News Arm to launch first self-made processors, poaching employees from clients: Reports

[Admin]

Arm lands first order from Meta for its first in-house designed datacenter processor, which will rival CPUs from Arm's partners.

Arm to launch first self-made processors, poaching employees from clients: Reports : Read more

 

[bit_user]

This decision places Arm in direct competition with some of its customers that develop Arm-based processors for data centers, including Ampere and, to some degree, Nvidia

Ampere already switched to using its own cores, last year. And Nvidia's Grace is quite specialized for integrating into a NVLink fabric with Hopper and Blackwell nodes.

Ampere might be a bit threatened by this move, just because they had recently been the only independent ARM server CPU maker, but it's probably not even on Nvidia's radar screen.

the Neoverse N3 CSS allows the building of CPUs with up to 8 N3 cores per die.

Not according to this:

"the N3 CSS design supports a range of CPU cores, from 8 up to 32"

Source: https://www.anandtech.com/show/21270/arm-announces-neoverse-v3-and-n3-cpu-cores
​

the final core count of Arm's offering remains to be seen.

The above article claims up to 128 V3 cores can be included per package, if spread across at least two dies.

Arm also has Neoverse E3 cores, which seem aimed more towards embedded servers.

This hiring activity contradicts CEO Rene Haas's testimony in December in the company's trial against Qualcomm, where he denied that Arm was building its own CPUs

Their previous announcement of silicon design sounded like they were focused just on CPU chiplets. So, that statement could've been true in the most literal sense that their customers would be others making their own CPUs (probably already IP customers of ARM's), and not systems companies that would be interested in buying only full CPUs.

 

[pug_s]

I think this is a slippery slope that ARM is doing. by competing against its customers. I can't wait until RISC V matures to a point where we see them in phones and laptops.

 

[bit_user]

I think this is a slippery slope that ARM is doing. by competing against its customers. I can't wait until RISC V matures to a point where we see them in phones and laptops.

If ARM establishes itself as a CPU maker, then they are no longer tied to the AArch64 ISA. That means they can pivot to making RISC-V CPUs like everyone else, when the time is right.

I think that's one reason they're making this move, but it's also to offset the huge hit they took on revenue from Qualcomm switching to its own core designs, instead of licensing them from ARM.

Besides, as long as this is limited to server CPUs, they're not really competing against anyone licensing their cores. It's mainly just architecture licensees, from whom Arm is making very little money anyway.

 

[Mindstab Thrull]

Just want to say that this isn't the first tech company to compete against their customers. NVIDIA already does this, as do some PSU manufacturers. Also doesn't Foxconn do this as well, or at least did at one point? So if they find a good reason to do so, and are using their own designs and whatnot, why not?

Mindstab Thrull

 

[bit_user]

Just want to say that this isn't the first tech company to compete against their customers. NVIDIA already does this, as do some PSU manufacturers. Also doesn't Foxconn do this as well, or at least did at one point? So if they find a good reason to do so, and are using their own designs and whatnot, why not?

Perhaps the grand daddy of all examples: Microsoft! Back in their heyday, they would look at which applications were popular and successful, then write their own version. That's where most of the software in MS Office came from.

 

[stuff and nonesense]

Arm choosing to make their own cores is a huge change to their established business model. A gamekeeper turning poacher.

If it is accurate that they are poaching their client’s customers then it is somewhat immoral.

If their clients have full access to the features, architecture (everything ARM is selling to the users) and has had time to generate a competitive solution - a level playing field then all is fair.
If ARM is abusing a position of privilege that’s their business. Their clients will decide whether they want to keep supporting them.

 

[bit_user]

Arm choosing to make their own cores is a huge change to their established business model.

They already sell their own core designs. So, it just seemed like the next logical step to get them manufactured into chiplets and save customers the trouble of managing layout and manufacturing of that part of their processors, instead just concentrating on their unique value-add. That was rumored sometime last year. Going from chiplets to entire CPUs is the next logical step. The only significant competitive threat I see is to their architectural licensees, but this part isn't really new.

