News Datacenter chipmaker Ampere, once valued at $8 billion, explores possible sale: Report

[Admin]

Oracle-backed Ampere is exploring potential sale, Bloomberg says.

Datacenter chipmaker Ampere, once valued at $8 billion, explores possible sale: Report : Read more

 

[Kamen Rider Blade]

This sounds like a perfect target for Qualcomm.

Qualcomm already has Nuvia for Server chip Designs, but buying out Ampere, it could kick start it's rush into the DataCenter by having a ready to go product that can be iterated on.

This would give it CPU products in Enterprise Server, Consumer LapTops, & Mobile SoC's.

Covering 3x very different sections.

 

[thestryker]

Ampere is in a really unfavorable position unfortunately. They're the only company making general purpose enterprise Arm CPUs and their business model is being attacked from all sides. As mentioned in the article AMD and Intel are finally getting competitive in core counts, but they're also providing more performance. The other side that's squeezing them is companies doing their own custom Arm architecture chips. I think this is a bigger problem than AMD/Intel as a lot of these custom solutions can effectively be drop in replacements.

Qualcomm would seemingly be the obvious choice for acquisition, but they've indicated they have no interest in enterprise general purpose. This may just be due to the spectacular failure of Centriq. Marvell and MediaTek would be the other two positioned in the right place, but I'm not sure what sort of greater asperations either has towards that sort of market.

Hopefully they're able to find a stable way forward though as the last thing the industry needs is more consolidation.

 

[Kamen Rider Blade]

Ampere is in a really unfavorable position unfortunately. They're the only company making general purpose enterprise Arm CPUs and their business model is being attacked from all sides. As mentioned in the article AMD and Intel are finally getting competitive in core counts, but they're also providing more performance. The other side that's squeezing them is companies doing their own custom Arm architecture chips. I think this is a bigger problem than AMD/Intel as a lot of these custom solutions can effectively be drop in replacements.

Qualcomm would seemingly be the obvious choice for acquisition, but they've indicated they have no interest in enterprise general purpose. This may just be due to the spectacular failure of Centriq. Marvell and MediaTek would be the other two positioned in the right place, but I'm not sure what sort of greater asperations either has towards that sort of market.

Hopefully they're able to find a stable way forward though as the last thing the industry needs is more consolidation.

Are they really the only one doing a generic Server/Enterprise CPU?

Nobody else in the ARM eco system is doing a standard CPU?

Why can't the big "Custom Arm Guys" join hands with Ampere to make a General Purpose Arm Enterprise/Server grade CPU that would perform for all their requirements?

Wouldn't the shared R&D along with combined buying power get them better Waffer Starts?

 

[thestryker]

Are they really the only one doing a generic Server/Enterprise CPU?

Nobody else in the ARM eco system is doing a standard CPU?

All of the companies who did got bought (with no products being released to the market), failed (like Qualcomm with Centriq), or went out of business.

There are plenty examples of standard Arm IP being used in enterprise but these are all customer specific like nvidia with Grace or amazon with Graviton.

Why can't the big "Custom Arm Guys" join hands with Ampere to make a General Purpose Arm Enterprise/Server grade CPU that would perform for all their requirements?

Wouldn't the shared R&D along with combined buying power get them better Waffer Starts?

Apple and Qualcomm are the only two running fully custom Arm based CPUs right now. Ampere's first is the AmpereOne which sounds very impressive for their future (these are the ones going up to 256 cores, up to 192 currently).

Ampere does have a deal with Qualcomm to utilize some of their AI technology for use with AI specific workloads (breaking out of just general computing). I think the biggest question is really what the larger players in the Arm SoC game are interested in doing and how much of a market they see. There are certainly benefits to be had with larger budgets and better access, but it's mostly a question of whether or not they see it as worth it.

 

[jackt]

Ampere should make riscv too.

 

[bit_user]

However, now that AMD has 192-core x86 CPUs and Intel has 244-core x86 processors

Neither of these have launched, yet. They're imminent (more so, in AMD's case; I believe the 244-core Sierra Forest isn't due until Q1 2025), but they shouldn't be put on par with the 192-core AmpreOne CPUs that are currently shipping.

Oracle could also be a potential buyer for Ampere. However, given the fact that in the past it abandoned the development of proprietary Sun UltraSPARC processors (after it took over Sun Microsystems in 2009 - 2010) and favored industry-standard Intel Xeon, it does not look like it is interested in developing CPUs in-house.

There's a big difference between merely designing CPUs like the Altra, which used basically all off-the-shelf IP from ARM, and designing custom ISA processors, like UltraSPARC. Sun went even a step further and developed its own OS (Solaris), although most of the focus had shifted to Linux, in the latter days of UltraSPARC. The benefit of ARM is that virtually all of the toolchain and OS support is handled by others, including ARM themselves.

If Ampere stepped back from designing custom cores, then Oracle would have in its hands a capability of customizing ARM IP similar to what Google, Microsoft, and Amazon have. That makes sense to me for Oracle to want, but it's worth a heck of a lot less than $8B.

I think Ampere made a big gamble with designing its own cores. They actually got this capability from AMCC, which designed Ampere's early eMAG processors:

https://www.anandtech.com/show/15733/ampere-emag-system-a-32core-arm64-workstation/3
​

No, they weren't very good. Nothing like Altra. Ampere should've taken Altra's success to heart and just focused on providing a pathway for small & medium-sized businesses to buy their own ARM server hardware (containing ARM IP). I know that's not a big value-add, but competing with ARM on core design is neither easy nor cheap.

Oracle might prefer to keep away from CPU development to maintain its relationships with AMD, Intel, and Nvidia

That's silly. Amazon, Microsoft, and Google (also Facebook?) assemble their own ARM CPUs from ARM IP, and yet they also source hardware from the x86 folks and Nvidia. A company like Oracle is more than big enough to do the same.

 

[bit_user]

Qualcomm would seemingly be the obvious choice for acquisition, but they've indicated they have no interest in enterprise general purpose. This may just be due to the spectacular failure of Centriq.

Centriq didn't fail, it was killed in the cradle! It was a casualty of the takeover bid from Broadcom. For Qualcomm to fend off a hostile takeover, it had to convince shareholders and the board that it could be more profitable and that the bid from Broadcom was too cheap. That meant significant cost-cutting, which included Centriq.

Marvell and MediaTek would be the other two positioned in the right place, but I'm not sure what sort of greater asperations either has towards that sort of market.

Marvel already bought Cavium. They decided those weren't terribly competitive on the open market, so they took their server business semi-custom:

https://www.anandtech.com/show/1604...ver-platforms-towards-custom-silicon-business
​

I think I heard a rumor they're doing contract CPU designs for Google, actually. Perhaps I'm getting the companies mixed up... Regardless, it seems their custom datacenter CPU/DPU business is still a thing:

https://www.anandtech.com/show/21294/marvells-2nm-ip-platform-enables-custom-silicon-for-datacenters
​

Apple and Qualcomm are the only two running fully custom Arm based CPUs right now.

Aside from Ampere, right? AmpereOne uses custom cores.

Also, Fujitsu is still around and working on future CPUs. I forget whether they've said which ISA they'll be using.

And didn't Huawei (HiSilicon?) claim to have their own ARM server CPUs with custom cores?

Pretty much everyone else doing custom cores has switched over to RISC-V, aside from the oddballs at Tachyum. Then, IBM is still designing mainframe CPUs and you've got China's Loongson with their own ISA. Any others?


BTW, Ampere recently made a presentation on AmpereOne at Hot Chips:

AmpereOne at Hot Chips 2024: Maximizing Density

Ampere Computing focuses on developing cloud-native processors, and aims to serve many users while providing security, privacy, and consistent performance, without relying on process node scaling.

chipsandcheese.com

 

[bit_user]

This sounds like a perfect target for Qualcomm.

Qualcomm already has Nuvia for Server chip Designs, but buying out Ampere, it could kick start it's rush into the DataCenter by having a ready to go product that can be iterated on.

This would give it CPU products in Enterprise Server, Consumer LapTops, & Mobile SoC's.

Covering 3x very different sections.

A fun fact about Nuvia is that they started out trying to build server CPUs. When Qualcomm acquired them, they retargeted their cores at the mobile/laptop segment. I'm sure the team from Nuvia is still wanting to do server CPUs and their cores would probably do well in that segment. When asked about it, Qualcomm hasn't said "no", but more of a non-committal: "not yet, laptops is where we're currently focused".

Perhaps the main thing Ampere could contribute is experience with all the system-level stuff and the customer relationships, in order to more quickly deploy Nuvia's IP into that market. However, the price they'd have to pay for Ampere would likely be much too high, just for that.

