News Arm aims to capture 50% of data center CPU market in 2025

ARM is definitely growing in the service space, we use ARM a far bit these days for our backend services. However, I struggle to see how ARM is going to increase their market share 3.3x in 9 months to 50% from 15%. I doubt enough applications will come online that can run on ARM in that short period of time while replacing x86/x64 versions of their services.
 
The article said:
Google and Microsoft have also started designing data center processors with Arm's technology, although their projects are at an earlier stage compared to Amazon.
It's true that they're at an earlier stage of maturity, but both have had cloud instances available to customers for a while. Microsoft's is called the Cobalt 100:


Google has its Axion-branded ARM server CPUs, which Phoronix even benchmarked against an x86 instance and an Ampere Altra:

Like Amazon and Nvidia, Google went with Neoverse V2 cores, while MS broke from the pack and used a greater number of the smaller Neoverse N2 cores.

The article said:
The company's Grace CPUs with 144 Arm Neoverse V2 cores
No, it's only 72 cores per CPU. Same as Google Axion. Amazon's Graviton 4 went up to 96 cores per CPU.

The way Nvidia reaches 144 cores is by putting two Grace CPUs on a module. Each has its own LPDDR5X DRAM and they communicate via NVLink. So, they're every bit as distinct as two x86 CPUs would be, when you have dual-CPU server mainboard. Remember: Nvidia always inflates their numbers. The only thing that differs is how.

BTW, I think AMD is going to have a major ARM server CPU announcement later this year, as well. I have no specific information, so call it a hunch.
 
Last edited:
While x86 dominates the server market and will likely continue for a while, Arm adoption is growing.
The revenue for amd and intel for datacenter in all of 2024 was 12,5 and 12,8 bil
Let's face it, servers are dominated by GPUs, or now NPUs I guess will be the hot thing.
Nvidia made 130 bil revenue last year, I don't know how much of that was gaming GPUs but, come on.
 
  • Like
Reactions: artk2219
Considering the short time horizon, and the time it takes to fab, package, and integrate CPUs in to product (DC stuff takes more engineering and integration time), 2025 is mostly set in stone already. So, this guy either knows something or is utterly clueless.
 
Call me crazy, but doesn’t that seem a bit optimistic?
Certainly, if we're talking about installed base. However, he could mean as a share of sales. With virtually all the big cloud providers transitioning to ARM, it does sound like it might be at the optimistic end of plausible that 50% of all server CPUs deployed in 2025 could be ARM-based.

Also, keep in mind that Intel won't have any new server CPUs until 2026 and I think the same is basically true of AMD. That makes x86 basically a known quantity and a fixed target. If these Neoverse V2 & N2-based CPUs are offering better TCO, then there's a motive to pair with the means (above).
 
ha, good joke arm, almost had me.
Why, you like to pay for more electricity because you enjoy creating unneeded heat ?

I am reminding you that the chip that has the best efficiency is an ARM chip
The chip that has the best single thread performance is an ARM chip.
The chip that has the best multi thread score is an ARM chip.

x86 is only here because of inertia.
 
Why, you like to pay for more electricity because you enjoy creating unneeded heat ?

I am reminding you that the chip that has the best efficiency is an ARM chip
To be fair, we don't have energy efficiency data on the latest round of these server CPUs, because (except for Nvidia's), they're all in the cloud. And Nvidia's isn't optimized for energy efficiency.

The chip that has the best single thread performance is an ARM chip.
When you're comparing against x86 running in SMT mode, yes. But, not when x86 is running one thread per core. So, which makes sense in perf/$ really depends on cloud provider pricing, which does usually favor their in-house ARM CPUs.

The chip that has the best multi thread score is an ARM chip.
Which multi-threaded score are you talking about, now?
 
Why, you like to pay for more electricity because you enjoy creating unneeded heat ?

I am reminding you that the chip that has the best efficiency is an ARM chip
The chip that has the best single thread performance is an ARM chip.
The chip that has the best multi thread score is an ARM chip.

x86 is only here because of inertia.
I said that because the article was posted on April 1st.
 
  • Like
Reactions: bit_user
Call me crazy, but doesn’t that seem a bit optimistic?

Certainly, if we're talking about installed base. However, he could mean as a share of sales. With virtually all the big cloud providers transitioning to ARM, it does sound like it might be at the optimistic end of plausible that 50% of all server CPUs deployed in 2025 could be ARM-based.

Also, keep in mind that Intel won't have any new server CPUs until 2026 and I think the same is basically true of AMD. That makes x86 basically a known quantity and a fixed target. If these Neoverse V2 & N2-based CPUs are offering better TCO, then there's a motive to pair with the means (above).
Arm server chips are a known quantity too. Both 2nd gen Neoverse versions have already been tested.
 
Arm server chips are a known quantity too. Both 2nd gen Neoverse versions have already been tested.
Sure. My point was that when all of the quantities are known, you can solve for the remaining unknown, which is the acceleration in their sales growth. In fact, ARM is probably basing this assertion on known order volumes that cloud operators might've placed, based on things like TCO advantages.

So, I was saying that neither AMD nor Intel appear to be launching anything which could disrupt those plans, making such a prediction actually quite easy.

BTW, I do wonder when Graviton 5 will launch and what it'll look like. Graviton 4 is at least 15 months old, so it seems quite likely we'll be hearing about it before the year is up.