News for hackers. Stuff that matters.
Re:In Defense of RAM by RamRodification on Friday November 03, 2023 @12:18PM (#38130910)
> Embedded RAM, like or not, is the future of CPUs unless you want memory performance/power efficiency to freeze near current levels

Why is that?

Re:In Defense of RAM by DannyBee on Friday November 03, 2023 @12:40PM (#38131241)
It's remarkably difficult to maintain signal integrity/etc if it's not very close to the CPU at the signal rates they are getting to.

This is already a serious problem.

For places where you are willing to give up power efficiency, or latency, or both, this is not as big a deal - this is why you see CXL/etc.

But for a laptop, where people want power efficiency, latency, etc, this is a serious issue.

Re:In Defense of RAM by brucethemoose2 on Friday November 03, 2023 @12:34PM (#38131151)
Because the signal integrity across the pins and traces becomes a serious bottleneck.

SODIMMs are already "past" the limit, requiring 1.35V for modest DDR5 speeds at terrible timings, hence vendors are trying to migrate to this:

https://www.anandtech.com/show/21069/modular-lpddr-becomes-a...

I suspect regular desktop/server DIMMs are getting close to their limit as well

Essentially, the closer the RAM sits to the CPU, the faster/lower latency/lower power it gets. This is especially important in graphics heavy chips, where the wide, high frequency busses can eat loads of power. In fact, I wouldn't be surprised if Apple transitions to something "new" like a newer generation of Samsung's Wide I/O if they can get memory vendors to produce it in bulk.

Re:In Defense of RAM by bryanlarsen on Friday November 03, 2023 @12:30PM (#38131087)
M2 pro/max/ultra bandwidth is 200/400/800 GBps.

Ryzen dual channel bandwidth is 100GBps. EPYC 8 channel bandwidth is well under 400GBps.

Re:In Defense of RAM by brucethemoose2 on Friday November 03, 2023 @12:40PM (#38131243)
And more importantly, M2's energy/bit is pretty low.
Re:In Defense of RAM by matja on Friday November 03, 2023 @12:57PM (#38131541)
* Shared bandwidth between CPU and GPU - as soon as you use up one, you're bottlenecking the other.

* No M2 memory benchmark I've seen gets above ~250GB/s once cache effects are exceeded.

* Random access latency is >~110ns which is no improvement over even mid-range DDR4 designs from 5 years ago.

TL;DR: It's about cost

Re:In Defense of RAM by sliken on Friday November 03, 2023 @01:03PM (#38131647)
Lets compare. The m3 max has up to 128GB of ram, 400GB/sec bandwidth, and uses that memory for both CPU and GPU. It fits in thin (under an inch thick) laptops and is efficient enough to have great battery life on the order of 10-20 hours depending on use cases. So efficient that performance is identical when running on battery or wall power, unlike many of it's competitors. The power savings is even more impressive once you consider you don't need a discrete GPU for many use cases.

The AMD Epyc has 12 channels of DDR5-4800 and manages 460GB/sec. It's a physically large chip, can handle a ton of ram, but also uses a ton of power. You can fit a crippled version (without all the memory channels) on a micro atx motherboard that's several inches thick. Normally you'd need an ATX, EATX, or rack mount motherboard. It needs several 100 watts for just the socket, not to mention the numerous dimms (12 minimum for 12 memory channels). It's the kind of thing you see in a data center, after you put ear protection on.

If you compare the M3 max, which fits in a laptop or small desktop, to Intel and AMD competition that fits in a thin laptop you end up with 1/4th the memory bandwidth. The competition either matches the power use and is much slower, or is as fast or faster at a huge power penalty. AMD and Intel laptops need discrete GPUs to compete and often require wall power to run at full speed.

All in all the M3 max is an impressive chip, fast, and very efficient. But part of that is the soldered in dimms that allow for great bandwidth and low power use.

Re:In Defense of RAM by brucethemoose2 on Friday November 03, 2023 @01:15PM (#38131824)
This is Apples and oranges.

A better AMD comparison is Van Gogh (the Steam Deck chip) to the base M1, and on paper they are very similar. The M1 is just a more "premium" version of the budget AMD chip.

AMD could make a good M3 competitor (and maybe even improve on it by tiling the memory controller/cache like the 7900), but... There is just no market for it. Windows OEMs dont want a big expensive SoC, they want a cheap CPU to hang a power hungry graphics card off of. Hence the successor to Van Gogh was canceled because no one wanted it.

Re:In Defense of RAM by sliken on Friday November 03, 2023 @01:26PM (#38132020)
AMD/Intel iGPUs stink, mostly because the standard 128 bit wide memory system from the lowest end Intel/AMD laptop/desktop chips to the high end like the Ryzen 9 7950x have the same memory bandwidth.

AMD even makes better memory systems for the XboxX and PS5, which run games great. For years gamers drifted from gaming desktops to consoles because GPUs were insanely priced or unavailable.

However, finally, many years after the GPU shortage there's GASP going to be a laptop/desktop chip with iGPU with more than a 128 bit wide memory interface. The AMD Strix Halo. I hope it does well, it's roughly similar to the M2 Pro. Still half a M2 Max or 1/4th a M2 ultra.

Re:In Defense of RAM by brucethemoose2 on Friday November 03, 2023 @01:40PM (#38132250)
> The AMD Strix Halo

Eh, I was excited at first, but rumors suggest its going to be a tiled, power hungry SoC with high clocked cores. Its like the desktop-class 7000 series monsters they stuff into laptops, but with a decent IGP hanging off... All performance and zero efficiency.

I have higher hopes for Arrow Lake: https://videocardz.com/newz/intel-arrow-lake-p-with-320eu-gp...

Re:In Defense of RAM by sliken on Friday November 03, 2023 @02:02PM (#38132542)
Dunno, AMD is making progress with it's APUs, they seem pretty competitive and well reviewed. AMD does seem to be getting more design wins in the laptop space. Makes me optimistic for the next gen.

Hawk Point, Strix Point, Strix Halo, and Dragon Range look to have a decently wide range of performance and power points coveraged. The top end will have the X3D (extra cache) and the next tier down will have double the memory bandwidth.

Or maybe pigs will fly and Apple will lend a hand to Asahi to get Linux working well. Not that Asahi hasn't been doing a good job, I do wonder how long the new GPU features will take to show up in Linux.