Would have been good if just to drive down Intel's Xeon pricing.
Industry whispers: Qualcomm mulls Arm server processor exit
Servers powered by CPUs based on Arm Holdings IP sound like a good idea. They’re a good chance of being denser, more powerful and less costly to run than the Intel CPUs that handle the overwhelming majority of server workloads (and scoop most of the server CPU profits too). Because Arm dominates the mobile world, server CPUs …
COMMENTS
-
-
-
Tuesday 8th May 2018 14:32 GMT FIA
Microsoft already have the top-tier ARM architectural licence AFAIK. [...]
Indeed they do if this random internet news site is believable. ;)
And lets not point out the issue with "do an Apple and control the hardware"...in that they use ARM designs as well.
No, Apple control the hardware, they use the Arm instruction set, not Arm CPUs.
Like MS they have an architectural licence, which means they're allowed to design their own CPUs to implement the Arm instruction set, rather than designing SoCs using an Arm CPU core, such as the Cortex, like most Arm licences do.
There's 2* levels of Arm licence, core and architectural. A core licence allows you to take an Arm designed CPU core and add your own IP around it to create a SoC. The (much pricier) architectural licence allows you to design the CPU portion too; it just has to pass the Arm validation tests.
This is why Apple's CPUs are so quick compared to the competition.
* Okay, that's not quite true.
-
-
Tuesday 8th May 2018 08:10 GMT Anonymous Coward
Why should ARM Holdings help?
The market for server CPUs will never be remotely as large as the market for CPUs for portable devices. If IoT grows to a size anything like analyst wet dreams, it will be orders of magnitude larger.
Before anyone suggests that they could charge much higher royalties for server CPUs than they do for e.g. smartphone CPUs, the whole premise driving ARM server CPUs is that they'd be a lot cheaper than x86 CPUs when you're talking about filling a cloud datacenter with hundreds or thousands of racks full. Maybe they could get away with a few bucks each, but start talking about tens of dollars and they are going to make ARM server proponents a lot more shy. At $5 each there would have to be 50 million sold to offset the design costs - and that money is wasted if instead there are only 5 million, when those engineers could have been designing mobile cores to beat back custom mobile cores from Samsung, HiSilicon, etc.
ARM Holdings is right to concentrate on mobile, and designing cores that go in even smaller and lower power IoT devices, and leave it to others to take the risk of designing server cores that currently don't have a market. If someone else creates that market, maybe they will think it is worth getting involved.
-
Tuesday 8th May 2018 08:22 GMT ToddRundgrensUtopia
Re: Why should ARM Holdings help?
@Doug,
I think your £5 per server chip is out by a couple of orders of magnitude. For something that competes on power/watt with Intel, they cab charge £500 and be very very competitive with Xeon. Two problems remain however, a) how much does the server motherboard cost and b) where are the applications outside of webservers?
-
Tuesday 8th May 2018 14:46 GMT Nate Amsden
Re: Why should ARM Holdings help?
El reg reports the new ARM server chip is going for $800(16 core) $1,800(32 core) in 1,000 unit quantities. Also Qualcomm's server chip going for $2,000(48 core)
https://www.theregister.co.uk/2018/05/08/cavium_thunderx2/
I have found it interesting that the server arm chips do draw a lot of power - the article above mentions power from 75W to 180W(32 cores - which seems to be in the same ballpark as a 32-core AMD Epyc?)
-
Tuesday 8th May 2018 21:19 GMT Anonymous Coward
Re: Why should ARM Holdings help?
You think they could get away with charging £500 per chip in licensing costs? Pull the other one! Those prices are higher than an Intel CPU for 1S servers that are used in cloud environments!
You could charge that sort of licensing cost for 4S/8S I suppose, but ARM has no chance of ever breaking into that market.
-
-
Tuesday 8th May 2018 08:37 GMT Alan Johnson
Re: Why should ARM Holdings help?
Cost is a factor, as is support/infrastructure, familiarity and perceived risk but the costing model you are talking about is far from reality. A high end server processor must have a price measured in hundreds of dollars. An embedded processor especially in high volume is going to be a dollar or two at most and exists in a market place overflowing with viable alternatives and cut throat competition. The embedded processor is vastly different from the server processor.
The issue any ARM based device faces is the difficulty of displacing/competing with an established dominant supplier. A small technical/commercial edge is not enough in this situation, a sustained and major advantage is needed to vercome the infrastrcuture, support and familiarity issues.
