Apple M1 - How did we get here...

Anthony Medici wrote:

E.J. Peiker wrote:When the more powerful 16" models come out, possibly using the higher end M1 that is under development, I would expect discrete graphics with its own memory and 32GB as options.  

Why would you want discreet memory when by putting more shared ram on the chip, the GPUs and CPUs don't need to spend the time copying things from one memory pool to another?

The memory used by the GPU is due to its own needs for computation and rendering - there is no copying from one memory pool to another.  In fact by having memory dedicated for graphics, the CPU and GPU can work autonomously without needing to schedule memory calls, they both have simultaneous access to their own memory pools.  The memory is already off chip, the 8 or 16GB of RAM is not on the M1 processor.  Currently if the CPU and GPU need memory, they have to negotiate it.  If you want to keep it as an SOC where the GPU and CPU are on the same chip you can do that but for maximum performance you still want dedicated memory to each that can be simultaneously accessed.

I'm reading that the ram is on the M1 chip. Do you have sources that have stated that the memory for the M1 is off chip?

Anthony Medici wrote:I'm reading that the ram is on the M1 chip. Do you have sources that have stated that the memory for the M1 is off chip?

 Memory is on the chip:

https://www.apple.com/mac/m1/

Joe

Anthony Medici wrote:I'm reading that the ram is on the M1 chip. Do you have sources that have stated that the memory for the M1 is off chip?

That's just the cache.  All CPUs have some SRAM cache which is MUCH faster than DRAM primary memory and is used to insure that the processor does not suffer from too many wait states.  It has 16MB of SRAM in two caches.  If you actually built 8 or 16GB of DRAM on a CPU, the die size, even with 5nm technology would be huge.

In general DRAM and CPU uses very different semiconductor processing.  It would be really complex and expensive to put DRAM on a CPU chip and neither would be optimized.

You can also just look at the chip itself and see that there is no DRAM.  You would have to have a field of perfectly repeated structures that is bigger than the M1 chip itself on the die - you can see a photo of the die here:
https://www.apple.com/mac/m1/

In that photo, you see two separate chips on the mini board to the right of the processor - those are the DRAM memory modules.

Finally, it says the M1 has 16B transistors.  How do you have 16GB of RAM in 16B transistors?

Here’s a link to an article on the M1:

https://www.zdnet.com/article/apple-m1- ... nd-memory/

Joe

E.J. Peiker wrote:

Anthony Medici wrote:I'm reading that the ram is on the M1 chip. Do you have sources that have stated that the memory for the M1 is off chip?

.............
In that photo, you see two separate chips on the mini board to the right of the processor - those are the DRAM memory modules.

Finally, it says the M1 has 16B transistors.  How do you have 16GB of RAM in 16B transistors?

I know nothing about computers, but it seems to me in the color diagram, it shows the DRAM modules on the chip, not separately. The text seems to indicate that as well. Am I just misreading something there?

Joe

SantaFeJoe wrote:

E.J. Peiker wrote:

Anthony Medici wrote:I'm reading that the ram is on the M1 chip. Do you have sources that have stated that the memory for the M1 is off chip?

.............
In that photo, you see two separate chips on the mini board to the right of the processor - those are the DRAM memory modules.

Finally, it says the M1 has 16B transistors.  How do you have 16GB of RAM in 16B transistors?

I know nothing about computers, but it seems to me in the color diagram, it shows the DRAM modules on the chip, not separately. The text seems to indicate that as well. Am I just misreading something there?

Joe

That's a diagram of the system which is all on that little board, not a diagram of the actual processor die.  If you go further down the page where they describe each quadrant of the processor, you can clearly see that.  Trust me, there is no DRAM on that die, just the SRAM cache.  The DRAM is the two black modules to the right of the processor.

The M1 package contains more than one die -- the memory is stacked and is in the same package.


Brian

Brian Stirling wrote:The M1 package contains more than one die -- the memory is stacked and is in the same package.


Brian

Yes, the actual M1 Arm processor and GPU, which contains no DRAM is the visible die.  The two DRAM modules are to the right in this picture and are separate memory chips (die).  They are then coupled on this small system board which forms the "computer" part of the system that then things like system storage, ports, etc are attached to...

I wonder though, what effect data transfer speeds are -- not the best case but the typical case. With a typical PC the memory reside off the CPU die and require a bus with signal lengths on the order of 10-30cm or maybe longer whereas with the M1 package the signal lengths might well be 1-2cm. All things being equal the shorter path should permit faster transfer speeds. And, that does count the even longer path length to the GPU via PCIe.

Yes, once the GPU has data to work on then all the memory transfers can be done on a separate memory set, but, you have to get some data to the GPU and then some from the GPU via the PCIe.

