[aix1]

> Google has an incentive to go all in on CPU design as that helps their data centers and GCP both, so we might see significant gains over time.

It seems that CPU design for phones and for data centres are quite different problems, or do you see them as facets of the same problem?

[welkie]

Apparently they used to be quite different problems, but with today's focus on "performance per watt-hr" as the most important metric, they've merged together. GCP's new AMD CPUs are popular because they're x86 but do more per watt. And all the big cloud platforms are looking into ARM64 chips for the same reason. Arm architecture is crucial in phones. Voila.

[actuator]

At least my understanding was that the SoC would be different but the CPU core of phone SoC itself would be similar with the laptop/desktop/server die. The server SoC would probably feature more cores, with more memory interfaces and higher clock speed due to more thermal allowance.

At least that's what Apple seems to have done between M1 and A series processors.

[qmarchi]

I'd imagine that if Google were to ever start deploying ARM servers, it could be a big boon. Amazon Web Services and Oracle Cloud already have ARM servers available, and Azure is supposedly working on their own ARM chips, wouldn't be that much of a stretch.

(Disc: I work for Google, but not in Cloud nor Pixel)

[the-rc]

It doesn't need to be public facing servers. It could be also for internal uses such as the storage nodes or YouTube transcoding (which has custom decoding/encoding HW as well). I doubt they'd keep the Tensor subsystem for applications like those, though... Except perhaps for YT, where one day you could feed models with frames as they get converted.

[ksec]

>It seems that CPU design for phones and for data centres are quite different problems

Still limited by the same TDP. Just as data point and prospective Apple spend more TDP budget on its mobile CPU core than any current ARM Server CPU core.

[jsjohnst]

> Just as data point and prospective Apple spend more TDP budget on its mobile CPU core than any current ARM Server CPU core.

Are you talking TDP per core? Because there are many ARM server processors with much higher total TDP than Apple's A/M series processors. Even per core, I find it hard to believe that's true for even most ARM server chips, let alone all, given the total TDP of an M1 (10-15W, 4/4 big.LITTLE cores) or an A14 (6W, 2/4 big.LITTLE cores) chip.

[ksec]

Per Core, The Max TDP per core of M1 pushes to 5W. All current ARM server based on N1 has a per core maximum of 3W max. You have specialised core like those on Fujitsu A64FX but they dont push to 5W either with when running with SVE.

i.e What ever core Google decide to work on tensor could easily be used on Server as well. Whether they are the best for it would be a different question though.

[jsjohnst]

> All current ARM server based on N1 has a per core maximum of 3W max.

The N1 based Altra Q80-33 and the Q64-33 both have a per core max over 3W, with the Q64-33 approaching 3.5W/core. Still way under your figure for M1 of course.

[wmf]

The Nuvia Phoenix core is designed for servers but Qualcomm is putting it in phones. That didn't work well for Mongoose but we'll see.