Watt the heck is this? A 32-core 3.3GHz Arm server CPU shipping? Yes, says Ampere

Carlyle Group-backed Ampere Computing, run by ex-Intel president Renée James, says it is, at last, shipping its 64-bit Arm-compatible server processor. It represents another attempt by the Arm world to grab a chunk of the lucrative data-center server market, which is virtually 100 per cent locked up by Intel and its x86-64 …

[Charlie Clark]

If you're used to cross-compiling code, playing with exotic architectures, and are patient, then you've probably had a smooth journey.

The problem isn't really cross-compiling. The toolchains for this across x86, x86_64 and ARM v7 and v8 are pretty well-established. And if the target market is large scale then compile-time tweaks are less of an issue than ease of deployment and power consumption. The bigger problem that has held ARM back in data centre is about drivers, which can't just be cross-compiled.

[Christian Berger]

"The bigger problem that has held ARM back in data centre is about drivers"

Yes and that's actually the main advantage of "x86" these days. With "x86" you have the "IBM-PC" a well defined hardware platform which allows you to boot your OS and ennumerate the hardware and talk to screen and keyboard in a rudimentary way without any special drivers. For many applications you can "port" an image designed for a server from one vendor to another one just by swapping the disk.

[HmmmYes]

No - OSes are not blocked by drivers. Drivers are portable across CPUs too.

Its testing thats the killer.

[Joe Montana]

Drivers?

"The bigger problem that has held ARM back in data centre is about drivers"

It's nothing to do with drivers, an ARM server will have drivers for the hardware fitted in it, and it's rare that a server will have anything else installed into it that would require drivers. The number of servers having nonstandard cards installed is very small. In fact, even on x86 servers all i ever see in the expansion slots are vendor-supplied storage and network cards.

The problem is application code, and specifically closed source applications. Linux has supported ARM for a long time, and the vast majority of open source code has already been compiled for ARM by various distributions.

Running Linux and open source apps on ARM is just as easy as x86, and has been for a long time. Many people are running applications on raspberry pi and other similar boards, the only thing missing is higher end ARM hardware aimed at datacentres.

[gnasher729]

Re: Drivers?

How difficult is it to compiler for ARM?

Every iOS developer knows it's trivial. There's the iOS simulator on your Mac running x86 code (32 bit or 64 bit, depending on the emulated device), and there's the real devices running ARM code (32 bit or 64 bit, depending on the emulated device). There are no problems with endianness, and you can even use SIMD code if you use the compiler functionality and not x86 or ARM specific extensions.

So unless your application has lots of x86 assembler code, there's no problem at all.

[Christian Berger]

Re: Drivers?

"It's nothing to do with drivers, an ARM server will have drivers for the hardware fitted in it, and it's rare that a server will have anything else installed into it that would require drivers."

That means you'll be limited to the kernel versions the manufacturer supports, which means that you'll have you typical "Smartphone" situation where you need manufacturer support to get updates.

[Anonymous Coward]

Re: Drivers?

That is the common model for enterprise platforms already.

[Loud Speaker]

Re: Drivers?

NO SCSI (SAS) in your servers?

Do you mean no tape backup??

Perhaps for a compute server, but database server without backup? I hope not!

[Alan Brown]

Re: Drivers?

WTF are you doing running backups directly on your database server? Are you a holdout from the 1960s?

These days you have a central backup server attached to the tape library and a client running on everything else. The backup server doesn't even need to be in the same building as everything else.

[Hans 1]

So this is targetted at the Xeon Platinum 8180 (https://ark.intel.com/products/120496)

$800 vs $10 000 ?

125W vs 200W ?

32 cores vs 28 ?

1TB vs 768 GB (RAM) ?

Can it run Crysis ?

$800 ? Wow, nice and cheap ... I want one - a nice toy for Xmas!

[Tom 7]

RE $800 ? Wow, nice and cheap ... I want one - a nice toy for Xmas!

I'd guess with the ARM NN software this would come close to some high end GPUs for performance for fewer beer vouchers and running costs?

[Anonymous Coward]

I was thinking that 42 PCIe lanes was a bit on the low side these days but I notice that the Xeon Platinum 8180 you used for comparison only has 48. Epyc has 128.

[Anonymous Coward]

The comparison Ampere make is to the Gold 6120:

$800 vs $2000 ?

