backtop


Print 14 comment(s) - last by gamilonman.. on Apr 5 at 10:15 AM

Barcelona-based design also uses NVIDIA low-power GPUs, similar to current integrated laptop designs

Apple Inc.'s (AAPLhot new A5X processor found in the third-generation iPad gave NVIDIA's (NVDATegra 3 chip a run for its money, blowing it away in graphics tests.  However, in CPU tests, NVIDIA's higher clocked processor -- which uses licensed Cortex-A9 IP (intellectual property) cores from ARM Holding plc (LON:ARM) -- felled its fruity foe.

I. Mont-Blanc is an Energy-Efficient Powerhouse

Thus it's not surprising that The Barcelona Supercomputer Center (BSC), an internationally renowned supercomputing center, turned to the Tegra 3 when they were looking for a lightweight processor to serve the role of communications mediator in a new energy-efficient GPU computing design.

The new Spanish supercomputer will be dubbed "Mont-Blanc".

In the new design, the bulk of computation will be carried out using either OpenGL or NVIDIA's proprietary CUDA API, with kernels (pieces of parallel work) running on unannounced NVIDIA low-power GPUs (the center says are similar to the GeForce GT 520MX).  The GPUs will likely be based on the new Kepler architecture (GeForce 600 series), which proved extremely energy-efficient in AnandTech's benchmarking.

But something will need to ferry data to the GPU, while performing I/O between GPU-enabled computing units and offloading results to storage devices.  That something will be the Tegra 3.

Tegra 3
Tegra 3 chip package and die image [Image Source: AnandTech]

While the typical Xeons gobbles 50 to 100 watts, the Tegra 3s in Mont-Blanc are expected consume around 4 watts when loaded.  The finished design is expected to pack anywhere from 2,000 to 4,000 processors inside, along with corresponding GPUs.

The BSC project aims to become the first ARM-powered supercomputer to make the Top500 list of supercomputers, which is currently topped by Japan's 10.5 petaflop "K computer."  However, the BSC aims to be at the very top of another newer list -- the Green 500, a list measuring the world's most power efficient supercomputers.

That list is currently topped by International Business Machines, Inc.'s (IBMBlue Gene computer at the Thomas J. Watson Research Center, which can do 2 quadrillion calculations per second (2 gigaflops) per watt.  Mont-Blanc aims to earn 7 gigaflops per watt, making it more than three times as energy-efficient as any design in the world today.

Alex Ramirez, product manager at BSC who's working on the project, in an interview with Wired states, "Instead of using very few — but very big performance — processors… we’re going to be using a lot of very low-power — but middle performance — processors."

II. Tegra ARMv8-powered Successor Design Aims to Top Top500

But Mr. Ramirez's team is far from finished when they complete the lightweight design sometime in the next couple years.  They're already planning its successor, a next generation design, which will be build on NVIDIA's 64-bit ARMv8 instruction set Stark series (or possibly a later variant).

That 64-bit design is aimed at becoming the world's most powerful supercomputer, with a planned processing power of 200 petaflops (twenty times the K Computer's capacity).  That design will be completed in 2017 (approximately), according to plans shared with Wired.

Mont Blanc
A new 64-bit design to launch in 2017 will be even beastlier than Mont-Blanc. 
[Image Source: Flickr/pirinenc]

ARM Holdings itself has invested casually in pushing out power-efficient ARM servers to IT users.  However, its focus remains on its core markets -- smartphones, tablets, and embedded devices -- as well as its push into the personal computer space.  Stated new ARM president Simon Segars last month, "Supercomputers, for ARM, is not a high volume market,” he said. “It’s not something we spend a lot of time talking about. Ours is a business that is royalty and unit driven, so we’re interested in high-volume markers."

For now, he says, ARM supercomputers are an "interesting" diversion, but not a business focus.

Source: Wired



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: Tegra3 or Tegra4?
By nafhan on 4/4/2012 1:46:04 PM , Rating: 2
Like Flunk said: they'll be using Tegra 3 for the CPU's.

The use of low power GPU's is interesting. It makes me think they may be putting a high priority on available GPU memory (i.e. high ratio of memory to compute resources). High end GPU's have orders of magnitude more compute power, but only 2 or 3 times as much memory. Interesting stuff...


RE: Tegra3 or Tegra4?
By JasonMick (blog) on 4/4/2012 3:02:29 PM , Rating: 2
quote:
The use of low power GPU's is interesting. It makes me think they may be putting a high priority on available GPU memory (i.e. high ratio of memory to compute resources). High end GPU's have orders of magnitude more compute power, but only 2 or 3 times as much memory. Interesting stuff...

Having worked with CUDA code myself, I would guess they're going to go for some sort of custom Kepler GPU solution, with a lower clock, but similar cache, but potentially shared global memory.

One of the biggest problems with CUDA speed-wise is that it overwhelms on the computational front getting work done very fast, for parallelizable problems (so clock speed could be lowered) and has typically more global memory than you need. The limitation is in SMP count (which constrains how much work you can do @ once) and the speed cost of ferrying results off the GPU via the PCI bus.

If Mont-Blanc or its successor use a unified global memory with the CPU, latency might increase slightly on the GPU side (as discrete card GDDR5 is relatively low-latency), HOWEVER, you would eliminate a costly transfer step.

In other words, most current CUDA schemes operate like:

Results (GPU)
|
PCIe
|
V
Results (CPU)
|
\--> SATA storage,
|
\
|\ ----> Memory
|
\---> Other processor (via optical interconnects, etc.)

By merging the GPU and CPU memory, you would eliminate the PCI transfers, which can be a major program cost.


RE: Tegra3 or Tegra4?
By nafhan on 4/4/2012 3:41:25 PM , Rating: 2
I was making my assumption based on this: "the center says are similar to the GeForce GT 520MX", a card which is probably pretty low on SP resources.


RE: Tegra3 or Tegra4?
By JasonMick (blog) on 4/4/2012 4:30:10 PM , Rating: 2
quote:
I was making my assumption based on this: "the center says are similar to the GeForce GT 520MX", a card which is probably pretty low on SP resources.
I think they meant that in the sense of a non-card option with unified memory (hence the memory might actually be a bit LESS).

On the other hand Kepler iGPUs, while not fully all out there yet are expected to pack much more SMPs... plus this may possibly be a semi-custom design.


RE: Tegra3 or Tegra4?
By nafhan on 4/5/2012 9:28:53 AM , Rating: 2
Either way... should be interesting :)


"We basically took a look at this situation and said, this is bullshit." -- Newegg Chief Legal Officer Lee Cheng's take on patent troll Soverain














botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki