Huawei - 64-Core Arm Server CPU with CCIX and PCIe Gen4 Launched

[ub4ty]

https://www.servethehome.com/huawei-kunpeng-920-64-core-arm-server-cpu/

Today Huawei announced a new 64-Core Arm Server CPU. The Huawei Kunpeng 920 is being billed as the fastest Arm CPU to date. One thing is for sure, there is going to be interest in this CPU not just for the 64 custom Arm cores, but also the I/O that the chip has onboard. We thought we would take a moment to show what Huawei is building, and why it is important.

Beyond the CPU cores and 8 channel DDR4-2933 memory controller, the company is also introducing PCIe Gen4 and CCIX support. We know that AMD EPYC Rome will support PCIe Gen4 this year (likely with CCIX), and Intel Cascade Lake will not. Like we saw Cavium do with the original ThunderX, Huawei is also integrating networking with a 100GbE RoCE NIC in the Kunpeng 920.

And now comes the new competition.... Good luck to anyone who isn't innovating and pushing boundaries. The standards are being set.

 

[SarahKerrigan]

Just looking at this, and AMDs configuration for ROME, how difficult, or even useful in this market, would it be for AMD to spend some money developing 7nm chiplets based on ARM cores for an ARM version of ROME? It seems that by divorcing the I/O chip from the CPU cores, they gain not only flexibility in the x86 market, but ALSO the ability to use other CPU core technologies to create products for other markets as needed. I'm not even proposing a super complicated hybrid model with x86 and ARM cores on the same package, just a package that has choices for the manufacturer.

Always nice to see someone starting to appreciate Skybridge (or, I guess, Boxboro, if you're into older and more obscure systems...)

 

[darkswordsman17]

If Apple has any common sense, they'll push their own IP into the same space somehow. At the present Apple isn't even in the server market right now.

I agree that I think Apple should (and will) expand their CPU core use beyond mobile. I don't think they'll license it out though. They have plenty of need/demand in their own datacenters (and that's going to grow a lot/quickly), so I think they lock it up tight, both to keep others from getting details on their designs, but also for security (by that I mean software side). If anything, I'd expect them to buy a company like Cavium to apply their knowhow with Apple's cores. Or, like allegedly has been occurring with GPU development, maybe they'd partner with AMD.

Apple is targeting services more, and I think they'd move to licensing the server use versus the server hardware (so more like Amazon or Microsoft where they handle the hardware and software and you tell them what you need and they can tailor a hardware and software solution for you). They'd probably offer incentives for apps to use their backend as well (more ability for in app purchases and the like, maybe discount rates or better cut of app sales).

Would there be any benefits to matching products with cloud, by that I mean, for every SoC they have in a product, what would having an equal SoC in the cloud enable them? Could they do like some type of caching or offloading? So that on the cloud side they process as much as possible and then just send the minimum amount of data to the person's device. Or for some security sandboxing or something.

 

[DrMrLordX]

If anything, I'd expect them to buy a company like Cavium to apply their knowhow with Apple's cores. Or, like allegedly has been occurring with GPU development, maybe they'd partner with AMD.

Dunno about AMD partnerships, but I do suspect that they'll look to Cavium sooner or later. Cavium has the interconnect chops to put together a CPU based on parts of the A-series SoCs. A 64-core chip full of Vortex cores would be a beast.

Apple is targeting services more, and I think they'd move to licensing the server use versus the server hardware (so more like Amazon or Microsoft where they handle the hardware and software and you tell them what you need and they can tailor a hardware and software solution for you). They'd probably offer incentives for apps to use their backend as well (more ability for in app purchases and the like, maybe discount rates or better cut of app sales).

Would there be any benefits to matching products with cloud, by that I mean, for every SoC they have in a product, what would having an equal SoC in the cloud enable them? Could they do like some type of caching or offloading? So that on the cloud side they process as much as possible and then just send the minimum amount of data to the person's device. Or for some security sandboxing or something.

Interesting possibility. Apple does rely on an enormous back-end for their services. If they want to use A-series-based machines for service hosting, then it might be useful for them to have essentially one uarch or uarch family for SaaS/cloud applications.

 

[SarahKerrigan]

I agree that I think Apple should (and will) expand their CPU core use beyond mobile. I don't think they'll license it out though. They have plenty of need/demand in their own datacenters (and that's going to grow a lot/quickly), so I think they lock it up tight, both to keep others from getting details on their designs, but also for security (by that I mean software side). If anything, I'd expect them to buy a company like Cavium to apply their knowhow with Apple's cores. Or, like allegedly has been occurring with GPU development, maybe they'd partner with AMD.