125W vs 125W ?

32 cores vs 16 (32 with hyper-threading) ?

1TB total vs 768 GB (RAM)/socket ?

Performance on Ampere's benchmarks favours Ampere by around 20%, but there are no details on either what performance test this was on or it's ability to run Crysis.

[a handle]

"32 cores vs 16 (32 with hyper-threading)" but since Aug 1st 2018 hyperthreading needs to be turned off in environments such as hypervisors.

Our servers dropped from 32 to 16 logical cpus without any noticeable performance drop, according to tests by vmware (or was it intel) some environments run faster with hyperthreading disabled.

Hyperthreading is hype, it isn't helping much in many or most environments.

[guyr]

Re: $800 ? Wow, nice and cheap ... I want one - a nice toy for Xmas!

Meh, that pricing is in the neighborhood of AMD Threadripper: $899 US for 16 cores / 32 threads, 180 watts.. And those are full fat x86 cores. We'll need to see some *server* benchmarks to see if these ARM CPUs are indeed competitive.

[Alan Brown]

So this is targetted at the Xeon Platinum 8180

I was thinking the same thing (having just installed a dual-socket Xeon 8180 box for a project), but the acid test would be throughput per core.

Then again, if they could put 32 cores in for $850 then they can probably put in 128 cores and start running rings around Intel.

[jms222]

Drivers ?

> The bigger problem that has held ARM back in data centre is about drivers

Let's just assume that unless you're tasked with building drivers, this has been done and you have a working network stack and filesystem.

[Charlie Clark]

Re: Drivers ?

Why should we assume that? The lack of an ISA for ARM has been the biggest hindrance to wider adoption, of which I'm a great fan.

[joeldillon]

Re: Drivers ?

I'm sorry, what? ARM /is/ an ISA (instruction set architecture) and seeing as we are talking data centre the drivers are all part of the Linux kernel and generally portable. What 'drivers' do you think are lacking? Please be precise.

[Charlie Clark]

Re: Drivers ?

ISA = Industry Standard Architecture as in the IBM PC. This is what enabled the clones and hence entrenched the x86 eco-system. SoCs, which make up the majority of ARM chips, don't come with an ISA which is one of the reasons why it's so hard to get drivers for them.

[Adrian 4]

Re: Drivers ?

It's a valid argument for consumer devices (though designing to a decades-old standard creates many problems) but is irrelevant to a data centre, where you need only the peripherals you bought from the manufacturer (and often only disc and network).

And android phones seem to have very little difficulty running arbitrary Android apps. Because the OS isolates the user from the hardware, as it's meant to do.

We really, really don't want people coding applications for specific hardware any more. Meanwhile, with PCI and USB, drivers are getting more portable. They're more tied to the OS they run under than the platform the hardware is attached to.

[Peter Gathercole]

Re: Drivers ?

Probably crossed terms.

ISA also means Instruction Set Architecture, which is what the ARM ISA is.

Nowadays, pretty much all devices work through PCIe 3, so device drivers are much less of an issue than they were.

Most people building x86 systems see the legacy BIOS, keyboard, serial and parallel ports as being something that ought to be culled from modern systems (some have done it already), and I really don't think you mean that there is still support for the 8-bit 'ISA' adapters that were in the original IBM 5150 PC.

[joeldillon]

Re: Drivers ?

Again, what drivers exactly do you think these servers are missing? For what, your 30 year old Soundblaster that Linux probably does have drivers for in any case?

[Yet Another Anonymous coward]

Re: Drivers ?

How good are Supermicro's 10G ethernet drivers?

Are they 1% less efficient for ARM than Xeon?

Do I need to even worry about this?

Do I want to worry about this?

Do I need to suddenly care what motherboard is in my cloud providers servers ?

In the margin small differences make a difference

[Wayland]

Re: Drivers ?

I thank that's a fair assumption. If it's aimed at the server market then having network ports is the top requirement. If they use any of the existing network chips then the drivers are already written.

I think if you're not sure what you will use this server for then there is some concern that there is something it might not do. If the only way you can do the job is by running Windows Server then you've bought the wrong hardware. Same would go for running Windows VMs. However if it can be done on Linux then this can do it.

