Ampere computing will ship their 32 core 3.3Ghz ARM Server CPU this week

[Master Disaster]

This was talked about at the start of the year and now it's here. Ex Intel Boss Renée James and her company Ampere Computing will aim their new chip at the top 8 web companies as cheaper (both to buy and run) server chips for running things like web servers as well large scale network appliances and other targeted servers.

Quote

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 Xeon bruisers.

 

This plucky upstart will be pitched at the hyperscalers – the cloud world's Super Eight in the US and China: Google, Amazon, Microsoft, Facebook, Baidu, Alibaba, Tencent, and China Mobile. The processor could be used to power their web servers and infrastructure boxes.

 

The technology will also be dangled in front of anyone who wants to build edge servers, software-defined storage and networking appliances, and other targeted solutions.

The specs are fairly impressive for what's essentially a low power chip,

Quote

In summary: it's a 32-core 64-bit Armv8 CPU clocked up to 3.3GHz in turbo mode, with a shared 32MB L3 cache. It supports up to 1TB of DRAM from 16 DIMMS plugged into eight DDR4-2667 memory controllers, has 42 lanes of PCIe 3.0, draws up to 125W, and is a single-socket chip. It'll be made using 16nm TSMC FinFETs. The cores were designed by Silicon Valley-based Ampere, and the whole package is branded the eMAG.

 

The single socket approach is not seen by Ampere as a particular weakness at this stage, by the way: it avoids NUMA headaches, and not every application needs dual or more sockets. The 16nm is also not a typo: it was apparently deemed by Ampere to be the most robust process node available at the time the chipset was being designed.

 

A 32-core eMAG part will set you back $850, or $550 for a 16-core version.

They're already teasing a 7nm second generation model with dual socket support.

Quote

Interestingly enough, the second-generation processor will support multiple sockets, use 7nm TSMC FinFETs, and is due to land next year. Another two further generations are being planned.

 

"Hyperscalers buy roadmaps, not products, and they've told us they were impressed by the quality of our roadmap," Matt Taylor, Ampere's senior veep of worldwide sales and business development, told The Register.

You will be able to order servers from Lenovo by the end of this week with Ampere having their own servers ready sometime soon.

Quote

Right now, if you want to get your hands on the first generation, you can from this week, we're told, order servers powered by the chipset from Lenovo, or soon white boxes direct from Ampere. Other system manufacturers are being wooed to ship the processor family, too. Oracle, for one, seems particularly keen to run its flavor of Linux on the hardware.

It seems as though Intels current struggles with die shrinkage makes this a perfect time for the Arm chips to launch a counter attack

Quote

Arm server chips have come and gone – the highest profile casualty recently being Qualcomm's beleaguered Centriq – and no one's made a dent in Intel's monopoly. While IT buyers await benchmarks and machines to experiment with, to ascertain the eMAG's crucial performance per watt per dollar stats, what on Earth is Ampere doing to ensure it too isn't another footnote in semiconductor history?

Taylor told us it comes down to three things. One being that the Armv8 CPU architecture is at the stage where it is ready to be used in server-class processors.

 

The second is that, in Taylor's view, the likes of TSMC are now churning out pretty decent chips in high volume for customers of all sizes that, in theory, can compete with semiconductors built by the traditionally dominant factory: Intel. It used to be that Chipzilla led the pack, and everyone else followed. Now, especially with TSMC's 7nm on the horizon, even small upstarts can or will be able to get hold of enough quantities of near-leading-edge or leading-edge chips while Intel stumbles around in its 10nm quagmire.

 

Intel is losing its leading edge, in other words, while fabless outfits like Ampere can go to contract manufacturers like TSMC and get volume amounts of competitive chips fabricated. This may give them a fighting chance against rival Xeons. Again, in theory: it all depends on how performant the final silicon is, of course. Ampere claims the eMAG's performance-per-watt will be "compelling."

 

The third point is that the Arm software ecosystem is maturing for enterprise-class workloads. Up until around about now, it's not always been an easy ride if you've tried to build and run non-trivial tools and applications for and on Armv8 server boxes. If you're used to cross-compiling code, playing with exotic architectures, and are patient, then you've probably had a smooth journey. For folks used to x86, though, it can get a little hairy if you go off the beaten track.

 

Now, according to Taylor, it's easier to build and run server-grade applications and stacks for the 64-bit Arm architecture.

 

Also, there is a fourth point we should mention: Ampere has been hoovering up chip designers from Qualcomm, Intel, and Marvell – so the team, while fairly small and nimble, has a lot of experience.

 

"It's the right time for us to puncture the market," Taylor concluded.

 

It may also be the right time for AMD's Epyc or Cavium's ThunderX2 to swoop in, or for Intel to respond with another Xeon D family – its answer to the last time rivals tried to rattle its money-stuffed cage in the data center.

https://www.theregister.co.uk/AMP/2018/09/18/ampere_shipping/

 

I would be very interested to see Linus put one through its paces and see how it performs. When they say "upto 125w" is that an under 100% load number or an Intel number? We've all seen Intel chips draw upwards of 400 watts.

[vanished]

I can't help but wonder how the performance compares.  I know it would be difficult to test since you're going across instructions sets but I do wonder.

I can see the big companies like Amazon, etc. being interested in this what with the chaos that the many Intel vulnerabilities have caused for them, but at the same time, what Linus says about the server market valuing reliability and proven results over the latest thing is true and for good reason, so that might hurt adoption of this for the next several years.

[Sauron]

If the performance is there it sounds great!

[Master Disaster]

I just realised something, zero mention of ECC RAM support. That's kind of a big deal when running pretty much any mission critical server.