-
Tuesday 8th May 2018 10:44 GMT Flocke Kroes
Perspective from the small side
I am fairly sure my experience is the complete opposite of people responsible for racks of servers with high utilisation.
I have no problem finding the software I need for ARM. I am familiar with it. I can rarely find an Intel small enough or cheap enough for the task at hand. When Intel comes within a mile of selling something suitable, it either needs a fan or turns out to be broken with a "Won't fix because embedded has no budget". This has given me a strong preference for making the problem smaller rather than buying Intel. The three ARMs I have an ssh connection to have uptimes of over 200 days. The Intel has an uptime of 5 days (to be fair it often lasts 30).
There is a good reason I do not expect big ARM servers this year: If Google or Amazon do not like Intel's prices they can buy an ARM license and new Intel prices will arrive promptly. (My prediction for next year has about the same chance of being right as Gartner.)
-
Tuesday 8th May 2018 21:56 GMT Richard Plinston
Re: Perspective from the small side
> The Intel has an uptime of 5 days (to be fair it often lasts 30).
Some of my client's machines (Intel and AMD) have had runs of over 1,000 days, for example just now:
[root@nzedi00 ~]# uptime
09:53:16 up 1069 days, 54 min, 1 user, load average: 0.17, 0.15, 0.10
[edi@nzedi01 ~]$ uptime
09:54:38 up 848 days, 2:00, 1 user, load average: 0.09, 0.04, 0.01
-
Wednesday 9th May 2018 08:20 GMT Anonymous Coward
Re: Perspective from the small side
>The three ARMs I have an ssh connection to have uptimes of over 200 days. The Intel has an uptime of 5 days (to be fair it often lasts 30).
If uptime was the main metric (and it is a very high priority for mainframes) then VAX/VMS seems hard to beat with uptimes typically measured in years. I used VAX/VMS in cluster years ago and it was indeed very reliable.
Time to resurrect VAX architecture?
-
Wednesday 9th May 2018 08:50 GMT Anonymous Coward
Uptime is not a measure of CPU quality
I think I only remember a couple cases were systems crashed or needed to be taken down due to CPU problems. Pretty much every other hardware component is a more likely culprit for problems.
Of course hardware itself is rarely the issue, 98% of the time your uptime is reset to 0 because of patches. Speaking of which, I sure hope those Intel servers with 1000 days of uptime are on an isolated network, because they are obviously not at all current on patches no matter what OS they are running!
-
Wednesday 9th May 2018 19:58 GMT Richard Plinston
Re: Uptime is not a measure of CPU quality
> I sure hope those Intel servers with 1000 days of uptime are on an isolated network,
Absolutely. Headless and well behind the firewall(s).
> because they are obviously not at all current on patches no matter what OS they are running!
With Linux, unlike Windows, it is only new kernel versions that would require a reboot. Other patches and updates are installed without a reboot*.
* the inode system caters for files being replaced even when they are currently open. The old files (program/library) are still used by the processes that have them open until each is all closed and then that file space is recovered. It is only necessary to restart the updated servers.
-
-
-
-
Tuesday 8th May 2018 21:48 GMT Richard Plinston
Re: Why should ARM Holdings help?
> the whole premise driving ARM server CPUs is that they'd be a lot cheaper than x86 CPUs
No. The whole premise for using ARM server CPUs is that the savings in electricity costs will be greater than the cost of replacing all the x86 based hardware. Not only will the CPUs use less watts per MIPS but most CPU cores can be shut down to use zero watts when load reduces.
-
-
Wednesday 9th May 2018 21:45 GMT doublelayer
Re: Why should ARM Holdings help?
"[T]he WHOLE point of cloud is too pool customers at such scale that they can keep all cores utilized."
Is that really the case? Even cloud must have comparative downtimes. If some data center in Europe needs 100 units of performance for peak consumption by clients, which probably happens during working hours if these are business clients, then what would the requirement be at 3:30 in the morning. I don't doubt that there are people using those systems then, nor that there are tasks the servers perform off hours, nor that there may be businesses in other areas that are operating during the day and therefore using more power at European night than in European working hours, but even so I'd guess that there is a noticeable difference. If only 85 units are required, then using 85 or 90% of the cores and shutting the rest down for power savings can't hurt, right? Also, this could be a factor for companies that have systems not on the cloud but still need quite a bit of performance.