Some of the real world testing of the M1 based laptops, including video editing and rendering, which is extremely CPU/GPU and memory intensive appear to show, even in emulation mode, that it's a pretty fast system. Once native software for apps like video editing/rendering is available eliminating the emulation penalty it will be interesting to see more honestly where it is. And, while the M1 isn't Apples first silicon this is the first generation of this new line and let's see what a 32 core M1 can do in a year or so.


Brian

Oh definitely, there are some serious potential memory access advanteges to this approach - expandability isn't one though and that would have to be addressed in an M1 chip for something like a Mac Pro.

The expandability thing is just Apple being Apple -- you want more will sell you a new Mac with what you need! There appears to be a roadmap for the high end that might include 128 cores and if memory scales approximately, then we could see 64GB to over 128GB of RAM, though fitting that into a single package would be a tall order. But, if that is possible, and over time it would be, the potential speed advantage, particularly if the GPU's have dedicated RAM, could well be quite substantial. One alternative would be a unified memory approach were all memory talks over the same bus but a redundant bus permits the GPU to access a portion of that memory without stealing bandwidth from the unified bus.

One other thing that should have Intel, AMD and Microsoft worried is the use of this silicon technology in server farms were being 5X or better more efficient could save boat-loads on electricity and cooling. In fact, I can see a variant of the M1 with higher core count and greater memory as well as large SSD's forming the core of that new server farm. I have to believe Apple is aware of that use case and is working on it so it is interesting that there's been no public talk about it.


Brian

Some objective benchmarks that show that maybe all of the superlatives for the M1 are overhyped just a bit. Yes on a power per watt basis the M1 is great but it is nowhere near the best performing CPU/GPU combination for laptops. Be careful when you look at these graphs and be sure to read the text as in some of the graphs a larger number is better (Cinebench for example) and in some a smaller number is better (Premiere, Photoshop, Gigapixel). For example the M1 holds it's own on the Photoshop benchmark but I would never want to run Topaz Gigapixel AI on it - you might as well let the uprez run overnight. Some interesting reading here for those that are interested in this stuff:
https://www.pcworld.com/article/3600897 ... -4000.html

I have no doubt that the current M1 will not fare well against high end CPU/GPU combinations from Intel or AMD for some benchmarks and real world applications, but what I was getting at is the roadmap that indicates upwards of 128 cores with an unspecified RAM quantity and configuration. What now seems apparent is that Apple has had this game plan from a long time, perhaps dating back to when they switched from IBM/Motorola to Intel, to begin to make there own silicon so that one day they could push Intel out the door. Apple isn't going to do that until they have Intel beat in most all metrics and that day isn't here ... yet.

So, I expect Apple to keep chugging along upping the capabilities and squeezing everywhere they can until there's nothing left to squeeze.

I mentioned in a prior that not only is Intel in trouble but that Microsoft is as well, but recent news from Redmond indicates they are working on there own silicon -- no idea when they started that push but best guess is that they are new comers and have a decade or more to be where Apple is now. Frankly, Samsung would be a more likely player with alternate silicon for higher end boxes given there semi talent.


Brian

Intel today replaced CEO Bob Swan, with Pat Gelsinger. Pat was the CEO of VMWare and widely seen as one of, if not the best tech CEO in the USA. Pat worked for many years at Intel in the 80's, 90's and early 2000's. I worked with him on numerous projects including the 486 and original pentium - he on the design side, me on the process side. Even back in 2013 when then CEO Paul Otellini (RIP) retired I wrote on the Intel retiree board that they should back a Brinks truck up to Pat Gelsinger's house and start unloading money bags until he succumbs and says yes to the CEO job offer.

I truly believe that if anyone can bring Intel back from the serious slump it has been in, it is Pat. For an Intel retiree who wants the company he invested 27 years, this is excellent news and also for anyone that is into technology or a user of technology as almost all of us are...

Do you think Intel can be competitive against the Apple's ARM architecture ? Would Intel need to fundamentally change something, or would catching up with TSMC's 5nm (or is it 3nm now) be enough to get serious performance gains and be back in the race ?

lelouarn wrote:Do you think Intel can be competitive against the Apple's ARM architecture ? Would Intel need to fundamentally change something, or would catching up with TSMC's 5nm (or is it 3nm now) be enough to get serious performance gains and be back in the race ?

Can they be?  Absolutely given that leadership makes that decision firmly without being wishy washy and letting go some of their past paradigms.  Will they be? I have more confidence now with Pat Gelsinger at the helm where my answer would be a firm no with Swan at the helm.  The question will come down to if the board gives him the leeway to do what must be done, even if that means scrapping a big part of IA or relegating it mostly to the server space.

The other side of the argument:
https://www.tomshardware.com/news/intel ... benchmarks