These processors will have very decent performance for many applications, very similar to that of AWS Graviton5, but with more cores per socket.
However, the claim made by Arm: "the Arm AGI CPU, for agentic AI infrastructure, delivering more than 2x performance per rack compared with x86 platforms" is obviously false.
The new Intel Clearwater Forest Xeon processors use Darkmont cores, which have approximately the same performance per core, the same die area per core and the same power consumption per core as the Neoverse V3, but Intel offers 288 cores per socket and 576 cores per board, in comparison with only 136 cores per socket for Arm.
Therefore there is no chance that these new Arm processors can provide more performance per rack than Intel Clearwater Forest.
For applications that benefit from array operations, the AMD Zen 5 compact cores have much more performance per core than Neoverse V3 and AMD has provided 192 cores per socket for a long time. There is no chance for the new processors to exceed the performance per rack of Zen 5, but for those applications that do not benefit from array operations, these new Arm CPUs should have better performance per watt than Zen 5. But by the end of the year AMD should have Zen 6 Epyc CPUs, with more cores per socket, enhanced performance per core and improved performance per watt, so then there would be even less opportunities for these Arm CPUs to be better at something.
The only way how the claim of Arm can be true is if they have compared their new CPUs with antiquated CPUs like the Intel Granite Rapids Xeon CPUs, instead of comparing with state-of-the-art Intel Clearwater Forest and AMD Zen 5.
Aren't Intel Xeon Rapids and Intel Xeon Forest just different target markets? Rapids has fewer but faster cores in general, and more special-purpose accelerators (e.g. AMX, QAT), while Forest is focused on maximum compute density (just pack in as many fast-enough cores as you can).
IIRC Granite Rapids is also not _that_ old, and either current or a single generation behind. (Has its successor landed yet? IIRC GNR is the same generation as Sierra Forest).
Validating a core to server standards takes significantly longer.
V4 cores should be out this year using X925 and C1 Ultra-based V5 will probably be 2027-2028.
I suspect that X4 is already fast enough to beat EPYC in per-core performance when using the whole chip. ARM caught up/passed x86 in IPC all the way back around A77/78 in 2019-2020. They are now much faster per clock and hitting about the same all-core clockspeeds as standard EPYC (let alone zen5c EPYC).
The big issue is that Graviton5 is already starting to hit the market and uses the same v3 cores. A lot of marketshare for this chip will probably come from taking Ampere customers.
Cortex-X4 a.k.a. Neoverse V3 has significantly lower performance per core than Zen 5.
However, Neoverse V3 has a lower die area, so you could implement more cores per socket than with Zen 5, but this has not been done yet, as these new CPUs have only 136 cores per socket versus 192 cores per socket for Zen 5.
For programs that do not use array operations, i.e. which do not use AVX/AVX-512 instructions, Neoverse V3 has better performance per watt than Zen 5. But that changes for programs that benefit from AVX/AVX-512, where Zen 5 has better performance per watt.
Moreover, Zen 5 is already old. By the end of the year there will be Zen 6, which will be the real competitor for these new Arm CPUs, and Zen 6 will have better performance per watt, even more cores per socket and even more performance per core.
I think the only ARM licensee going for the hyperscaler CPU market is Ampere. Amazon and Microsoft make CPUs for themselves and Nvidia’s are aimed exclusively at AI workloads driving their GPUs.
They use ARM cores designed by the Arm company, but the complete chips are designed by AWS/Microsoft.
Ampere had previously used cores designed by Arm, but their latest CPUs (which do not impress much) use a custom core, like the Apple and Qualcomm CPUs.
I know, but I can't buy a Cobalt or Graviton workstation. Ampere has been the only way I could lay my hands on a nice workstation-grade ARM chip (unless you count Apple, but they also don't sell chips)
Not sure how to feel about this. Does this mean ARM is slowly moving from just licensing IP to actually competing with companies building on top of it?
This seems bad, doesn’t it? I already know that there has been friction between arm and their customers over higher licensing fees since the IPO just trying to put this in context.
The ones who recognize standards as a good thing. ARM making their own CPUs shifts their focus from making a good ISA for people to use to making a good ISA to use in their own CPUs.
Softbank still owns 90% of ARM and they finished their acquisition of Ampere only a few months ago in November 2025.
I'm a chip designer and a chip this complicated takes about 3 years from start to actual silicon so it would have started well before Softbank started their acquisition process of Ampere.