-
-
-
-
Tuesday 8th May 2018 09:05 GMT mark l 2
it will take a big player such as Google, Microsoft or Facebook with deep pockets to announce they are going to go with their own custom ARM server CPUs for others to wake up and take ARM servers seriously.
I am a little surprised after the whole meltdown fiasco that people are still buying Intel hardware at what appears to be the same rates as before, despite the fact that the meltdown bug still exists and at current is only fixed with a patch that slows down the hardware.
Although AMD have made some in roads I think it would take a hardware flaw that couldn't be patched and required physical CPU replacement to shift Intel out of the data centre.
-
Tuesday 8th May 2018 13:31 GMT Anonymous Coward
"I am a little surprised after the whole meltdown fiasco that people are still buying Intel hardware at what appears to be the same rates as before, despite the fact that the meltdown bug still exists and at current is only fixed with a patch that slows down the hardware
We buy several thousand processors a year. It's not just a simple case of not buying them or "simply swapping to AMD", there is all the associated "crap" to go with them. We do buy AMD, just not in the same amounts.
-
-
Tuesday 8th May 2018 10:09 GMT Malcolm 1
Cloudflare
CloudFlare seem to be going all-in for ARM. Naively I would expect them to have a sufficiently signficiant scale to cause manufactures to take notice?
-
Tuesday 8th May 2018 12:17 GMT Anonymous Coward
Respect Intel's intellectual property rights..?
Maybe Qualcomm has to pay to-much to Intel to emulating its (x86) instruction set on there ARM-server CPU's
and can't compete on price for the big cloud server-farm boys ?
https://www.theregister.co.uk/2017/06/09/intel_sends_arm_a_shot_across_bow/
-
-
Tuesday 8th May 2018 16:25 GMT Anonymous Coward
Re: RISC-V is the future
Not sure why that post about RISC-V was down voted. After all Qualcomm has joined as one of the main partners in RISC-V. Also ARM is RISC but RISC-V is an based on the experience gained since ARM was launched more than 20 years ago and not encumbered by need for any backward compatibility.
And Qualcomm is not alone, Andes Tech who also have their own preexisting designs has joined RISC-V.
Perhaps the downvoters could explain to us just why RISC-V is not the future?
-
-
Wednesday 9th May 2018 08:25 GMT Anonymous Coward
Re: RISC-V is the future
>Vested interest in either their employer or their big bad geek game rig powered by X at home?
OK, so a baseless guess is enough to downvote then. The sneering in here is reaching new depths. I am perhaps old fashioned when I believe in facts rather than suspecting underhand motives behind every posts.
-
Wednesday 9th May 2018 09:38 GMT isogen74
Re: RISC-V is the future
I suspect the issue is exactly the same issue as this article. The problem isn't the instruction set - the problem is getting the software ecosystem built around it.
What does RISC-V bring that Arm doesn't have already? As far as I can tell, the only benefit is that it is "free", so the saving is a little bit of licensing cost. However, in the grand scheme of things the Arm fees are pretty inexpensive compared to everything else it costs to build a server grade processor (best guess, given their publicly stated revenue breakdown and shipped CPU volumes every year, so I doubt anyone with enough money to build a competitive chip here really worries about it - it's in the noise.
What does RISC-V not have? A mature architecture (it's fragmenting rapidly with extensions), or any form of mature software ecosystem. The Arm architecture is mature, and has been around at scale for at least 20 years, and yet according to this article is still struggling with a mature-enough software ecosystem in new markets.
Hell, it's exactly the same problem as mass-adoption of desktop Linux too, or VHS vs Betamax. Just because something might (and that's still unproven) be technically good doesn't mean it's going to be well adopted if the alternatives are easier to use and/or marketed better and/or better established.
-
-
-
Tuesday 8th May 2018 20:42 GMT Anonymous Coward
Re: RISC-V is the future
> Perhaps the downvoters could explain to us just why RISC-V is not the future?
Before we talk about the future, let's focus on the present.
Can you provide a reference to a RISC-V server board available for purchase now?
Is there any compiler support for RISC-V that is more than a toy/experimental hobbyist project?
Is there a RISC-V Linux port?
Without a working board, without a working compiler, and without a working OS, it seems kinda premature to declare that RISC-V is the future.
-
-
Tuesday 8th May 2018 22:49 GMT doublelayer
Re: RISC-V is the future
>Perhaps the downvoters could explain to us just why RISC-V is not the future?