I suspect the only thing stopping people choosing these is lack of experience with non-x86 CPUs. Anyone who has used a Raspberry PI as a server will be quite confident to use this to scale up the project.

[Alan Brown]

Re: Drivers ?

"Let's just assume that unless you're tasked with building drivers, this has been done and you have a working network stack and filesystem."

The bigger problem is generally manglement resistance and/or x86-tuned environments.

[alain williams]

It runs Linux

Almost a year ago Red Hat announced Arm server support for their Linux. So, all the hard work is long done. I notice that CentOS (aka Red Hat) was working on this in February, so they have probably knocked the bugs out by now and this is ready for real customers.

[bombastic bob]

Re: It runs Linux

well, couple that with *NO* 'Management Engine' [or its equivalent] and I'll be wanting one for my desktop.

It'll need to run FreeBSD though... maybe I'd have to wait longer for that one

(nobody said - IS IT infected with a 'management engine' like Intel and AMD? Or were the designers security conscious enough to AVOID that trap?)

[Anonymous Coward]

Re: It runs Linux

> It'll need to run FreeBSD though... maybe I'd have to wait longer for that one

NetBSD?

[John Sager]

Re: It runs Linux

If it's a server chip it's almost certainly got a little ARM core to do low-level management stuff. Whether it's completely sealed off & spies on you I have no idea though.

[pwl]

Re: It runs Linux

... or SUSE Linux, which has been available with aarch64 support since 2016 <https://distrowatch.com/?newsid=09630>, including on Raspberry Pi 3 <https://www.raspberrypi.org/blog/suse-linux-enterprise-server-for-raspberry-pi/>

[Ilgaz]

Re: It runs Linux

I think the issue is the support from Oracle, IBM kind of companies. That is actual, perfect support not just press releases.

[Cronus]

...but does it suffer from Spectre et al?

[EnviableOne]

RE: @cronus Spooknotes speak volumes

As for the Spectre and Meltdown CPU vulnerabilities – which affects a range of CPU architectures, from Intel'x x64 to Arm's Cortex-A families – Ampere's Taylor told us this:

Patches have been installed. As with other Arm-based processors, there are vulnerabilities. For eMAG and all future generations, architectural changes were made and will be made now to address the Arm fix for Spectre and Meltdown.

[defiler]

125W for ARM?

Ouch! Ee when I were a lad and all that...

Not saying it's not doing something useful with that energy, but still. Ouch.

[Anonymous Coward]

Re: 125W for ARM?

Well there are 32 of them in there. So about 3.9W each.

[Anonymous Coward]

Re: 125W for ARM?

"Well there are 32 of them in there. So about 3.9W each"

The upcoming AMD TR2 with 64 cores should equal that if it is a 250W part.

[Anonymous Coward]

Re: 125W for ARM?

The upcoming AMD TR2 with 64 cores should equal that if it is a 250W part.

The TR2 is a 32 core chip with 64 hardware threads.

[Anonymous Coward]

Re: 125W for ARM?

The TR2 is a 32 core chip with 64 hardware threads.

Yes, that's the current model. That's why I said "upcoming".

https://wccftech.com/amd-7nm-zen-2-rumors-16-core-am4-32-core-threadripper-64-core-epyc/"

[Anonymous Coward]

Re: 125W for ARM?

Hmmm

The 7nm Based EPYC Rome 64 Core (128 core logical) CPU is already in the pipe and is scheduled for release in 2019. Oh and the author is wrong instating that Intel still has 100% of the market. Not even close. AMD has been whittling away at Intel's hold on the market for some time now.

[gnasher729]

Re: 125W for ARM?

"Well there are 32 of them in there. So about 3.9W each."

I heard there is an iPhone XI prototype with this chip. Empties the battery in 16.9 seconds. Explodes your charger after 1 min 15 seconds.

[JohnGrantNineTiles]

Re: 125W for ARM?

What voltage do they run at? Guessing around 1.2V, 3.9W per core must be about 3A. Whatever happened to "MIPS per milliwatt"?

[Alan Brown]

Re: 125W for ARM?

"Ouch! Ee when I were a lad and all that..."

That's at full song. The idle power will be interesting to see.

[Wolfclaw]

Another point, if Windows On Arm ever gets a running start, thes ewill make very power WIndows workstaions at a fraction of the price Intel and even maybe AMD can match. Image one of these power your home server setup ??

