Hacker News Reader

Comment by inasio | Hacker News Reader

Assuming the M1 performance extrapolates to desktop power/cooling it's going to be a monster chip. If that is the case I think apple will not stop at having the fastest watch/phone/tablet/laptop/PC. Why not go after the datacenter and leave money (a ton of money) on the table?

View on HN

How you intend to extrapolate that? A seemingly much larger thermal/cooling headroom usually does not gain you that much in absolute performance. Consider the mobile/desktop Zen3 Ryzen geekbench (S)ingle/(M)ulti scores:

- 5800U, mobile (10-25W), 1400(S), 7000(M)

- 5800, desktop (65W), 1600(S), 9000(M)

- 5800X, desktop (105W), 1700(S), 11000(M)

and

- M1, mobile (10W), 1700(S), 7000(M)

Everyone's going gaga over the fact that the passively cooled M1 can trade blows with big desktop towers. But mobile Zen3 is not that far behind those towers either, so I think much more thermal headroom buys you less than people assume.

One of the big reasons the M1 is so fast is its RAM is on-die. Latency and throughput to memory are dramatically better and you don't have to drive those long traces, saving power as well. Servers need to scale to memory sizes in the terabyte range. Apple will need a new design that uses conventional memory before they can go after the big workstation market, let alone server.

It's not on die arghhhh why does this misinformation keep happening. It's on package. It's two separate LPDDR4X chips. Latency is not dramatically better. Other current LPDDR4X laptops run the same clock speeds off package. Maybe Apple has more aggressive timings than them, but the difference would be tiny.

Reality distortion field.

> Why not go after the datacenter

For certain applications, the migration will be challenging. It's also a slowly dying market. Amazon has an ARM instance, already. I'm sure they're already tearing the M1 apart to see what they can use. It's also no Apple's core competency; enterprise sales is a different beast.

Edit: I wasn't quite clear: it's dying because businesses are migrating to the cloud, and cloud providers are already working on/support ARM CPUs. Google might be willing to buy M1 chips. Not servers or boards because everything is probably custom by now, just the chips.

In what way is the datacenter market dying? Are Googles and Facebooks and Cloudfares and Microsofts and Amazons suddenly decommissioning huge racks of computers while their cloud businesses take off? What is that cloud made of?

Those all owns their datacenters (and some can afford to design their own CPUs for those). Renting racks isn't dying yet, but it is on a downward trajectory.

Maybe because every tier 1 cloud provider is starting to design its own CPU. If this is the trend there will be little left to sell them.

Can you name a cloud provider that has recently launched their own CPU?

Graviton's been out for a while.

https://www.zdnet.com/article/microsoft-is-designing-its-own...

I haven't seen anything on this from GCP, but Google builds so much in-house, I'm sure they've at least considered it.

AWS is on its second in-house ARM chip. You can launch an M6g/R6g/C6g right now. If you have workloads on the cloud, you should be benchmarking already, as they will likely be cheaper for any given workload.

Reply on news.ycombinator.com