Apple is targeting services more, and I think they'd move to licensing the server use versus the server hardware (so more like Amazon or Microsoft where they handle the hardware and software and you tell them what you need and they can tailor a hardware and software solution for you). They'd probably offer incentives for apps to use their backend as well (more ability for in app purchases and the like, maybe discount rates or better cut of app sales).

Would there be any benefits to matching products with cloud, by that I mean, for every SoC they have in a product, what would having an equal SoC in the cloud enable them? Could they do like some type of caching or offloading? So that on the cloud side they process as much as possible and then just send the minimum amount of data to the person's device. Or for some security sandboxing or something.

Marvell is a large semiconductor company that does a lot of stuff beyond just ThunderX. I have a hard time imagining Apple being interested in inheriting all of those product lines (and their long-term parts availability obligations.)

 

[DrMrLordX]

Marvell is a large semiconductor company that does a lot of stuff beyond just ThunderX. I have a hard time imagining Apple being interested in inheriting all of those product lines (and their long-term parts availability obligations.)

That is why I think Apple is more likely to license their core designs out to someone already competing in the server room.

 

[SarahKerrigan]

That is why I think Apple is more likely to license their core designs out to someone already competing in the server room.

I would be surprised if that's a path Apple is interested in taking, although their cores are definitely amazing. At that point it would probably make sense to add SMT, too.

 

[Arachnotronic]

That is why I think Apple is more likely to license their core designs out to someone already competing in the server room.

Why would they do that? For a small licensing fees and pennies per unit? Come on.

 

[DrMrLordX]

Why would they do that? For a small licensing fees and pennies per unit? Come on.

Who said anything about small licensing fees and pennies per unit? Besides, Cavium has already tossed their own design once for a design they acquired from an outside firm (Cavium picked up Vulcan from Broadcom and based ThunderX2 on that). It should be obvious by now that, without outside help, Cavium can't really keep up with something like the monster that Huawei just launched.

I agree that I don't think Apple wants to buy out Marvell or try to snatch away Cavium (considering Marvell just dropped $6 billion on Cavium recently). It is also unlikely that Apple wants to manage their own server CPU division. I do think Marvell wants their ARM server tech to be as competitive with Huawei as possible. And the strongest ARM core out there right now is Apple's Vortex. Until something major changes at Apple, it is probable that A-series chips will continue to dominate in the ARM world. If Marvell/Cavium want to win, they know where to go once the time comes to replace Vulcan. Apple will charge accordingly.

 

[Ajay]

If Apple has any common sense, they'll push their own IP into the same space somehow. At the present Apple isn't even in the server market right now.

They are a consumer product and services company it would make no sense whatsoever for them to move into the server market. Their custom arm development is merely an enabler for Apple's current product line where they see the benefit of a more tightly integrated platform.

 

[beginner99]

They are a consumer product and services company it would make no sense whatsoever for them to move into the server market. Their custom arm development is merely an enabler for Apple's current product line where they see the benefit of a more tightly integrated platform.

Fully agree. Apple has no interest in server market and it's a completely different market. They would probably fail just like intel failed in mobile.

 

[DrMrLordX]

They are a consumer product and services company it would make no sense whatsoever for them to move into the server market. Their custom arm development is merely an enabler for Apple's current product line where they see the benefit of a more tightly integrated platform.

You've just described fully half of what Huawei does, and here they are with a 64-core ARM server chip.

 

[SarahKerrigan]

Who said anything about small licensing fees and pennies per unit? Besides, Cavium has already tossed their own design once for a design they acquired from an outside firm (Cavium picked up Vulcan from Broadcom and based ThunderX2 on that). It should be obvious by now that, without outside help, Cavium can't really keep up with something like the monster that Huawei just launched.

I agree that I don't think Apple wants to buy out Marvell or try to snatch away Cavium (considering Marvell just dropped $6 billion on Cavium recently). It is also unlikely that Apple wants to manage their own server CPU division. I do think Marvell wants their ARM server tech to be as competitive with Huawei as possible. And the strongest ARM core out there right now is Apple's Vortex. Until something major changes at Apple, it is probable that A-series chips will continue to dominate in the ARM world. If Marvell/Cavium want to win, they know where to go once the time comes to replace Vulcan. Apple will charge accordingly.

Huh? Huawei claims to have 25% higher throughput than TX2, while also having double the core count and a far newer node. The idea that Marvell can't keep up with that is ridiculous. TX3 is going to be sampling this year, and I fully expect that it will match or beat this Huawei chip's performance.