A contrasting view would be that for Microsoft, Amazon, Google, and others all to take ARM IP and individually go through the motions of integrating it and getting it manufactured is a duplication of labor. If ARM subsumes that function as well, then it should drive efficiencies.

 

[stuff and nonesense]

They already sell their own core designs

License to clients, e.g. Qualcomm.

trouble of managing layout and manufacturing of that part of their processors

Licence available to clients

Going from chiplets to entire CPUs is the next logical step.

ok, but selling to ultimate customer potentially breaks the business models of their clients.

A contrasting view would be that for Microsoft, Amazon, Google, and others all to take ARM IP and individually go through the motions of integrating it and getting it manufactured is a duplication of labor. If ARM subsumes that function as well, then it should drive efficiencies.

ARM, if they subsumed the whole design and manufacturing procedure would revolutionise the ARM marketplace.
The duplication of labour is between competitors. Differing requirements based around the common core foster advances in design and improve the product. Competition advances the art. Though not a perfect example look at Apple v Qualcomm et al.

 

[bit_user]

License to clients, e.g. Qualcomm.

Qualcomm? Not any more!
; )

The loss of IP licensing revenue from Qualcomm is one of the reasons Arm is having to do this.

ok, but selling to ultimate customer potentially breaks the business models of their clients.

The thing is, the article seems to focus on server CPUs. The only ones who currently license ARM core designs and sell server CPUs to a 3rd party (i.e. as opposed to just using them internally, like Microsoft, Amazon, and Google all do), is Nvidia and Huawei. Nvidia's Grace is heavily customized for use within their GPU servers, and there's no way ARM is selling drop-in replacements because NVLink is proprietary. And Huawei is selling to Chinese customers who probably need or want to source from a Chinese brand, no matter what else might be on the market. Not to mention that I think Huawei is using their own cores, like Ampere has been doing since last year. So, I really don't see a single instance where someone who's currently licensing ARM cores would be directly threatened by this move, providing that it's contained only to the server market.

If ARM wanted to branch into selling their own mobile or laptop SoCs, that would be a different story, as it would put them in competition with MediaTek, Rockchip, Samsung, etc. However, ARM doens't have all the necessary IP needed for that. One glaring omission from their IP portfolio is a 5G modem, which (as Apple seems to have demonstrated by counterexample) is quite a non-trivial undertaking.

The duplication of labour is between competitors.

Yeah, but with scale comes efficiency. Arm can do it cheaper and possibly by enough that their customers would end up no worse off.

Differing requirements based around the common core foster advances in design and improve the product.

Arm has always allowed customers to order up their IP with customized parameters. Intel and AMD make custom Xeon and Epyc models that are made to order, for their biggest customers. There's no reason to think Arm won't continue this practice.

Competition advances the art.

Sure, but I don't really see anything in MS, Amazon, and Google's server CPUs that's such a special sauce. Not like what Nvidia is doing, at least. And if they want to incorporate some custom IP in the CPUs ARM would make for them, I'm sure Arm will oblige. Minimum order 10M.

Though not a perfect example look at Apple v Qualcomm et al.

It's not an example, at all. Both design their own cores. Amazon, Microsoft, and Google don't. They just pick parts out of Arm's existing IP catalog.

 

[pug_s]

Perhaps the grand daddy of all examples: Microsoft! Back in their heyday, they would look at which applications were popular and successful, then write their own version. That's where most of the software in MS Office came from.

I think a better example is what Microsoft is doing with making surface laptops/tablets and Intel is making with Intel Nucs. Both companies want to make appealing PC/laptop form factors so that other manufacturers would copy them and it works. Very much different thant what is ARM is doing.

Arm choosing to make their own cores is a huge change to their established business model. A gamekeeper turning poacher.

