Arm releases Chiplet System Architecture spec beta version | Hacker News Reader

Arm releases Chiplet System Architecture spec beta version

newsroom.arm.com

66 pointsfork-bomber23 comments

Can someone explain, why this is needed and why ARM is involved?

I understand the desire for standardization in the Chiplet space, but why is this something ARM is concerned about? Does the transition to Chiplets impact ISA? I would have assumed it being transparent from a system architecture's perspective, a mere technical detail, nothing within ARM's scope.

More than the ISA, its the memory interconnect that require standardization. At SoC level, ARM is already a de-facto standard (ACE-Lite, CHI, ...), but its only a standard for communication inside a chip, to interconnect varius IPs.

I guess this standard aim to keep being a standard interconnect even in multi-chiplets system, to create/extend the whole ecosystem around ARM partners.

I'm guessing but it may make it easier to sell CPU cores if you can sell physical dies instead of IP blocks.

Anyway, this is presumably no different than why Intel took an active role in the specification of what a motherboard looks like.

The strange thing to me is why ARM, which is about chips with small area, where you don't need chiplets, cares.

If you do get involved with chiplets, which ARM says will be "because AI" (sigh), you need physical and logical interconnect standards, something ARM has been doing for a long time for SoC-internal connections: the AMBA bus standard.

ARM have been moving away from chips with small area for a long time (see server SoC which are huge beasts), and are trying to become the standard platform for everyone trying to have custom hardware.

In this space, chiplets makes a lot of sense: you can have a compute chip with standard arm cores which is reused across your products, and add an extra chiplet with custom IPs depending on the product needs. That is for example what (as far as I'm aware) Huawei is doing: they reuse the chiplet with arm cores in different product, then add for example an IO+crypto die in the SoC in their routers/firewalls products, etc.

Have you seen the "Ultra" versions of the Apple Silicon? They are gigantic chips. And many other competitors make server-class ARM based processors, so having a chiplet architecture as part of the ecosystem makes a lot of sense.

Not to mention the GB10 where Nvidia mates a 20 core arm chip to a blackwell chip and puts it in a widget the size of a mac mini.

> ARM, which is about chips with small area

Huh? Is that their goal? Recent high power ARM chips do seem to have quite some area and chiplets enable increasing power further, which, looking from the outside, seems much higher on the agenda than keeping chip area small.

Arm has been enabling server/data-center class SoCs for a while now (eg Amazon Graviton et al). This is only going to pick up further (eg Apple Private Cloud Compute).

Also, there's nothing fundamentally stopping chiplet pick-up in traditional embedded domains. It's probably quite likely.

They have been "enabling" them but not designed the best of them(°), and I'm not sure how serious they are about the top end because their results are rather half-assed compared to Apple, AMD and Intel. As is, their bread and butter and main focus is still mobile and embedded chips.

(°) The best of them also seem to use barely any ARM standards except for the ISA itself

Arm's definitely trying to push on the laptop, tablet, desktop, and server markets. The fastest cluster on the top500 was arm for several years, most of the big clouds either have home grown arm servers (like graviton) or will soon.

They are definitely making progress.

Maybe ARM sees a life for ARM outside of mobile? Someplace where you might need a lot of cores and a chiplet approach could play out.

I've heard rumours of AMD working on something like this. They're going to call it Zen1. Could be a total nothing burger though, we'll see.

ARM haven't made all that much headway there. AMD designed an ARM ISA Zen 1, yes, but they didn't release it. I haven't heard anything about it since. I'd be not too surprised yet excited if AMD released an ARM Zen tomorrow, but there don't seem to be any recent rumors.

I was making a joke about how people also downplayed the benefits of chiplets when AMD released their Zen1 cpus. I am aware they also had an arm design cooking, but that wasn't the direction I was going. Apologies if this was a whoosh.

You mean like over half of hardware added to the amazon cloud in the last year? Graviton, now in it's 4th gen, seems to be doing quite well. I believe other large cloud providers are working on similar home grown arms.

Or maybe all the apple desktops? (Imac, Studio, and Mini)

Or maybe one of the largest HPC clusters on the top 500 list, fugaku? Was number 1 for several years.

AMD had and killed an arm project, but rumors claim they are working on sound wave APU that has an arm chip combined with a GPU.

Or similar Nvidia's GB10, which is their new AI dev kit the size of a mac mini with "1 petaflop" that combines 20 arm cores with a blackwell GPU.

Seems like arm is doing just fine outside of mobile.

See my other comment in this thread.

Arm doesn't only do ISA. It essentially wrote the standards for the AMBA/AXI/ACE/CHI interconnect space. Standardizing chip-to-chip interconnects is very much in Arm's interests. It is a double edged sword though since Chiplets will likely enable fine grained modularity allowing IP from other vendors to be stitched around Arm (eg RISC-V IOMMU instead of Arm SMMUv3 etc).

Just curious: is this hardware having its "micro-services" moment?

basically yes :)

chiplets seems the hot new thing, with an important difference - it's not just a fad: a lot of investments and production plans are moving in this direction. Expensive stuff. So IT HAS to work. Failure would be very expensive. But it's still totally possible.

imagine your services crashed at runtime 5% of the time due to cosmic rays.

actual number in hardware is a closely guarded secret, at least for the most recent process, but the high level solution is the same... build smaller boxes so the whole system can handle one of them failing gracefully.

That has been the way of how things were done in hardware from the start I'd say, just look at any moderately complex PCB - you can see each IC as a "microservice" of sorts, even analog function groups (oscillators, bandpasses, filters, ...) made out of discrete components can be considered "microservices" of sorts IMHO.

All this is is moving the definition scope of an individual "service" from the IC to the chiplet level.

That's how it feels to me as well. It's just a matter of cyclic evolution.

We can specialize hardware at the chiplet level and at some point, I expect more players and more standards out there that "synchronization" and "compatibilty" would become a bottleneck/burden.

We will go back to monoliths eventually again is my guess.

Anybody knows if "Silicon Box" is involved as well?

Reply on news.ycombinator.com