimized for sequential access. The CPU needs RAM that's not quite as fast but much more random in its access. And they both need plenty of bandwidth and power. PS5 and other serious shared-memory architectures use GDDR memory for the whole system, and the CPU just has to live with being second fiddle. That's fine because consoles do most of their work on the GPU and the CPU has gobs of cache and just doesn't need the system memory all that often. PS5 can stream data right off its SSD into graphics memory. What bottleneck? PCs that care about graphics performance use a discrete GPU with its own memory that, sure, copies data because it's a separate bank of RAM, but which means the CPU and GPU can each have exclusive access to their own memory, optimized for their own uses and without having to share bandwidth except when they load textures (which doesn't happen too often). The other reason is that discrete chips also have discrete power connections and discrete heat sinks. This part doesn't matter as much in a laptop where you're limited by battery and getting the heat out of the case, but desktop gamers, crypto miners, protein folders, and neural net scientists use discrete GPUs and blow all the integrated setups out of the water.
M1 is a medium-grade mobile GPU, intended for rendering the desktop. Up against AMD or nVidia, it competes with laptop or low-end desktop graphics from two generations ago. A current-tech APU runs away from it, and it's an order of magnitude behind the Playstation 5, which itself still lags the new discrete GPU cards released for PC from both major players at pretty much the same time. M1 will run FaceTime just fine, but it is not any kind of wonder chip, except in the power consumption area. This isn't useless, but it benefits its battery life, not its performance.
It remains to be seen if this neural processing unit means anything. If it does, Intel and AMD will copy it. There's a reason they have recently purchased the two biggest FPGA makers, and it's to hedge against things like this.