[Ken Hagan]

"very powerful WIndows workstaions at a fraction of the price Intel and even maybe AMD can match"

Citation needed, I think. Admittedly the article only quotes a price for a very powerful chip, but it is still pretty eye-watering. I see no reason to assume that an ARM-based chip would be cheaper than anything AMD might offer in their x86 line, at a given level of performance.

Unless someone comes up with mid-price and low-price variants, and delivers an ISA (see earlier comments), the notion of buying a cheap, generic, ARM-based "PC" and sticking a standard OS distribution on it (either Windows or Linux) without *then* spending weeks hunting the internet for drivers to run all the interesting in-box peripherals (like sound and graphics) ... is just a notion.

[Oneman2Many]

CPU is just small part of the overall cost. And as others have said, there is nothing to suggest the CPU is going significantly cheaper,

[LeahroyNake]

AMD

Nice to see AMD getting a mention in the last but one paragraph.

Fingers crossed Arm and co get to take a slice out of Intel... until they start paying suppliers for 'preferential treatment' sorry I meant something else, apparently you can milk anything with nipples.

[Crazy Operations Guy]

Don't go after the Hyper-scale market

They should go after the start-up market. Worm into the small, growing markets, then profit off them once they are behemoths. Really, the best tactic would probably be to tweak the architecture to be optimal for Machine Learning, Node.js, software-defined-whatever, or whatever today's fancy new shiny technology people are using to spackle over poor planning and terrible coding practices. And if it helps sell it, call it something like the "Raspberry Wedding Cake" so they can exploit how much the mainstream tech press loves to fawn over the RasPi while also showing how much more powerful it is (IE, its like a Pi(e) but can serve many, many more people with one).

Intel unseated the big players in the IT space not by marketing to the customers that were perfectly content with their rooms full of mating dinosaurs and dumb terminals. But rather they went after the home user, the small businesses, the people and organizations that could never even do more than dream about having a computer of their own.

[Oneman2Many]

Re: Don't go after the Hyper-scale market

don't you know startups only use the cloud. Why would they want to buy servers ?

[Anonymous Coward]

Has no one learned the Calxeda lesson?

1. There's no price benefit over x86 CPUs: this comparison "Intel = $2000", "32-core ARM = $800" is ridiculous. Intel's cost is maybe $200, Ampere's cost is maybe $1000. If there's one thing Intel excels at is economies of scale for a general-purpose CPU.

2. The entire benefit of ARM is being a system-on-chip. So you can integrate it with specific peripherals or specific hardware offloads relevant for your application. E.g. you can create a dedicated CPU for AI or for databases or for storage. *Not* having anything specific on the ARM server makes it a general purporse CPU where it will always lose to the Intel. That's why Calxeda, well-funded, failed abysmally, and Annapurna, with zero marketing, became the back-bone of AWS. Annapurna integrated the ARM CPU to a NIC and specific hardware offloads to solve a specific problem. Calxeda was a general purpose CPU.

3. There are many CPU performance benchmarks yet none are mentioned here. A 32-core ARM even at the same clock will lose to the more sophisticated Intel because of the complexity of the Intel pipeline, the way its integrated into L1/L2/L3 caching, extremely intensive investment into PCIe/DDR/etc. Am I wrong? Where's the benchmark comparing this CPU to the Intel?

4. So this startup I never heard of wants to succeed where *Qualcomm* failed? Please. What is the innovation here?

[Rajesh Kanungo]

Re: Has no one learned the Calxeda lesson?

First sensible response. Maybe there (4) Where is the innovation? (Business, power, performance, management, etc.) should be mentioned more clearly. Not clear if having a totally different architecture works for businesses as it competes with an already x86 humming server ecosystem.

Maybe the authors or inventors haven't been able to articulate the win or they are keeping that under wraps.

[John Smith 19]

So if everyone runs Linux, why has Intel still got a lock on the data centre?

Because Windows only runs on x86 architecture?

Or is it this particularly dysfunction co-dependent relationship got something to do with it?

Maybe hiring an ex Intel bigwig might give them some insight into what's really needed to make a difference?

[P. Lee]

Re: So if everyone runs Linux, why has Intel still got a lock on the data centre?