The press release says it was co-developed with Meta who has a growing custom chip team. Normally these chips like Amazon's Graviton or Google's Axion are designed for their own data center use only and rented to customers. This ARM chip sounds like Meta and other companies will all be able to buy chips for their own data centers.
I'm guessing Softbank will get ARM and Ampere to align on future chips or just merge Ampere completely into ARM.
Agreed. The ARM AGI CPU supports a newer version of the vectorized instructions and has matrix math extensions that the AmpereOne M doesn’t. Also has almost twice the memory bandwidth. One paper at least, the AGI CPU seems like a better choice for AI workloads. Ampere is really pushing the AI workload use cases for the AmpereOne M, so this really makes their lives a lot harder.
Neoverse V3 is better than any core used or designed by Ampere, but it does not have matrix math extensions.
Neoverse V3 is the server version of Cortex-X4 and it is an Armv9.2-A CPU, with SVE2, but without SME/SME2.
The matrix math extensions, i.e. SME/SME2, are present only in the latest generation of Arm cores (the C1 cores), which implement the Armv9.3-A ISA, and also in recent Apple cores.
Neoverse V3 is also used in AWS Graviton5 and in a few NVIDIA products, e.g. in Thor, and it is also pretty much equivalent with the Skymont and Darkmont Intel E-cores, which are used in the Lunar Lake, Arrow Lake, Panther Lake and Clearwater Forest CPUs.
These processors will have very decent performance for many applications, very similar to that of AWS Graviton5, but with more cores per socket.
However, the claim made by Arm: "the Arm AGI CPU, for agentic AI infrastructure, delivering more than 2x performance per rack compared with x86 platforms" is obviously false.
The new Intel Clearwater Forest Xeon processors use Darkmont cores, which have approximately the same performance per core, the same die area per core and the same power consumption per core as the Neoverse V3, but Intel offers 288 cores per socket and 576 cores per board, in comparison with only 136 cores per socket for Arm.
Therefore there is no chance that these new Arm processors can provide more performance per rack than Intel Clearwater Forest.
For applications that benefit from array operations, the AMD Zen 5 compact cores have much more performance per core than Neoverse V3 and AMD has provided 192 cores per socket for a long time. There is no chance for the new processors to exceed the performance per rack of Zen 5, but for those applications that do not benefit from array operations, these new Arm CPUs should have better performance per watt than Zen 5. But by the end of the year AMD should have Zen 6 Epyc CPUs, with more cores per socket, enhanced performance per core and improved performance per watt, so then there would be even less opportunities for these Arm CPUs to be better at something.
The only way how the claim of Arm can be true is if they have compared their new CPUs with antiquated CPUs like the Intel Granite Rapids Xeon CPUs, instead of comparing with state-of-the-art Intel Clearwater Forest and AMD Zen 5.
Aren't Intel Xeon Rapids and Intel Xeon Forest just different target markets? Rapids has fewer but faster cores in general, and more special-purpose accelerators (e.g. AMX, QAT), while Forest is focused on maximum compute density (just pack in as many fast-enough cores as you can).
IIRC Granite Rapids is also not _that_ old, and either current or a single generation behind. (Has its successor landed yet? IIRC GNR is the same generation as Sierra Forest).
It really is a choice for Arm to use their 2023 based mobile X4 cores instead of their current C1 Ultras for this. Hopefully they step up quickly
Validating a core to server standards takes significantly longer.
V4 cores should be out this year using X925 and C1 Ultra-based V5 will probably be 2027-2028.
I suspect that X4 is already fast enough to beat EPYC in per-core performance when using the whole chip. ARM caught up/passed x86 in IPC all the way back around A77/78 in 2019-2020. They are now much faster per clock and hitting about the same all-core clockspeeds as standard EPYC (let alone zen5c EPYC).
The big issue is that Graviton5 is already starting to hit the market and uses the same v3 cores. A lot of marketshare for this chip will probably come from taking Ampere customers.
Cortex-X4 a.k.a. Neoverse V3 has significantly lower performance per core than Zen 5.
However, Neoverse V3 has a lower die area, so you could implement more cores per socket than with Zen 5, but this has not been done yet, as these new CPUs have only 136 cores per socket versus 192 cores per socket for Zen 5.
For programs that do not use array operations, i.e. which do not use AVX/AVX-512 instructions, Neoverse V3 has better performance per watt than Zen 5. But that changes for programs that benefit from AVX/AVX-512, where Zen 5 has better performance per watt.
Moreover, Zen 5 is already old. By the end of the year there will be Zen 6, which will be the real competitor for these new Arm CPUs, and Zen 6 will have better performance per watt, even more cores per socket and even more performance per core.
