NVIDIA Shows Off "Kayla" Running On Ubuntu

[phoronix]

Phoronix: NVIDIA Shows Off "Kayla" Running On Ubuntu

Announced at NVIDIA's GPU Technology Conference (GTC) 2013 event today was the "Kayla" ARM development board...

http://www.phoronix.com/vr.php?view=MTMzMTU

[zanny]

Prediction: this will flop.There are two use cases for GPGPU : massively parallel workloads in server farms where GPUs crush traditional generic core processors, and accelerating workloads for consumer apps.

The first segment is already covered by Quadro and its ilk, and nobody is going to stick Tegra chips in a server because they waste budget, resources, heat, and space on ARM cores that are much less efficient for the workloads they target than big beefy gpu cards. If they are doing CUDA compute in a server environment, it is on dedicated CUDA hardware, not some APU. Understandably, though, they pretty much dominate this segment with their current crop of "business" class gpus. FireGL isn't even close to the market penetration Nvidia has, and CUDA is hyper-optimized by them on purpose whereas openCL targets the second segment.

The second class is composed of developers who won't use a non-open GPGPU standard to write apps for in an environment even more hostile to Nvidia than the desktop where their only competitors are Intel and AMD. In mobile they don't have close to the market segment they do on desktop and nobody will platform lock themselves to Nvidia hardware, especially when every other player in the room ships openCL.

I think this announcement might be even worse for Nvidia than the mediocre performance figures on Tegra 3 and the lack of enthusiasm from manufacturers to adopt Tegra 4. It is obvious ARM in the next ~5 years will become the new mainstream compute platform for the consumer market, and GPGPU on these devices as they become more powerful is an obvious optimization path for hardware that needs to be exceedingly power efficient. Not using the industry standard and trying to stuff their (albeit, solid and well supported) 6 year old GPGPU implementation is sealing their fate if the rest of the mobile world starts adopting OpenCL in retaliation.

Going down a path to their own proprietary way to do things (as per their usual, to be honest) is going to alienate Nvidia from hardware segments they think this kind of move will get them a monopoly in.

[coder543]

Prediction: this will flop.There are two use cases for GPGPU : massively parallel workloads in server farms where GPUs crush traditional generic core processors, and accelerating workloads for consumer apps.

The first segment is already covered by Quadro and its ilk, and nobody is going to stick Tegra chips in a server because they waste budget, resources, heat, and space on ARM cores that are much less efficient for the workloads they target than big beefy gpu cards. If they are doing CUDA compute in a server environment, it is on dedicated CUDA hardware, not some APU. Understandably, though, they pretty much dominate this segment with their current crop of "business" class gpus. FireGL isn't even close to the market penetration Nvidia has, and CUDA is hyper-optimized by them on purpose whereas openCL targets the second segment.

The second class is composed of developers who won't use a non-open GPGPU standard to write apps for in an environment even more hostile to Nvidia than the desktop where their only competitors are Intel and AMD. In mobile they don't have close to the market segment they do on desktop and nobody will platform lock themselves to Nvidia hardware, especially when every other player in the room ships openCL.

I think this announcement might be even worse for Nvidia than the mediocre performance figures on Tegra 3 and the lack of enthusiasm from manufacturers to adopt Tegra 4. It is obvious ARM in the next ~5 years will become the new mainstream compute platform for the consumer market, and GPGPU on these devices as they become more powerful is an obvious optimization path for hardware that needs to be exceedingly power efficient. Not using the industry standard and trying to stuff their (albeit, solid and well supported) 6 year old GPGPU implementation is sealing their fate if the rest of the mobile world starts adopting OpenCL in retaliation.

Going down a path to their own proprietary way to do things (as per their usual, to be honest) is going to alienate Nvidia from hardware segments they think this kind of move will get them a monopoly in.

I don't have time to debate the finer details, but you're just wrong. GPGPU is extremely important for customer apps in the medium term, and you don't know nearly enough about their plans to talk about supercomputers. ARM with big GPGPU co- processors could be the next big advancement in supercomputer compute efficiency. You really don't know yet, no one does, but right now ARM is more efficient in performance per watt than x86, and therefore generates less heat. Space usage is all about the fabric.

Finally, this is a dev board, not a business class or consumer class product. You almost certainly cannot predict things about an unreleased and un-specced product.

[johnc]

Prediction: this will flop.There are two use cases for GPGPU : massively parallel workloads in server farms where GPUs crush traditional generic core processors, and accelerating workloads for consumer apps.