Because Intel does high-power systems well and DC's keep their CPUs busy. They do not race to idle.

[Rajesh Kanungo]

Why is it better?

Some of the issues I see are:

Negatives:

1. Software: The Servers including virtualization tools are all standardized around x86 (Intel, AMD). In fact, there is a lot more invested in software than in the HW.

2. Architecture: If the competition is based purely on architecture, the server team will look for highly dedicated (NVIDIA type), or the generalized architecture. A general purpose ARM CPU doesn't buy you anything special. Why not stick to AMD or Intel?

3. Silicon: TSMC may have an advantage in Silicon over Intel but Intel has been pumping out these processors for a long time and beating them on Silicon is an iffy strategy at best. If AMD is having a hard time competing then how can you expect a whole different architecture to win?

Positives:

1. Watts: Clearly ARM has an advantage when it comes to lower watt CPU but at the higher end, I don't see it being a deal breaker unless the difference is huge. Nothing seems to indicate otherwise.

2. Bottom up: Remember how DEC stole the market from IBM, Sun from DEC, and then Intel from SUN? My believe is that when the majority of the CPU's are ARM systems, which they already are, that they will slowly move into the mainframe.

However, this will take time.

3. Customization: ARM Cpu's can be customized at a faster rate than intel CPU's. ARM CPU's with full blown AI engines, DSP's etc. are very common.

4. Multiple vendors for CPU choices: Unlike Intel or AMD, ARM allows anyone to find a vendor with close enough specs to what they need. If there isn't one then you can have a vendor design and spin one for you. You want one with 3 DSP cores, 2 Neural network cores, etc. and can't find one? Ask a vendor to make you one.

Apple did something like this in-house for their iPhone since Intel can't put everything they want in the CPU at the rate they want it. (plus the power consumption controls ...)

Ultimately ARM will win because of the speed of customization is in months and it will keep on eating into the x86 market. It will, however, take time.

[Anonymous Coward]

Re: Why is it better?

> 2. Bottom up: Remember how DEC stole the market from IBM, Sun from DEC, and then Intel from SUN?

Intel did not win the race against SPARC. Intel was brought in data centres on the shoulders of Linux. The transition from SPARC to Linux@Intel was very noticeable at the and of '90, then the scientific community and universities began their migration from Unix to Linux@Intel. Why? For several reasons: a) system's price/performance was good enough; b) the existence of huge community around Linux minimized support risks c) $0 price tag on the operating system and compilers was more better choice than it was offered by commercial world.

So, Where is Intel's role? What Intel need to do, in Intel's HQ they must erect a monument with two words "IBM & LINUX" and preferably it should be done in gold.

m.r.p.

[Stuart Halliday]

You just know that somewhere, sometime someone will try to run RISC OS on it! :)

[Anonymous Coward]

Cloud Super Eight, and no mention of IBM!?

[Hstubbe]

IBwho? Do they still have skilled people working for them? I thought they fired everyone.

[Androgynous Cow Herd]

Nope

<quote>

IBwho? Do they still have skilled people working for them? I thought they fired everyone.

</quote>

Nope, just the old people.

[ssieler]

prior ampere

I wonder if Ampere knows that their name conflicts with a prior Ampere? (I have two Ampere laptops in my house :)

[John Savard]

Locked Up?

I was aware that IBM manufactired, and offered for sale, data centre servers built around the PowerPC architecture.

I don't know how many of them are in use, so it perhaps is entirely possible they haven't made a dent in an x86 near-monopoly. Obviously, the x86 world will be the most competitive, and likely the cheapest. Until, of course, ARM's entry.

[-tim]

Re: Locked Up?

"I was aware that IBM manufactired, and offered for sale, data centre servers built around the PowerPC architecture."

IBM and Sun opps... Oracle still make same fantastic insanely powerful hardware. The problem is nearly no one has a problem where they need that much power. Both of them also have the problem that you can no longer start small. Sun's last server comes in sizes that are the price point of a very nice new car, a very nice house or very nice house in the Bay area. There is no way the future decision makers will ever get to play with that sort of hardware so there isn't much research being done about making use of some of the newer concepts like fully compressed and encrypted memory. The very big machines from last year can map tends of terabytes of a file into ram and then go through it with a thousand threads. What that is cool, it isn't a problem most companies have.