Why do companies like Microsoft or Google develop their own CPUs?

[philb2]

I keep reading where leading Internet companies are building their own chips. I get it that if they use their own chips, they don't need to buy chips from merchant chipmakers like AMD (and Intel). But it is my impression that designing your own chip is extremely expensive and time-consuming, and these companies need to have a relationship with TSMC (or SamSung).

Please correct my misunderstandings.

 

[pendragon1]

extremely expensive and time-consuming

 

[Tsumi]

Typically when a company is designing a CPU, they're developing it to be highly efficient at a few particular tasks. The goal is a stripped down chip that can be manufactured very cheaply and use significantly less power. The hope is that the R&D of said niche chip is less expensive when compared to the cost of purchasing and energy bill of more general purpose chips.

 

[westrock2000]

Against Intel and AMD specifically, you can only make those type of processors with an x86 license from Intel. And Intel isn’t going to give you one of those without some major restrictions and prohibitions. But likely they won’t give you one at all.

Making one yourself, it’s a horrible process. Fabs take years to bring online, 5-10 years to get to a useful numbers of wafers per day with profitable yield numbers. And then you have to poach talent from other companies, which means $$$. And the big fab companies have the money to buy all the tools needed in the fab in order to starve the smaller companies. If Intel and TSMC buy all of ASML’s stepper capacity for the next 2 years, it means the small companies have no option to get the latest technology. Which gives Intel and TSMC a competitive advantage that can last years.

 

[LukeTbk]

Making one yourself, it’s a horrible process. Fabs take years to bring online,

They all go the fabless (like AMD) road and often start with an ARM or other pre-made design that they will tweak, they rarely do the whole thing because of how extremelly expensive, time-consuming/complicated, they would need to become Intel and than have CPU.

As mentionned by Tsumi , a custom CPU will often have specialized hardware accelerated path for specific workload those giant company make all the time and they will co-design those algorithm.

For example:

View: https://medium.com/@christopheradamson253/aws-graviton-processors-performance-and-use-cases-84cb1fad0e3b
https://docs.aws.amazon.com/pdfs/wh...ystem/security-design-of-aws-nitro-system.pdf

They end up being more what look like a soc than a CPU, that integrate many little ASIC for task customres (or amazon themselve) do a lot around the arm CPU:
The main components of the Nitro Cards are AWS-designed System on a Chip (SoC) package thatrun purpose-built firmware. AWS has carefully driven the design and implementation processof the hardware and firmware of these cards. The hardware is designed from the ground up byAnnapurna Labs, the team responsible for the AWS in-house silicon designs. The firmware for thesecards is developed and maintained by dedicated AWS engineering teams.

With custom accelerator for database, decompression/compression, crypto, like GPU in an AMD apu, a bit like the M chips from Apples, AWS gravitron are designed to accelerate their Nitro platform discrete card:

you can think of those card a bit like custom GPU to do common server task instead of rendering 3d graphic and AWS gravitron CPU/SOC that have custom hardware made specially to interact with them (and to not have on the cpu what those handle like virtualisation/hypervisor/etc... which save space and power on the chips, which is also a lot of it not just what those custom CPU have more than a Xeon/Epyc, but a lot of it is what they do not have because they do not need).

A Meta (and its instagram, reel, etc...) custom cpu are made for deeplearning/imagine-video processing, a bit like an amd APU on big SOC that have a CPU with a special chips doing those.

Nvidia is CPU customisation build around their NVLink and unified rack (and at that pace maybe eventually it will be at multi-rack/multi-building) level memory architecture, a giant rack act as a single GPU for the exterior, grace cpu integrate an NVLINK interconnect in their die directly.

 

[MrGuvernment]

Everyone covered it well, I think Meta also did some custom work, or at least they created their own servers/racks and such so that they can swap out / add new nodes more easily, taking up less space, power et cetera.