The first segment is already covered by Quadro and its ilk, and nobody is going to stick Tegra chips in a server because they waste budget, resources, heat, and space on ARM cores that are much less efficient for the workloads they target than big beefy gpu cards. If they are doing CUDA compute in a server environment, it is on dedicated CUDA hardware, not some APU. Understandably, though, they pretty much dominate this segment with their current crop of "business" class gpus. FireGL isn't even close to the market penetration Nvidia has, and CUDA is hyper-optimized by them on purpose whereas openCL targets the second segment.

The second class is composed of developers who won't use a non-open GPGPU standard to write apps for in an environment even more hostile to Nvidia than the desktop where their only competitors are Intel and AMD. In mobile they don't have close to the market segment they do on desktop and nobody will platform lock themselves to Nvidia hardware, especially when every other player in the room ships openCL.

I think this announcement might be even worse for Nvidia than the mediocre performance figures on Tegra 3 and the lack of enthusiasm from manufacturers to adopt Tegra 4. It is obvious ARM in the next ~5 years will become the new mainstream compute platform for the consumer market, and GPGPU on these devices as they become more powerful is an obvious optimization path for hardware that needs to be exceedingly power efficient. Not using the industry standard and trying to stuff their (albeit, solid and well supported) 6 year old GPGPU implementation is sealing their fate if the rest of the mobile world starts adopting OpenCL in retaliation.

Going down a path to their own proprietary way to do things (as per their usual, to be honest) is going to alienate Nvidia from hardware segments they think this kind of move will get them a monopoly in.

CUDA is the industry standard.

I don't know how much of this is going to flow over into mobile apps per se, but I think it'll be important in other, more integrated embedded systems like car tech, etc.


edit: here's the video they showed but it doesn't give much except that they appear to have an ARM binary kernel driver for the discrete Kepler GPU (I'm assuming). So that could also be an indication of things to come.

http://www.youtube.com/watch?v=qVsVtXUX5Bw

[Ibidem]

Prediction: this will flop.There are two use cases for GPGPU : massively parallel workloads in server farms where GPUs crush traditional generic core processors, and accelerating workloads for consumer apps.

The first segment is already covered by Quadro and its ilk, and nobody is going to stick Tegra chips in a server because they waste budget, resources, heat, and space on ARM cores that are much less efficient for the workloads they target than big beefy gpu cards. If they are doing CUDA compute in a server environment, it is on dedicated CUDA hardware, not some APU.

Really?
http://www.tomshardware.com/news/Nvi...ARM,13993.html
I've heard of at least one other supercomputer along the same lines.

[nightmarex]

but right now ARM is more efficient in performance per watt than x86, and therefore generates less heat.

I think you meant GPU compute is more efficient?
ARM in a GPU compute enviroment is fine, I wouldn't use the actual chips for number crunching though.

[zanny]

Really?
http://www.tomshardware.com/news/Nvi...ARM,13993.html
I've heard of at least one other supercomputer along the same lines.

They just want low power consumption. Efficiency wise, spending money on ARM cores that you aren't using in computation isn't giving you efficiency benefits. For a comparison of common gpgpu hardware, a Tegra 3 has 4 - 8 watt TDP and 4.8 - 7.2 gflops, whereas a GTX 580 (or comprable Quadro 6000) has 1581 GFLOPs on a 244 watt TDP.

Performance per watt of the gpu compute in single point floating operations, at least, heavily favors workstation graphics. And it should. The mobile part is an APU for low power devices and isn't meant to be a power efficiency workhorse dedicated server card. Real high end server hardware uses specialized interconnects, no generic CPU cores, and nothing like what most consumers here use in their desktops - they don't even use generic racks for compute class GPGPU tasks, they will link up dozens or hundreds of gpus into an IPC and run some task across all of them in massive parallel, no generic compute cores required with the exception of traffic routing in the network, maybe.

I just don't buy that you would get better performance / watt / heat out of mobile parts than hardware designed to do the workloads efficiently. Especially with the overhead of unused dangling ARM cores taking up die and current.

[TemplarGR]

CUDA is the industry standard.

....................... You serious?

[Licaon]

clearer pics http://www.pcgameshardware.de/GTC-Ev...M2-20-1061349/

[EpochDC]

CUDA is the industry standard.

OpenCL is the industry's standard, CUDA is a proprietary Nvidia technology.
Though it has a very large market penetration, it's not the standard...