Additionally, it strikes me as highly unlikely (for political and national-security reasons) that Huawei is going to make any inroads in, say, DoE labs - like the one currently operating a large TX2 supercomputer.

 

[DrMrLordX]

Huh? Huawei claims to have 25% higher throughput than TX2, while also having double the core count and a far newer node. The idea that Marvell can't keep up with that is ridiculous. TX3 is going to be sampling this year, and I fully expect that it will match or beat this Huawei chip's performance.

We'll see then. Maybe Marvell's deep(er) pockets can keep Cavium ahead. To me it looked like they were struggling when they had to scrap TX2 and start over using Vulcan.

Additionally, it strikes me as highly unlikely (for political and national-security reasons) that Huawei is going to make any inroads in, say, DoE labs - like the one currently operating a large TX2 supercomputer.

Anyplace that has restrictions on Huawei telecom equipment will have similar restrictions on their server CPUs.

 

[Arachnotronic]

You've just described fully half of what Huawei does, and here they are with a 64-core ARM server chip.

Huawei has long been a network infrastructure company -- long before it was anybody in the smartphone market.

 

[SarahKerrigan]

We'll see then. Maybe Marvell's deep(er) pockets can keep Cavium ahead. To me it looked like they were struggling when they had to scrap TX2 and start over using Vulcan.

It's a sign of struggling that a company expanding into a new market (servers) and having had previous issues with its own NPU-derived microarchitecture would pay for an experienced team with a mostly-complete core? And this means they're going to be behind forever?

 

[DrMrLordX]

It's a sign of struggling that a company expanding into a new market (servers) and having had previous issues with its own NPU-derived microarchitecture would pay for an experienced team with a mostly-complete core? And this means they're going to be behind forever?

To me it says, "we can build the interconnect, but we can't build a custom ARM core". If they picked up a team of designers from Broadcom then maybe their ThunderX efforts will have legs. Maybe. If not, they'll have to license/acquire yet another custom ARM core once Vulcan is played out.

Huawei has long been a network infrastructure company -- long before it was anybody in the smartphone market.

Correct. But they didn't have any products in the server market, at all. At least Apple has experience in that area, albeit in the past. Huawei? Nope. To the best of my knowledge, all of Huawei's CPU/APU/SoC designs come from HiSilicon, and HiSilicon launched their first CPU in 2012.

 

[SarahKerrigan]

To me it says, "we can build the interconnect, but we can't build a custom ARM core". If they picked up a team of designers from Broadcom then maybe their ThunderX efforts will have legs. Maybe. If not, they'll have to license/acquire yet another custom ARM core once Vulcan is played out.

Cavium got the former Netlogic team from Broadcom, as far as I can tell from raiding resumes - including David Hass, the architect on XLP and later Vulcan. Take a look at LinkedIn sometime. Check out the number of design engineers with both Marvell/Cavium and Netlogic listed.

When ATI bought BitBoys, did it indicate ATI was incapable of good GPU design and was screwed in the long term? Also, weren't you saying a few months ago that the entry of Qualcomm into the server market meant Cavium was screwed? How did that work out?

Correct. But they didn't have any products in the server market, at all. At least Apple has experience in that area, albeit in the past. Huawei? Nope. To the best of my knowledge, all of Huawei's CPU/APU/SoC designs come from HiSilicon, and HiSilicon launched their first CPU in 2012.

Huawei has been making servers for many years, as well as attempting some ambitious stuff like a large custom many-socket custom-interconnect system with Itanium that didn't ship. (You clearly don't mean server CPU's, since you mention Apple, which got its server CPU's elsewhere.)

 

[DrMrLordX]

Cavium got the former Netlogic team from Broadcom, as far as I can tell from raiding resumes - including David Hass, the architect on XLP and later Vulcan. Take a look at LinkedIn sometime. Check out the number of design engineers with both Marvell/Cavium and Netlogic listed.

Then maybe they can iterate upon Vulcan. I'm interested in seeing what they produce.

When ATI bought BitBoys, did it indicate ATI was incapable of good GPU design and was screwed in the long term?

Don't be ridiculous. If you compare ATI's product history with that of BitBoys, you will see that ATI had design chops. ATI bought out BitBoys to snap up some of their tech portfolio, nothing more:

https://arstechnica.com/uncategorized/2006/05/6737-2/

It's not like ATI cancelled their entire product lineup in 2006 and switched to selling Glaze3D cards. Compare that to when Creative Labs bought out Ensoniq and sold rebadged AudioPCIs for years until their own PCI sound cards were ready for market.

