Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/google-to-design-own-chips.14600/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Google to design own chips


While initially these chips will be designed for Pixel phones and Chromebooks, lets make no mistake about Google's long term plans here. They are aiming for the data center and at Intel.

This make sense no on a business level.

Do Amazon do such a thing?
 
Apple did start it, building their own chips, but now most system companies are making chips. Look at EDA revenue, more than half is now systems companies. I get notifications on LinkedIn when my connections move jobs. Facebook, Amazon, Telsa and Google, I get that, but Airbnb? What are they up to?
 
Too much Singlish , should be some punctuation n grammar

This makes sense no, on a business level?

They should be making own chips for specifics if they possibly can

Bad week?

Companies contract architecture to a third party and let tsmc do the rest. Google has the money to make their own chips for Chromebooks. No one should be surprised.

After echo 1 and the F-117 creatives were required in stem. The fabless ecosystem has them and Intel doesn't.
 
This make sense no on a business level.

Do Amazon do such a thing?
Yes, they bough a SoC shop which did SoCs for webcams, and threw them on making Gravitron CPUs. Anapruna labs was the name if I believe.

Serving web pages does not require really a lot of CPU power, but more RAM, and cache.

Here the math makes it very obvious that even a cookie cutter ARM V7 SoC beats anything on the X86 market on cost, even taking into consideration the huge premium they pay TSMC for making a low volume prop chip without market (It's not in TSMC's interest to give them a chance against somebody to keep them as a client.)
 
Last edited:
Apple did start it, building their own chips, but now most system companies are making chips. Look at EDA revenue, more than half is now systems companies. I get notifications on LinkedIn when my connections move jobs. Facebook, Amazon, Telsa and Google, I get that, but Airbnb? What are they up to?

The last name I heard was Netflix out of all things.

In our small EE shop, we lost a bit chunk of the embedded team to the car company who got them on relocation offers to Bavaria, 1 year of living expense, doubled-to-tripled salaries, and a high profile immigration lawyer put on their cases.

The situation is getting very common across Chinese tech industry, and especially bad across the part of tech industry mingling with foreign multinationals. Even $100k a year Chinese engineers are still "very cheap" by American, UK, Netherlands, or German standards.
 
Yes, they bough a SoC shop which did SoCs for webcams, and threw them on making Gravitron CPUs. Anapruna labs was the name if I believe.

Not correct. Annapurna was a company which did Arm-based data center Smart NICs.
 
Apple did start it, building their own chips, but now most system companies are making chips. Look at EDA revenue, more than half is now systems companies. I get notifications on LinkedIn when my connections move jobs. Facebook, Amazon, Telsa and Google, I get that, but Airbnb? What are they up to?

And don't forget, if you design your own chip you can start software development in parallel before the chip is finished. A big deal in this software centric world, absolutely.
 
  • Like
Reactions: VCT
The funeral is coming. 60% of all semis are Taiwan as well as 90% of all "advanced semis." There is no interest in continuing the rohler acres relationship.
 
It seems like a tread for the largest tech companies...

The biggest question is who are going to be the winners and losers?

I read that CNBC article, and shook my head. Poorly written by a non-expert. Some primary reasons the big cloud companies design their own chips seem to be:

1. They have sufficient demand within their own company to justify the expense of the design teams and other development requirements (mask sets, etc). These companies can justify vertical integration.

2. TSMC has made using their foundry easy to use by creating a valuable ecosystem of hard and soft IP, development tools, etc. The richness of their environment is, frankly, a wonder in the semiconductor world. (Intel had better have one hell of an ecosystem strategy in mind.)

3. Custom designing chips, as Apple did with the M1 for example, can yield power and performance efficiencies beyond what merchant CPU, networking, and storage chip vendors can usually achieve while trying to please a broader market to maximize sales volumes. Flexibility for a merchant market usually has a price in some sort of inefficiency for any specific application.

4. Because the cloud computing companies have their own software ecosystem, Daniel's issue about software development has an easier solution when the chip design is in-house for these companies. Inclusion in the chip design process with top-down direction is the best possible scenario. Product-specific enhancements are more practical, like the inclusion of DRAM in-package in the M1. Will any merchant CPU vendor consider that? I'd bet against it. Not a big bet though...

5. Some fields of chip design, AI comes to mind quickly, certainly quantum computing too, do not yet have best known architectures that meet most needs. Where there's rich and diverse innovation there's more room for proprietary designs. Post-IEEE Ethernet networking appears to be another tempting field for the cloud vendors.

There are more reasons I think of, but I'm even boring myself.
 
