[alwillis]

>M1 is using standard LPDDRx. >It's not "very high performance".

AnandTech disagrees with you: https://www.anandtech.com/show/16252/mac-mini-apple-m1-teste...

Besides the additional cores on the part of the CPUs and GPU, one main performance factor of the M1 that differs from the A14 is the fact that’s it’s running on a 128-bit memory bus rather than the mobile 64-bit bus. Across 8x 16-bit memory channels and at LPDDR4X-4266-class memory, this means the M1 hits a peak of 68.25GB/s memory bandwidth.

Later in the article:

Most importantly, memory copies land in at 60 to 62GB/s depending if you’re using scalar or vector instructions. The fact that a single Firestorm core can almost saturate the memory controllers is astounding and something we’ve never seen in a design before.

[ianhowson]

Anandtech is comparing M1 vs A14. It's high performance for a cellphone part.

Dual channel DDR3L or DDR4L also has a 128 bit bus. 4200MHz DDR4 is clocked on the high side for most laptops, sure, but it's hardly unusual.

Run the numbers and you get the exact same throughput figure as for M1, which isn't surprising, because we're just taking width * rate = throughput.

So I'll repeat my assertion, downvotes be damned: the memory on the M1 is not special. The packaging and interconnect is interesting. It might reduce latency a little; it probably reduces power consumption a lot. But there's nothing special about it. The computer you're on right now probably has the same memory subsystem with different packaging.

[alwillis]

Anandtech is comparing M1 vs A14. It's high performance for a cellphone part.

That’s where they started, but their conclusion was beyond that.

Did you miss the part where they said the fact that a single Firestorm core can almost saturate the memory controllers is astounding and something we’ve never seen in a design before?

This isn’t only about A14 vs M1.

It’s not that LPDDR4X-4266-class memory is special; it’s been around for a while. What is special is that the RAM is part of SoC and due to the unified memory model, the CPU, GPU, Neural Engine and the other units have very fast access to the memory.

This is common for tablets and smartphones; it’s not common for general purpose laptops and desktops. And while Intel and AMD have added more functionality to their processors, they don’t have everything that’s part of the M1 system on a chip:

* Image Processing Unit (ISP) * Digital Signal Processor (DSP) * 16 core Neural Processing Unit (NPU) * Video encoder/decoder * Secure Enclave

There’s no other desktop such as the M1 Mac mini that combines all of these features with this level of performance at the price point of $699.

That is special.

[ianhowson]

> Did you miss the part where they

I don't think that's notable, sorry. I would expect that of any modern CPU.

> it’s not common for general purpose laptops and desktops

Well, yeah, because "memory on package" has major disadvantages. You (laptop/desktop manufacturer) are making minor gains in performance and power and need to buy a CPU which doesn't exist. Apple can do it, but they were already doing it for iPhone, and they must do it for iPhone to meet space constraints.

I think unified memory is the right way to go, long term, and that's a meaningful improvement. But as you point out, there is plenty of prior work there.

> they don’t have everything that’s part of the M1 system on a chip

They actually do! The 'CPU' part of an Intel CPU is vanishingly small these days. Most area is taken up with cache, GPU and hardware accelerators, such as... hardware video encode and decode, image processing, security and NN acceleration.

Most high-end Android cellphone SoCs have the same blocks. NVIDIA's SoCs have been shipping the same hardware blocks, with the same unified memory architecture, for at least four years. They all boot Ubuntu and give a desktop-like experience on a modern ARM ISA.

> There’s no other desktop ... at the price point of $699

Literally every modern Intel desktop does this.

[sliken]

I've seen latencies when pointer chasing (in a relatively TLB friendly pattern) of 30-34ns. Have you seen similar elsewhere?