David Kanter: AMD's ARM core will be 10% faster than their x86 one, ditch Bulldozer

[Vesku]

If you parse the Dell comments from this month it seems like there is interest in 64 bit ARM servers.

“We’re continuing to play close attention to it. We’re moderating end customer interest. We’ll continue to stay engaged with the ARM ecosystem ... in time for end customer adoption,” Norrod said. http://www.pcworld.com/article/2461180/processor-delays-hurt-arm-server-adoption-dell-exec-says.html

Meanwhile AppliedMicro is saying it's ready to supply: http://www.apm.com/news/apm-announces-readiness-64bit-arm-based-soc-high-performance-computing/

Hints at pricing hitch between Dell and AppliedMicro. Will be interesting to see what AMD is targeting for K12 ARM, small-medium or medium-large cores.

 

[CHADBOGA]

Did you every consider that these companies just don't have what it takes to make a competitive ARM server processor? They have to spend a lot more money than AMD does to compete. AMD already has the IP developed with its x86 processors.

Yes, the market won't be huge but it's still a place to make money. Samsung can't even make its own processor for its own phones not alone produce a server chip. Qualcomm has no server experience, nVidia has server experience but not in CPU...

And yet so many people were projecting that Intel was going to be swept away by the ARM hoardes.

Looks like reality is slowly sinking in.

 

[teejee]

And yet so many people were projecting that Intel was going to be swept away by the ARM hoardes.

Looks like reality is slowly sinking in.

Who has said that ARM will sweep away Intel in server space in short time? Just a strawman from you.
The competition starts late this year and ARM will step by step gain marketshare during the coming ten years.

 

[underclockers]

What a mess in this thread. Full of off-topic nonsense isn't it.

x86 need something to simplify. ARM need something to complexity.

Thet's why ARM is smaller and much efficient when comes to mobile. In heavy workload those differences still exist although not that huge anymore like mobile.

 

[MiddleOfTheRoad]

Our company has 0 opteron processors.

Yeah, and the DOE decided to build Titan with 18,688 of them.
http://en.wikipedia.org/wiki/Titan_%28supercomputer%29

AMD is doing just fine with the Opteron line considering their limited resources.

Titan was the most powerful supercomputer in the world when it debuted
according to Top500. Tianhe-2 has now replaced it as #1 in the world.....
But seriously, the world's #2 most powerful supercomputer is AMD powered.
The best Intel can claim is #6. Opterons are excellent performers for their
respective price. I'd be curious to see if ARM can break into supercomputers
like Sparc or PowerPC has.

 

[ShintaiDK]

Yeah, and the DOE decided to build Titan with 18,688 of them.
http://en.wikipedia.org/wiki/Titan_(supercomputer)

AMD is doing just fine with the Opteron line considering their limited resources.

Titan was the most powerful supercomputer in the world when it debuted
according to Top500. Tianhe-2 has now replaced it as #1 in the world.....
But seriously, the world's #2 most powerful supercomputer is AMD powered.
The best Intel can claim is #6. Opterons are excellent performers for their
respective price.

Cray was locked to AMD. How many Opterons do Cray sell today? 0?

 

[mrmt]

Yeah, and the DOE decided to build Titan with 18,688 of them.
http://en.wikipedia.org/wiki/Titan_(supercomputer)

AMD is doing just fine with the Opteron line considering their limited resources.

Titan was the most powerful supercomputer in the world when it debuted
according to Top500. Tianhe-2 has now replaced it as #1 in the world.....
But seriously, the world's #2 most powerful supercomputer is AMD powered.
The best Intel can claim is #6. Opterons are excellent performers for their
respective price. I'd be curious to see if ARM can break into supercomputers
like Sparc or PowerPC has.

Fine under what parameters? Because the only important parameters, sales and profits, are an outright disaster for them.

 

[Cerb]

I am sure thats just coincidence... one way or the other.

1. If you need the most RAM per U per Watt per dollar, AMD still has good CPUs, and RAM will idle high enough in power to make up for the CPU.
2. If you're doing rare DC-like clock speed and ALU bound (non-FP) work in clusters, AMD still has good CPUs.

Neither of those are remotely common, and TCO is high enough for any big fast server that CPU pricing barely matters, even using all Free software (if you need more sockets for more RAM, FI, increasing the sockets and paying for the DIMMs dwarfs a few thousand bucks for a CPU). The companies buying those AMD servers I'm sure are getting their money's worth out of them, but they represent tiny niches.