Yeah seems like competing with your customers is a bad idea.
I think the only ARM licensee going for the hyperscaler CPU market is Ampere. Amazon and Microsoft make CPUs for themselves and Nvidia’s are aimed exclusively at AI workloads driving their GPUs.
AWS Graviton and Microsoft Cobalt are arm-based.
They use ARM cores designed by the Arm company, but the complete chips are designed by AWS/Microsoft.
Ampere had previously used cores designed by Arm, but their latest CPUs (which do not impress much) use a custom core, like the Apple and Qualcomm CPUs.
I know, but I can't buy a Cobalt or Graviton workstation. Ampere has been the only way I could lay my hands on a nice workstation-grade ARM chip (unless you count Apple, but they also don't sell chips)
Arm and ampere are both majority owned by the soft bank now too…
[dupe] https://news.ycombinator.com/item?id=47506251 (18 minutes older, 6 comments)
This is my condensed version for the SBCwiki documentation focused on the key facts without all the unnecessary marketing around it
This is substantially more useful than the marketing fluff in the press release. Probably would have made sense to post this in that thread though
I added a cost/performance analysis for that at https://news.ycombinator.com/item?id=47509236 in case anyone's interested.
ARM naming a chip AGI is either the most confident product launch in history or the best marketing we have seen in years. Probably both.
Or for certain people, it makes them cringe a little whenever they see it..
Waiting for ARM-AGI-2
They actually said it’s coming in 2027 and ARM-AGI-3 is marked as future https://xcancel.com/mecoscorner/status/2036510613350785418
I wonder if it's a joke like Arm-Generative-Intelligence or something like that.
They say that it stands for AGentic AI.
They certainly are optimistic
It’s so deliberately misleading.
I can only imagine that their boardroom minutes included heated exchanges between their legal and marketing teams.
Not sure how to feel about this. Does this mean ARM is slowly moving from just licensing IP to actually competing with companies building on top of it?
I figured they would do this when SoftBank acquired them.
SB likes 10x buys.
Their marketing department is smoking a lot of hopium. I will now think of it as the ARM MatMul Unit.
No, the Arm TPU is coming later.
This seems bad, doesn’t it? I already know that there has been friction between arm and their customers over higher licensing fees since the IPO just trying to put this in context.
This is great for consumers. More options leads to more choices means more competition means lower prices. Which group is it bad for?
The ones who recognize standards as a good thing. ARM making their own CPUs shifts their focus from making a good ISA for people to use to making a good ISA to use in their own CPUs.
I wonder what the people at Ampere are thinking right now
Softbank still owns 90% of ARM and they finished their acquisition of Ampere only a few months ago in November 2025.
I'm a chip designer and a chip this complicated takes about 3 years from start to actual silicon so it would have started well before Softbank started their acquisition process of Ampere.
The press release says it was co-developed with Meta who has a growing custom chip team. Normally these chips like Amazon's Graviton or Google's Axion are designed for their own data center use only and rented to customers. This ARM chip sounds like Meta and other companies will all be able to buy chips for their own data centers.
I'm guessing Softbank will get ARM and Ampere to align on future chips or just merge Ampere completely into ARM.
Damn, no wonder Ampere stopped releasing chips. They were churning them out and then just crickets.
Agreed. The ARM AGI CPU supports a newer version of the vectorized instructions and has matrix math extensions that the AmpereOne M doesn’t. Also has almost twice the memory bandwidth. One paper at least, the AGI CPU seems like a better choice for AI workloads. Ampere is really pushing the AI workload use cases for the AmpereOne M, so this really makes their lives a lot harder.
Neoverse V3 is better than any core used or designed by Ampere, but it does not have matrix math extensions.
Neoverse V3 is the server version of Cortex-X4 and it is an Armv9.2-A CPU, with SVE2, but without SME/SME2.
The matrix math extensions, i.e. SME/SME2, are present only in the latest generation of Arm cores (the C1 cores), which implement the Armv9.3-A ISA, and also in recent Apple cores.
Neoverse V3 is also used in AWS Graviton5 and in a few NVIDIA products, e.g. in Thor, and it is also pretty much equivalent with the Skymont and Darkmont Intel E-cores, which are used in the Lunar Lake, Arrow Lake, Panther Lake and Clearwater Forest CPUs.
Wasn’t Ampere just bought by Softbank?
Yes, but they function as sister companies right now rather than one company.
[flagged]