Consider the sheer scale of hardware these companies have, going the custom route is a "big picture" type view of how much time/effort/maintenance down the road they can save on.

https://www.datacenterknowledge.com/servers/facebook-unveils-custom-servers-facility-design

https://engineering.fb.com/2016/03/...vers-on-performance-without-sucking-up-power/

In Facebook’s cluster architecture, each cluster consists of more than 10,000 servers. Most of our user traffic comes through the front-end clusters, and web servers represent a major portion of the front-end cluster. These web servers run HHVM, an open-source virtual machine designed for executing programs written in Hack and PHP. HHVM uses a just-in-time compilation approach to achieve superior performance while maintaining the development flexibility that PHP provides.

 

[LukeTbk]

sheer scale of hardware these companies have

A 25% less power usage/smaller building-cooling infrastructure density gain could pay for a lot of work here.... for sure.

 

[The Cobra]

Apple created their own Arm-based silicon and it has been really good in our video editing stuff.

 

[MrGuvernment]

Apple created their own Arm-based silicon and it has been really good in our video editing stuff.

Apple worked with ARM to create it I believe, which certainly gives all levels of performance gains vs using off the shelf hardware.

 

[Aurelius]

On top of better-fulfilling specific requirements, there's a larger theme: controlling your own destiny.

Apple is a good example. It got jerked around by Motorola, then again when it switched to IBM, and again when Intel began floundering. It always had to time products around another company's schedule, accept their design priorities, and hope they wouldn't lose interest in Apple as a customer. Windows vendors are experiencing that now with Intel in trouble — the ones that bet the most on Intel are sometimes the ones scrambling to adopt AMD and Qualcomm options.

Apple, Google, Microsoft, and others are now freer to launch on their own terms. Apple can decide what to focus on with its products, and has the most control it can get over when they ship. Say what you will about Tensor G-series chips, but they at least let Google avoid tying the Pixel's feature set and release window to Qualcomm's schedule.

You can look at the mobile industrry for an example of the problem with linking your fate to an outside chip vendor. Heavyweights like Samsung and Xiaomi aren't worried as they know companies like Qualcomm and MediaTek will bow to them, but if you're a smaller vendor... you're hosed. Your product schedules and feature sets are heavily dictated by someone else, and you're probably not going to stand out unless you're very clever (see: Nothing). HTC, LG, Sony, and others have arguably suffered precisely because they had neither their own chips nor industry dominance.

 

[philb2]

Windows vendors are experiencing that now with Intel in trouble — the ones that bet the most on Intel are sometimes the ones scrambling to adopt AMD and Qualcomm options.

Interesting. Which vendors? How is this affecting AMD market shares? Can AMD even keep up with additional demand?

 

[Aurelius]

Interesting. Which vendors? How is this affecting AMD market shares? Can AMD even keep up with additional demand?

I mainly think of companies like Lenovo that almost reluctantly use AMD in certain models. AMD isn’t surging that I know of, but it’s certainly doing well these days.

 

[Sycraft]

Also to add to the others: Sometimes it is arrogance/not invented here syndrome. Facebook has had that going on with some of their shit. I remember reading something from one of their nerds talking about how they had to make custom network equipment because nothing commercially available could meet their needs. That seemed extremely wrong to me, knowing the level of shit that is used in teir-1 ISPs, and I also hard real doubts that they had done what they said.

Ya... no. What they'd really done is to have Broadcom make them some semi-customized hardware, mostly to satisfy ego. It was the same ASICs as everyone else, and they were modifying existing network OSes. It was farm from "we built our own network hardware".


Yet one more, relating to the cloud in particular: Vendor lock in. They want to lock you in to their ecosystem. There are multiple ways they do that (huge data egress fees are a popular one) but this is another way. If your stuff runs on commodity hardware then you CAN just move it to another vendor. However if it was developed for their special, secret, "only we have it" CPU then it requires redevelopment to move it anywhere else, making it less likely you will.