Last edited:
I read that CNBC article, and shook my head. Poorly written by a non-expert. Some primary reasons the big cloud companies design their own chips seem to be:

1. They have sufficient demand within their own company to justify the expense of the design teams and other development requirements (mask sets, etc). These companies can justify vertical integration.

2. TSMC has made using their foundry easy to use by creating a valuable ecosystem of hard and soft IP, development tools, etc. The richness of their environment is, frankly, a wonder in the semiconductor world. (Intel had better have one hell of an ecosystem strategy in mind.)

3. Custom designing chips, as Apple did with the M1 for example, can yield power and performance efficiencies beyond what merchant CPU, networking, and storage chip vendors can usually achieve while trying to please a broader market to maximize sales volumes. Flexibility for a merchant market usually has a price in some sort of inefficiency for any specific application.

4. Because the cloud computing companies have their own software ecosystem, Daniel's issue about software development has an easier solution when the chip design is in-house for these companies. Inclusion in the chip design process with top-down direction is the best possible scenario. Product-specific enhancements are more practical, like the inclusion of DRAM in-package in the M1. Will any merchant CPU vendor consider that? I'd bet against it. Not a big bet though...

5. Some fields of chip design, AI comes to mind quickly, certainly quantum computing too, do not yet have best known architectures that meet most needs. Where there's rich and diverse innovation there's more room for proprietary designs. Post-IEEE Ethernet networking appears to be another tempting field for the cloud vendors.

There are more reasons I think of, but I'm even boring myself.
The simplest reason is that when companies become powerful enough financially and technologically they usually tend to vertically integrate some of the processes. That's what companies like Nvidia, Intel, and AMD are afraid of. Companies like Google and Apple can easily catch up in a few years just because they are financial powerhouses. Even Intel won't really stand a chance IMO.

I just rephrased and repackaged some of your points, but the fact that AMD is supposedly making 128 Bergamo chip should be an indication of how much they fear this scenario.
 
The simplest reason is that when companies become powerful enough financially and technologically they usually tend to vertically integrate some of the processes. That's what companies like Nvidia, Intel, and AMD are afraid of. Companies like Google and Apple can easily catch up in a few years just because they are financial powerhouses. Even Intel won't really stand a chance IMO.

I just rephrased and repackaged some of your points, but the fact that AMD is supposedly making 128 Bergamo chip should be an indication of how much they fear this scenario.

We call those domain specific ICs. We went through this in the early days of semiconductors. Computer companies made their own chips (DEC, Data General, HP, IBM, etc...) because nobody else did. Then came the IDMs Intel, Motorola, TI, etc... which replaced domain specific chips with off the shelf silicon. With the help of TSMC, now that you don't need to own a fab, systems companies are back in the chip business. Generally they start using the ASIC model. Google did this with their first chip with Avago I believe. Samsung ASIC did the first iPhone chips and eSilicon did the first iPod chip. The ASIC business really was the backbone of the semiconductor industry and still does have a big influence with the systems companies, absolutely.
 
Not correct. Annapurna was a company which did Arm-based data center Smart NICs.
Pre-acquisition Annapurna was building SOCs for ARM-based NAS systems, and would be competitor to Marvell and/or Intel in that space. There was some shipping solutions for 4C A15 based designs on IIRC 28HK or maybe 40? But post-acquisition one assumes that yes, they had a hand in Nitro as well as Graviton processors. I'm not aware of a different team doing this or as an ASIC through BRCM/Marvell or others? Google's TPUs have been ASIC programs through BRCM AFAIK, and speculation on this mobile SoC was that it is a co-development with Samsung, and likely that means Samsung does the SoC integration off an Exynos architecture, and Google provides some IP cores to be integrated. But teardowns should reveal this once launched.
 
Bad week?

And yes, Fire tablets.
Fire Tablets..? Try Graviton2, already used in AWS.

Serving web pages does not require really a lot of CPU power, but more RAM, and cache.
Amazon doesn't use Graviton2 to "serve web pages". It can run any of the standard AWS services (including compute). It's meant to be AWS's own ARM server chips and there's nothing slow about them.

Graviton instances are so effective, Amazon's internal software development is now standardized on it.
 
Pre-acquisition Annapurna was building SOCs for ARM-based NAS systems, and would be competitor to Marvell and/or Intel in that space. There was some shipping solutions for 4C A15 based designs on IIRC 28HK or maybe 40? But post-acquisition one assumes that yes, they had a hand in Nitro as well as Graviton processors.
That's not my understanding, as I met with them pre-acquisition. My current understanding is that the former Annapurna team essentially is the Graviton team (and I assume with added resources). Most impressively, Graviton2 is the platform for EC2, which must be a huge cost and power savings for AWS over any merchant solution I'm aware of.

 
Last edited:
Back
Top