If it is accurate that they are poaching their client’s customers then it is somewhat immoral.

If their clients have full access to the features, architecture (everything ARM is selling to the users) and has had time to generate a competitive solution - a level playing field then all is fair.
If ARM is abusing a position of privilege that’s their business. Their clients will decide whether they want to keep supporting them.

Arm is suing Qualcomm over licensing money over Nuvia and now wants to compensate by making its own CPU's is really stupid and nothing more than a cash grab. ARM should instead focus more attention on how to make ARM Server CPU's mainstream by providing more support in reference designs so more companies would make ARM cpu's and get more license fees.

 

[bit_user]

I think a better example is what Microsoft is doing with making surface laptops/tablets and Intel is making with Intel Nucs. Both companies want to make appealing PC/laptop form factors so that other manufacturers would copy them and it works.

FWIW, Intel sold off their NUC division to ASUS, last year. In an interesting and perhaps predictable twist, ASUS recently announced a NUC based on an AMD SoC.

Arm is suing Qualcomm over licensing money over Nuvia

That's over and done with. The major part of the lawsuit went Qualcomm's way. Arm could appeal or try to re-litigate the remaining point, but so far nothing has been announced.

ARM should instead focus more attention on how to make ARM Server CPU's mainstream by providing more support in reference designs

Well, the rumor of Softbank buying Ampere for their expertise in this area could go straight to that. However, it's probably more intended to sell their own CPUs than help their partners sell server CPUs.

so more companies would make ARM cpu's and get more license fees.

Qualcomm looks like they're moving towards introducing server CPUs. I think AMD will do it, as well. The market has adequate scale and the software ecosystem is ready. Now is the time to strike.

For AMD, they're also seeing their penetration in the x86 server market start to decelerate and that entire market is probably decelerating already. So, they need to do something to stimulate more growth in the datacenter. AI has yet to net them a huge windfall.

RISC-V won't be there for a few more years, at least. So, jumping to Arm is pretty much the only move to be made right now.

 

[jp7189]

Perhaps the grand daddy of all examples: Microsoft! Back in their heyday, they would look at which applications were popular and successful, then write their own version. That's where most of the software in MS Office came from.

Too bad that didn't repeat that in the mobile space. Windows phone was a great platform without and apps to go with it.

 

[suryasans]

Remember Calxeda? That's the reason why ARM decided to enter physical Server cores product. Ampere based Server has not reached server soc marketshare percentage it hoped to break, 33% and making ARM processor as a real challenger against Intel and AMD in data center market.

 

[bit_user]

Ampere based Server has not reached server soc marketshare percentage it hoped to break, 33% and making ARM processor as a real challenger against Intel and AMD in data center market.

Ampere's CPU design team formerly worked under the aegis of AppliedMicro and created their X-Gene CPUs. It was during that time when Arm announced their Neoverse line of server cores (which were directly derived from their Cortex series of mobile cores, as it turned out).

https://www.anandtech.com/show/1347...rse-infrastructure-ip-branding-future-roadmap
​

Ampere's most successful product was Altra, but this was little more than cores and interconnect IP licensed from Arm. Actually, they had a clever way of doubling the occupancy of the interconnect, which enabled them to reach higher core counts than others. However, the same assets which propelled Ampere to the height of their success also enabled Arm to more directly serve hyperscalers, which deprived Ampere of the market it hoped to dominate.

 

[stuff and nonesense]

Actually, they had a clever way of doubling the occupancy of the interconnect

Please clarify, was this a novel implementation by Ampere?

 

[bit_user]

Please clarify, was this a novel implementation by Ampere?

Altra used Arm's CMN-600 interconnect IP, but that's natively limited to only 64 cores. And yet Ampere managed to fit up to 128 cores per CPU! Here's where I think I read about their workaround:

https://www.anandtech.com/show/1697...x-review-pushing-it-to-128-cores-per-socket/3
​