Frédéric JacobsNov 20, 2023

Hyperscalers such as Microsoft, Google & Amazon are deploying an increasing amount of custom silicon.

But what's the case for actually doing that? Any good reads on how that silicon is more competitive than Qualcomm/Intel/AMD/... off-the-shelf silicon?
Is it the inability to get large supplies of ARM chips at competitive prices?

Show threadterenceNov 20, 2023

@fj most of the hyperscaler SoCs seems to be using just off the shelf designs (ie. ARM reference designs). And at their scale, it’s probably cheaper per unit than through a vendor.

It also changes the SoC design prioritities. An off the shelf part needs to be profitable as a standalone product for the vendor. A custom in-house design doesn’t need to be profitable on its own, as long as it is profitable as part of the services it provides.

Show threadDogeMocenigo
@terencey @fj The CPU cores are often Arm designs, sometimes with a few custom optimisations, less often so completely in-house implementations. The advantage here is that the companies can still integrate them in custom SoCs taylored to their own use cases.
Nov 21, 2023 at 8:56amWeb
Show threadFrédéric JacobsNov 21, 2023
@dogemocenigo Do you have a sense of what tayloring they are doing with custom SoCs in practice, that justify the additional complexity compared to taking off the shelf chips?
Show threadDogeMocenigoNov 21, 2023
@fj I cannot say much (obviously), but an example could be the on-chip interconnect: it may not make much sense to produce and sell chips with a crazy high bandwidth for the general purpose computing market (unless you are a fruit company from California), but if you need it for your own server farm with HPC capabilities, then it makes sense to build it yourself. Also, you may not need the onchip GPUs, NPUs and may want to deploy something from, say, NVIDIA. So, you may even save on area.