[Firewrath9]

I see an 8180 on it...

intel will sue

[jagdtigger]

4 minutes ago, Master Disaster said:

I just realised something, zero mention of ECC RAM support. That's kind of a big deal when running pretty much any mission critical server.

https://amperecomputing.com/product/

Quote

Enterprise class features including ECC (error-cor-recting code) and RAS (reliability, availability and serviceability).

 

[Master Disaster]

1 minute ago, jagdtigger said:

https://amperecomputing.com/product/

 

Ah ok, so it was an oversight by The Register.

[jagdtigger]

3 minutes ago, Master Disaster said:

Ah ok, so it was an oversight by The Register.

It wasnt, its just that the specs didnt changed so they only point it out an link to the spec sheet:

Quote

The specifications of this first-generation Ampere part haven't changed from what was teased in February.

 

[AlTech]

*Sees Ampere in Title*

*Expects Nvidia Topic*

*Opens Topic and sees ARM*

*Close Tab* /s

 

No but seriously when I clicked on it I thought this was Nvidia news Unless Nvidia is conspiring to make ARM server chips secretly

[AlTech]

Also, Intel is facing much more problems than just Die shrink right now .

 

HPE and Dell basically flat out told customers that it's almost impossible to buy a product with a Xeon based on 14nm+ right now.

 

I don't see 64 Bit ARM servers taking off unless a large majority of code is re-written to support ARM64.

[jagdtigger]

4 minutes ago, AluminiumTech said:

Also, Intel is facing much more problems than just Die shrink right now .

 

HPE and Dell basically flat out told customers that it's almost impossible to buy a product with a Xeon based on 14nm+ right now.

 

I don't see 64 Bit ARM servers taking off unless a large majority of code is re-written to support ARM64.

Well most(if not the entirety) of the Debian Software library is already done because of the Raspberry Pi taking off...

[Master Disaster]

7 minutes ago, jagdtigger said:

Well most(if not the entirety) of the Debian Software library is already done because of the Raspberry Pi taking off...

Would Raspbian run on the same repos as say, Debian Server? I've got to imagine theres some server specific applications not yet ported.

 

Still you're correct, most of the work is already done.

 

[jagdtigger]

3 minutes ago, Master Disaster said:

Would Raspbian run on the same repos as say, Debian Server? I've got to imagine theres some server specific applications not yet ported.

 

Still you're correct, most of the work is already done.

 

No, it has its own repos. But you can get armhf builds directly from the debian site, that one will use the same repo if im not mistaken.

[Guest]

125W though. That will be a bit naughty to cool. I think maybe everyone will just wait a year for the 7nm version rather than use the 16nm.

[Trixanity]

Renée James is a woman.

[PlayStation 2]

1 hour ago, RorzNZ said:

125W though. That will be a bit naughty to cool. I think maybe everyone will just wait a year for the 7nm version rather than use the 16nm.

In the grand scheme of things, would 125 watts be really high?

[vanished]

11 minutes ago, Dan Castellaneta said:

In the grand scheme of things, would 125 watts be really high?

Not really.  Intel's 8180 is quoted at 205 W and AMD's EPYC 7601 is 180 so obviously up to at least that much is considered acceptable for servers.

[PlayStation 2]

2 minutes ago, Ryan_Vickers said:

Not really.  Intel's 8180 is quoted at 205 W and AMD's EPYC 7601 is 180 so obviously up to at least that much is considered acceptable for servers.

That's what I was kinda figuring, I know there's other server parts out there which are rated for significantly higher, so if this can deliver reasonable performance it'd be something easier to cool.

[Drak3]

1 hour ago, RorzNZ said:

125W though. That will be a bit naughty to cool. I think maybe everyone will just wait a year for the 7nm version rather than use the 16nm.

My 5930K is 140W.

 

14 minutes ago, Dan Castellaneta said:

In the grand scheme of things, would 125 watts be really high?

Depends on the performance.

[vanished]

Just now, Dan Castellaneta said:

That's what I was kinda figuring, I know there's other server parts out there which are rated for significantly higher, so if this can deliver reasonable performance it'd be something easier to cool.

Yeah don't forget that running 2 or even 4 CPUs in a rack (specifically, the ones I mentioned above) is also a thing, so if that much heat can be handled I'm sure they can deal with 125 W, or eve 4x that without too much trouble.

[PlayStation 2]

Just now, Ryan_Vickers said:

Yeah don't forget that running 2 or even 4 CPUs in a rack (specifically, the ones I mentioned above) is also a thing, so if that much heat can be handled I'm sure they can deal with 125 W, or eve 4x that without too much trouble.

Yeah, even if I usually don't pay attention to the server scene I do know they tend to be ran in racks. Hopefully these turn out to be real winners.

[vanished]

2 minutes ago, Dan Castellaneta said:

Yeah, even if I usually don't pay attention to the server scene I do know they tend to be ran in racks. Hopefully these turn out to be real winners.

ah in my haste I worded that poorly... rack is the whole thing, I mean in like, a couple of rack units (one case)

[PlayStation 2]

2 minutes ago, Ryan_Vickers said:

ah in my haste I worded that poorly... rack is the whole thing, I mean in like, a couple of rack units (one case)

ah don't worry

everyone has long days

[Guest]

1 hour ago, Dan Castellaneta said:

In the grand scheme of things, would 125 watts be really high?

 

1 hour ago, Drak3 said:

My 5930K is 140W.

 

Depends on the performance.

Like for an ARM cpu it seems pretty high, although it is 16nm. Yes, we will have to wait for the numbers to really compare.

[Drak3]

Just now, RorzNZ said:

 

Like for an ARM cpu it seems pretty high, although it is 16nm. Yes, we will have to wait for the numbers to really compare.

Not really. 32 cores at 3.3GHz.