Cavium just chucked their own core design for ThunderX2 and used Broadcom's Vulcan instead. What does that tell you about the core design team at Cavium prior to their acquisition of the Vulcan? ThunderX wasn't a BAD product, but they knew they could do better on ThunderX2 with outside help.

Also, weren't you saying a few months ago that the entry of Qualcomm into the server market meant Cavium was screwed? How did that work out?

They would have been screwed if Qualcomm had backed their own effort. I have no doubt of that. For whatever reason, Qualcomm pulled the plug. I suspect it was done in anticipation of the Broadcom buyout.

Huawei has been making servers for many years

. . . using Intel chips, no? They've only recently been involved in designing their own server CPUs using custom ARM cores.

(You clearly don't mean server CPU's, since you mention Apple, which got its server CPU's elsewhere.)

Apple was a party to the design of the original PowerPC which they later used in their own XServe products. Granted, that is ancient history, but they have been involved in the design of server CPUs in the past.

 

[beginner99]

Also, weren't you saying a few months ago that the entry of Qualcomm into the server market meant Cavium was screwed? How did that work out?

That was simply because the guy owning broadcom doesn't like new experimental markets. Qualcomm ditched the server part for preparation for broadcom acquisition. Do you see the parallel and irony in it? Yeah broadcom ditched vulcan for the same reasons.

 

[DrMrLordX]

Yeah broadcom ditched vulcan for the same reasons.

Hock Tan was very forward about that in the press. To him, expanding market reach isn't worth it today. Better to improve efficiency in serving existing markets.

Having looked at some of the other ARM server offerings out there (Ampere, in particular), it looks like there are at least two designs moving forward that have been shed (Broadcom's Vulcan) or shuffled into a merger/buyout (Applied Micro's X-Gene). Now Ampere has their X-Gene3 while Marvell has ThunderX3 (codename: Triton), all aiming at CCIX and PCE4.0 functionality. If everything goes as planned, they should be making product announcements following on the heels of Huawei. It could be an exciting time in the ARM server space.

 

[moinmoin]

Correct. But they didn't have any products in the server market, at all. At least Apple has experience in that area, albeit in the past. Huawei? Nope. To the best of my knowledge, all of Huawei's CPU/APU/SoC designs come from HiSilicon, and HiSilicon launched their first CPU in 2012.

HiSilicon is and always has been a fully owned subsidiary of Huawei. That's like saying Samsung doesn't make any chips, after all it's actually SSI (Samsung Semiconductor Inc., wholly owned subsidiary of Samsung Electronics) that does them.

 

[DrMrLordX]

HiSilicon is and always has been a fully owned subsidiary of Huawei. That's like saying Samsung doesn't make any chips, after all it's actually SSI (Samsung Semiconductor Inc., wholly owned subsidiary of Samsung Electronics) that does them.

I think you are missing my point. Huawei's networking division has no history of custom CPU design. All their server designs and interconnect work have been on platforms from other companies (notably Intel).

HiSilicon introduced their first phone SoC back in 2012. Huawei/HiSilicon has more-or-less gone from being a networking and cloner phone company to a producer of server-class CPUs in six years.

 

[moinmoin]

Right, they did indeed. Your previous post made it sound like Huawei had no experience in any chip design which is of course located in their subsidiary HiSilicon.

 

[DrMrLordX]

Sorry if that didn't come across correctly.

 

[NTMBK]

With ARM cutting off Huawei, the future of their server CPUs seems in doubt: https://www.anandtech.com/show/14373/report-arm-suspends-business-with-huawei

In the latest event in the quickly moving saga that is Huawei’s technology export blacklisting by the United States Government, the BBC has published a report this morning claiming that IP vendor Arm has “suspend business” with Huawei and its subsidiaries. If this is correct, then it would represent a massive setback for Huawei’s hardware development efforts, as the company and its HiSilicon chip design subsidiary rely heavily on Arm’s IP for its products.

According to the BBC News report, Arm has almost entirely severed ties with Huawei, with the company instructing employees that they are not to “provide support, delivery technology (whether software, code, or other updates), engage in technical discussions, or otherwise discuss technical matters with Huawei, HiSilicon or any of the other named entities”.

I wonder what will happen next? Will they go rogue, and iterate on the ARM designs that they already have access to with their own silicon design teams? If they're already cut off from access to pretty much all Western tech, I doubt they will give a damn about breaking IP laws.