Shared hosting and bulk data processing might be able to use microservers, but with ARM implementers not being able to move fast enough to beat Intel to the punch, it may be a lost cause. They didn't see the writing on the wall when there was a demand for GPU cards in racks, and they didn't see it when the Atom came out and was almost too popular for all kinds of low-end uses. They only saw it long after Atoms and Raspberry Pis were being used for server duty where ECC and RAS aren't important, and by the time they put two and two together ("we can make real server SOCs with many times that performance at a low cost per unit, and people might actually want that..."), Intel was well on their way to releasing server grade Atoms, and improving power efficiency by leaps and bounds.

If just due to Intel's monopoly being disliked, I think they still have some window of opportunity left, but ARMv8 not being ready sooner, nobody wanting to use PAE to be quicker to market (surprise!), and Intel wanting in on it, they've all, AMD included, got their work cut out for them.

 

[Cerb]

Cray was locked to AMD. How many Opterons do Cray sell today? 0?

To add: with the Memory Wall being what it is, supercomputers are heavily limited by not merely memory speeds, like all of our gear, but also networking speeds. Not merely using special fast networking systems, but also fine tuning them, is a really big deal, and will make as much, if not a greater, difference, than the CPU, these days (at least with fast modern CPUs that handle lots of FP throughput). Being able to get those aspects right are a large part of why Fujitsu is still developing and making supercomputers, and why Cray still exists. They could build a very fast Opteron-based system today, too, it just wouldn't make any sense to, when Intel CPUs are generally more capable, have AVX2 (you know they've got engineering samples, already), and are such a small part of the cost.

 

[jpiniero]

The Cray supercomputer IIRC is mainly using the Teslas for the computation; the Opterons were chosen because they can use more ram per socket than Intel.

 

[MiddleOfTheRoad]

Fine under what parameters? Because the only important parameters, sales and profits, are an outright disaster for them.

Based on the size of the company in comparison to Intel -- Overall, they are actually still overperforming. Intel is a company that is 5 times larger.

Opteron did peak above 25% marketshare about 6 years ago.... That's pretty damn impressive considering the shoestring budgets they designed it with. Opteron has been in a period of decline recently mostly due to a lack of significant new offerings.

 

[Fox5]

And yet so many people were projecting that Intel was going to be swept away by the ARM hoardes.

Looks like reality is slowly sinking in.

Right now, I'd say Nvidia has the best bet.

If they put their ARM cores in something with the power efficiency of Tegra K1 or Maxwell but scaled up to the high end, they'd be way ahead of all competitors in compute density/w.

There's no reason AMD couldn't put out a similar product, but right now CUDA is a more compelling platform than OpenCL.

 

[monstercameron]

Right now, I'd say Nvidia has the best bet.

If they put their ARM cores in something with the power efficiency of Tegra K1 or Maxwell but scaled up to the high end, they'd be way ahead of all competitors in compute density/w.

There's no reason AMD couldn't put out a similar product, but right now CUDA is a more compelling platform than OpenCL.

what is your evidence for this? or is it just anecdotal?

 

[MiddleOfTheRoad]

what is your evidence for this? or is it just anecdotal?

I can't speak for why he thinks so.

I would at least suspect that he is referring to the fact that Cuda has available libraries.

This is a good read about OpenCL vs Cuda:
http://scicomp.stackexchange.com/questions/10354/cuda-vs-opencl-as-of-late-2013

 

[NTMBK]

what is your evidence for this? or is it just anecdotal?

CUDA is nicer to write code in than OpenCL C kernels, and has much better debugging and profiling tools.

 

[cytg111]

Problary not so much about the isa's as it is about the software.. Gaining access to the arm software eco system would be a big deal, it being iOS Android or whatever and running natively without a recompile .. especially if you bring bigcore x86-like performance to the "game". Markets could be breaking up.

And now we have ARM tapping into the x86 ecosystem as well

http://linuxgizmos.com/emulator-brings-x86-linux-apps-to-arm-devices/

 

[Fox5]

what is your evidence for this? or is it just anecdotal?

Anecdotal, but you certainly here about a lot more design wins for nvidia.

As a developer though, nvidia has way more libraries, training materials, and documentation available. Doing OpenCL is like doing some undocumented open source project, doing CUDA is like using Matlab.

 

[Arachnotronic]

Anecdotal, but you certainly here about a lot more design wins for nvidia.

As a developer though, nvidia has way more libraries, training materials, and documentation available. Doing OpenCL is like doing some undocumented open source project, doing CUDA is like using Matlab.

I have some (admittedly limited) experience with both OpenCL and CUDA, and I'd say this is a really apt analogy.