• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Nvidia Launches World’s First ARM 64Bit Powered Tesla K20 HPC Solution – ARM’s Entry into HPC

Gregster

Gregster

Gregster

Caporegime
Joined
24 Sep 2008
Posts
38,284
Location
Essex innit!
Nvidia unveiled a relatively new design two days ago, a design which green claims is a world’s first and it probably is. This venture marks the entry of the ARM Cores into the HPC Market for the first time. Basically Nvidia has coupled two ARM processors both with 8 cores with 2 Tesla K20 GPUs for a server ‘block’ style HPC Solution. Tesla K20

Nvidia Pairs 64 Bit ARM Octacore with the Tesla K20 GPU for World’s First ARM Powered HPC Solution
This information comes from the folks over at Computerbase.de who received the information in a press deck fo the International Supercomputing Conference 2014. Here Nvidia presented the ARM HPC with GPU Acceleration ‘block’ which you can see below. The block is actually from Cirrascale in a shallow 1U rack with two sections each containing one ARM + GPU couple. After some time 3U systems are planned as well and the company Eurotech promises to bring Liquid Cooler variants as well. These solutions can be used for Public or Private Cloud, HPC, and enterprise applications. Along with your usual seismic signal and image processing, video analysis, trace analysis and web applications. The list is endless, and this combo can pack a deadly punch. The GPU Accelerator will handle all the grunt work while the ARM fires off the orders (only Intel’s Co-Processors can task themselves). The Micro X-Gene ARM 64 CPU will be clocked at 2.4 Ghz and will be coupled to a Tesla K20. The server style is allegedly micro-ATX. The motherboard will feature support for the X-Gene ARM CPU (duh) along your usual suspects such as the PCI E 3.0 slot with 8 lanes.
nvidia-arm-hpc

cbfa04db1b1a3bd96a37b456a51a0767.jpg


ARM CPUs are surprisingly power efficient compared to their power hungry x86 siblings. In the HPC sector where the Performance Per Watt Per Dollar makes a helluva difference ARM should be rampant, but that is not the case since the small scale tech has not been adapted for HPC use. Untill now that is. Riding Nvidia’s GPGPU momentum into the market, ARM could bring abut a silent revolution in the same, since it offers nearly double the performance per watt. Nvidia already offers GPU acceleration for the x86 and POWER platforms but now ARM has been included as well, which is a very admirable step.


Read more: http://wccftech.com/nvidia-worlds-first-arm-64-bit-tesla-k20-hpc-solution/#ixzz35dfAPKGS
Click to expand...

This could be why the Maxwell only has a 256 bit bus. I personally think it is because it is mid range but time will tell. Maxwell not far off but an ARM chip coupled with a K20 does make for interesting reading.
 
Gregster said:
Maybe the ARM chip will help with boosting the memory speeds.
Click to expand...

What on earth, the arm chip has NOTHING to do with maxwell, the arm is the cpu on the mobo, the gpu is a gpu connected to the mobo via a pci-e bus. Nothing more. An ARM chip can't speed up memory speeds, and couldn't do so if it were on die either.

Nvidia are buying octo core chips from... Samsung maybe, can't remember if Qualcomm are making octo's yet, and putting in a lightweight mobo/cpu combination designed for throughput and just letting the gpu's do the work.

In boxes in which the cpu wasn't loaded anyway and all work was passed off to the gpu, it's about enabling the gpu to do this with the minimum power taken up by the cpu possible, but still having enough performance to process the data to send it to the gpu.

As for being first, I believe there is already ARM based supercomputers, with both a Tegra 2 and Tegra 3 based one in France or somewhere, which at some stage they got rid of Nvidia chips and put Exynos Octo cores in for more power.
 
drunkenmaster said:
What on earth, the arm chip has NOTHING to do with maxwell, the arm is the cpu on the mobo, the gpu is a gpu connected to the mobo via a pci-e bus. Nothing more. An ARM chip can't speed up memory speeds, and couldn't do so if it were on die either.

Nvidia are buying octo core chips from... Samsung maybe, can't remember if Qualcomm are making octo's yet, and putting in a lightweight mobo/cpu combination designed for throughput and just letting the gpu's do the work.

In boxes in which the cpu wasn't loaded anyway and all work was passed off to the gpu, it's about enabling the gpu to do this with the minimum power taken up by the cpu possible, but still having enough performance to process the data to send it to the gpu.

As for being first, I believe there is already ARM based supercomputers, with both a Tegra 2 and Tegra 3 based one in France or somewhere, which at some stage they got rid of Nvidia chips and put Exynos Octo cores in for more power.
Click to expand...

I am just spitting in the wind and speculating. It is possible that a ARM chip could work between GPU and memory. I could be miles off like you say but these are just my guesses.
 
It's for processing and I/O purposes. You need something to handle the disk drives, networking and the dispatching of tasks to the GPU. These things are generally not limited by CPU performance, since once started the GPU is capable of directly accessing memory as well as creating its own tasks.
 
This isn't the integrated ARM core as foretold by videocardz et al, that's supposedly coming with Maxwell. This is just a server rack. :D
 
Gregster said:
I am just spitting in the wind and speculating. It is possible that a ARM chip could work between GPU and memory. I could be miles off like you say but these are just my guesses.
Click to expand...

Oh it's totally possible, if it were on the card then it could perform hardware compression/decompression on the fly which means you could get much more data down the 256bit bus.

Of course a much much simpler and cheaper solution would be to just use a bigger bus :P
 
ubersonic said:
Oh it's totally possible, if it were on the card then it could perform hardware compression/decompression on the fly which means you could get much more data down the 256bit bus.

Of course a much much simpler and cheaper solution would be to just use a bigger bus :P
Click to expand...

Fair point. I was just looking for a reason for the 256bit bus was all and come up with 5 :D
 
You must log in or register to reply here.
Back
Top Bottom