NVIDIA ready for ARM supercomputing

NVIDIA ready for ARM supercomputing

Summary: CPU co-processing extends ARM capabilities in the datacenter

TOPICS: ARM, Data Centers

With the expected massive increase in the use of ARM in the datacenter and for general purpose computing, Nvidia is getting ready to take advantage of ARM’s most glaring weakness, abysmal floating point performance compared to the x86 architecture everyone is already familiar with.

If there is one thing that Nvidia GPUs do well, it’s math.  Many of the world’s top supercomputers already combine large numbers of the Nvidia Tesla co-processor GPU to aid them in achieving their top tier performance. And while Nvidia is already working on its own 640bit ARM designs, they announced this week that the software piece of the GPU puzzle, their CUDA programming environment was now available with top-level ARM support in version 5.5 as well as with support for a broad selection of Linux flavors. This allows developers to use the CUDA tools (compiler, debugger, performance analysis, etc.), to build new or port existing applications to the ARM environment that can make use of the GPU-accelerated math capabilities that the Nvidia GPUs provide.

From a datacenter perspective, GPU equipped ARM servers become a much more viable alternative to traditional servers if they are able to deliver equivalent or near-equivalent processing of IT workloads at significantly reduced energy budgets. Widespread adoption would also allow a more modular approach to IT workloads. For example, if you are doing low-level tasks that could be handled by ARM-based servers without GOPU support and such devices are cheaper in terms of both CAPEX and OPEX than even low-power x86 alternatives, then ARM is a natural fit.

But if your choice of x86-based systems was based on better math performance, ARM servers that allowed a modular upgrade path by adding GPU co-processing to offer equivalent capabilities might drive your basic purchase model more to the ARM side. If Nvidia and/or other ARM proponents are able to deliver a building block style server environment, where users are able to pick and choose components that can almost be dynamically added to the server farms to change the processing capabilities while retaining the same basic architecture, then the traditional x86 advantages in the datacenter may be lost.

It hasn’t happened yet, but the potential arrival of 64-bit ARM coupled with high-performance GPUs looks to change the landscape picture of the datacenter significantly.




Topics: ARM, Data Centers

Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.


Log in or register to join the discussion
  • Floating point math and precision

    Amd's GPU's are much better at single precision than double precision. While they achieve > 5 teraflop single precision they only get a 5th of that in double precision. In comparison the difference with Haswell processor is just 2x. A Haswell can achieve close to 500 GFlops (half of the Tesla) of double precision performance at much less cost and much less power. Those GPU coprocessors are very very expensive and consume huge amounts of power.

    And this is just FMA and AVX2.

    And it is simpler code.
  • Windows RT?

    I know that Microsoft has continued to say that they are commented to Windows RT that runs on ARMs. It would also make sense that as a large IT company they have the inside knowledge on where companies like AMD are putting their R&D. Microsoft also builds Windows Server and much of the kernel from Server ends up in Windows client.

    Is this why they are sticking with RT? They know it will have a more profound place in technology?
    Rann Xeroxx
    • Re: Is this why they are sticking with RT?

      You ascribe a degree of competence to Microsoft for something which can be more easily explained as based on stubbornness and pride.
  • ARM must more aggresive

    Intel seriously come into mobile market.
    I hope ARM also sriously aggresive into desktop and server market.
    If they can make cold server without many heat it will be great.
    ARM can also consider much bigger processor size, so we can reduce the processor number. whic mean big efficiency. ARM can do this because they did not produce much heat.
    Example for Server they can produce 20x20cm processor.
    Utomo Prawiro