I didn't vote either way, but I'll take a crack at it. I have nothing against that ISA. It's fine, and it shows some promise. That's not enough. We don't really need to keep inventing new ISAs for us to decide between; if we can get the ones we have running faster with less power, that gives us what we want. If this one will run really fast with low power and is relatively cheap to make, we will probably want it. However, it's at such an early stage that nobody can run on it now nor realistically plan on it for the near future. There have been enumerable new technologies that would have been great, if everything people said about them was true and we actually got it, but turned out never to become reality or was not that great compared to what we were told. For example, I've heard that basically every new file system for Linux/BSD servers would be the future. ButterFS was going to solve major problems. ZFS was nearly perfect. Even apple and microsoft did that. APFS will speed up the mac. REFS will do, actually I don't really know what it does but microsoft sent me a document about it, and I'm sure that it has some new features. Some of these will end up being very useful. Others won't be used in the longterm.
Thus, when you tell us that something is the future, without evidence of the future getting started now, a lot of us will assume that, whatever happens, it's not relevant now. In many cases, it won't be relevant at all. Everyone has at least one time where they said that something would be pointless and turned out to be wrong, but unfortunately everyone usually has many times where that turned out to be right.
-
-
-
-
-
Tuesday 8th May 2018 12:54 GMT Fenton
Memory Costs
These days with such high core counts it's generally the memory costs that determine the server cost.
Do you run a single server with high core counts and high memory requiring 64GB or 128GB Dimms or lost of small ARM servers with 16/32GB Dimms. All of a sudden real estate becomes an issue.
-
Tuesday 8th May 2018 13:56 GMT Lennart Sorensen
If the chip makers wanted there to be an arm server ecosystem, the first thing they should do is actually sell the stupid things to anyone that wants to buy one. So far it has been just talk with pretty pictures and lots of specs, but just about nothing anyone could actually buy if they wanted it.
If you think you can stick with only talking to the large cloud providers and people doing custom designs, well then you are not going to get any ecosystem going because the people that would actually play with it and get it working with lots of software can't get your system.
-
-
Tuesday 8th May 2018 14:51 GMT Nate Amsden
Re: It's the software.
I assume you are referring to Windows software? Since Linux has run on ARM for well over a decade I believe, looks like there is Java for ARM as well which covers a huge selection of software that run on servers.
Though I'd wager in many cases the code isn't as optimized for performance as on x86.
-
Tuesday 8th May 2018 16:18 GMT YourNameHere
Re: It's the software.
Yep. This is what people forgetting. All of the drivers, the plug and play, the interrupts and all of the fiddly bits. All the PCIe drivers, of the custom stacks that are designed for the features of x86. This all has be to tuned for any new ARM chip. Talk to Apple–IBM–Motorola about PPC's impact on the server space. Better HW, more SW support than ARM guys are giving but failed miserably. Envision a plane on fire crashing into the ground....
-
-
Tuesday 8th May 2018 16:37 GMT Anonymous Coward
The Political Angles
Business does not depend on tech alone. And here there is a lot of top level politics going on.
Qualcomm:
- has had a lot of conflict in China. The troubles are not over.
- has been forced to transfer tech to China
- has taken an architecture license from ARM
- has been forced to enter into a JV with Chinese companies
ARM:
- has been sold to a company that has racked up a lot of debt and might be forced to sell, possibly to Chinese companies
- has very recently entered a JV with China where China has 51% of the shares
China:
- is set on tech transfer from the West
- is willing to spend a LOT of money
- has received a lot of knock backs from the US
- has taken control over parts of the MIPS ecosystem
- forces foreign companies into JVs that tend to be dominated by the Chinese side
China is playing to win.
At the same time RISC-V is coming and is reaching industrial stage. The partner list looks like who-is-who of the CPU world. To them RISC-V is not just more modern but also less encumbered by politics.
If I am right Qualcomm will drop ARM and resurface with RISC-V. That is the test of this thesis.
-
Wednesday 9th May 2018 06:30 GMT Anonymous Coward
Re: The Political Angles
Since someone didn't like this I'll add some more links to back this up:
Softbank Has More Debt Than Venezuela https://www.electronicsweekly.com/blogs/mannerisms/dilemmas/softbank-debt-venezuela-2018-05/
And Softbank is the present owner of ARM.
Arm Transfers Microprocessor Technology to Chinese JV https://www.eetimes.com/document.asp?doc_id=1333267
ARM owns only 49% of this.
-