 

[thestryker]

Centriq didn't fail, it was killed in the cradle! It was a casualty of the takeover bid from Broadcom. For Qualcomm to fend off a hostile takeover, it had to convince shareholders and the board that it could be more profitable and that the bid from Broadcom was too cheap. That meant significant cost-cutting, which included Centriq.

That seems like a gross oversimplification of what was going on. They certainly did have the threat of Broadcom, but they were also trying to buy NXP (which failed after the failed hostile takeover attempt) and facing Epyc. All of this while attempting to break into a market where there wasn't guaranteed volume business. The Broadcom threat was publicized the same week they announced Centriq SKUs/pricing and that select customers were getting first shipments. This never materialized into volume products and servers using them weren't announced until after the server division was basically a zombie.

I have no doubt the Broadcom issues played a part, but the failure absolutely cannot be simply laid at the feet of it unless you have some sort of evidence to back that up.

 

[bit_user]

I have no doubt the Broadcom issues played a part, but the failure absolutely cannot be simply laid at the feet of it unless you have some sort of evidence to back that up.

I don't have evidence, but that's what the rumor mill was saying.

In contrast, your narrative that it simply flopped doesn't align with how they wound down the design team basically at the same time they were getting final silicon from the fabs. Early performance data looked promising and the product simply was never even given a chance to compete in the marketplace.

 

[thestryker]

I don't have evidence, but that's what the rumor mill was saying.

Good for the rumor mill, but there was so much going on with Qualcomm at the time hanging it on that alone simply isn't believable.

In contrast, your narrative that it simply flopped

I didn't say it flopped I said it failed to flop you have to have actually released a product which they didn't. By any definition it was a failure: they spent billions of dollars for no return.

doesn't align with how they wound down the design team basically at the same time they were getting final silicon from the fabs.

Your timeline here doesn't make any sense. They "launched" the product towards the end of 2017, except that they never actually shipped volume. Qualcomm went radio silent afterwards until finally dismantling what was left of the team around the end of 2018.

Early performance data looked promising and the product simply was never even given a chance to compete in the marketplace.

What performance data was this? The stuff they did on the run up to the launch? It looked like it could have potential then, but Epyc launched during the time between pre-release information and the launch that wasn't. If management thought they had enough of a place in the market there's no reason for them to not move ahead given the billions they'd spent getting to that point.

edit: I think cloudflare had some data around the launch as well.

 

[Kamen Rider Blade]

A fun fact about Nuvia is that they started out trying to build server CPUs. When Qualcomm acquired them, they retargeted their cores at the mobile/laptop segment. I'm sure the team from Nuvia is still wanting to do server CPUs and their cores would probably do well in that segment. When asked about it, Qualcomm hasn't said "no", but more of a non-committal: "not yet, laptops is where we're currently focused".

That's what I'm hoping for, combine the two Server Teams into one Team and make one Server CPU line.

Perhaps the main thing Ampere could contribute is experience with all the system-level stuff and the customer relationships, in order to more quickly deploy Nuvia's IP into that market. However, the price they'd have to pay for Ampere would likely be much too high, just for that.

That's what acquisition price negotiations are for.
Qualcomm's job is to pay less, Ampere wants as much $$$ as possible.

There's a compromise to be made somewhere.

 

[ezst036]

Ampere better make some of its CPUs for the AM4/AM5 platform.

I would buy one.

 

[Kamen Rider Blade]

Ampere better make some of its CPUs for the AM4/AM5 platform.

I would buy one.

Highly doubtful that it can just waltz onto another competitors platform.

It would have to create it's own platform from scratch.

 

[bit_user]

Ampere better make some of its CPUs for the AM4/AM5 platform.

I would buy one.

Zero chance of that happening, sadly.

Ampere's CPU are actually socketed, though. You can buy motherboards for them from Supermicro, Gigabyte, and maybe a couple others. There are a few system builders who sell full Ampere workstations (and many more which sell Ampere servers).

 

[thestryker]

Ampere's CPU are actually socketed, though. You can buy motherboards for them from Supermicro, Gigabyte, and maybe a couple others. There are a few system builders who sell full Ampere workstations (and many more which sell Ampere servers).

I don't know if it will continue with AmpereOne, but ASRock Rack sells a couple of Altra MATX boards with the CPU. They're pretty well featured and relatively speaking low cost.