Nvidia and AMD unveil new supercomputer GPUs

Nvidia and AMD unveil new supercomputer GPUs

Summary: Nvidia's Tesla K20 and Tesla K20X, and AMD's FirePro SM10000 GPUs can each deliver more than a teraflop of double-precision performance.


Chipmakers Nvidia and AMD have debuted new graphics processing units (GPUs) aimed at the supercomputer market.

Nvidia's GPUs are the Tesla K20 and Tesla K20X. The Tesla K20 features 2,496 stream processors and 5GB memory, all of which provides 3.52 teraflops of single-precision and 1.17 teraflops of double-precision peak performance. The Tesla K20X has 2,688 stream processors and 6GB of GDDR5 memory offers 3.95 teraflops single-precision and 1.31 teraflops double-precision peak floating point performance.

At the heart of these cards is Nvidia's new GK110 silicon packed with 7.1 billion transistor; the biggest GPU in terms of transistor numbers. Based on the Kepler architecture, the GK110 features 64 double-precision CUDA cores, giving it eight-times the double-precision performance of a earlier GK104 SMX.

The Tesla K20X GPU delivers three times greater energy efficiency than previous-generation GPUs, and offer better efficiencies compared to CPUs.

Because of the increased double-precision floating point performance, Nvidia says that the Tesla K20-series GPUs are ideally suitable for tasks such as seismic processing, financial computing, CFD, CAE, computational chemistry and physics, data analytics, satellite imaging, and weather modeling.

The Tesla K20 and Tesla K20X GPUs are also shifting fast, with Nvidia reporting that more than 30 petaflops of GPUs have already been delivered to various supercomputer clients during the last month. That 30 petaflops of power is equivalent to the computational power of last year's 10 fastest supercomputers combined.

Tesla Nvidia K20-series GPUs are shipping today and available for order from leading server manufacturers, including Appro, Asustek Computer, Cray, Eurotech, Fujitsu, HP, IBM, Quanta Computer, SGI, Supermicro, T-Platforms and Tyan, as well as from Nvidia reseller partners.

Also out is AMD's FirePro SM10000 dual-GPU workstation graphic processor, designed to be used in virtualized environments and targeted at high-performance computers and servers.

AMD reports that the FirePro SM10000 delivered 1.48 teraflops of peak double-precision performance and features 6GB of GDDR5 memory. The FirePro SM10000 delivers up to 1.3-times the single and up to 7.8-times the double peak precision floating point performance of Nvidia's Tesla K10 GPU.

The FirePro SM10000 can be used on cloud servers to accelerate graphics on the server side and deliver full high-definition virtual desktops to client devices.

Image source: Nvidia, AMD.

Topics: Hardware, Processors, Servers

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
  • Why "graphics"?

    Why are they called "graphics" processors? Are the Tesla's even used for that? I don't see any kind of I/O connector. If they're just incredibly fast massively paralleled floating-point math co-processor systems on a card, why not call them that?
    • Really

      You want me to call them "incredibly fast massively paralleled floating-point math co-processor systems on a card" every time I refer to them? ;)
      x I'm tc
    • It's really a GPGPU

      General purpose computing using graphics processor hardware - the K20 massively parallelizes 2500 stream cores (as are in an actual GPU you'd attach to a monitor) and you can inject vector and other parallel routines into it to either offload processing supporting high end visualization via a normal GPU or any other kind of calculations and manipulations you want to throw it
  • Power is the Key

    With the push forward in supercomputing power is the key. Not mentioned in this article is the power draw by each device. The AMD card draws much more power than the NVIDIA card and would likely not be used when trying to reach supercomputing levels.

    The AMD card produces 0.01576 teraflops single precision and 0.003947 teraflops double precision per watt of power consumption.

    The NVIDIA card produces 0.01681 teraflops single precision and 0.005574 teraflops double precision per watt of power consumption.

    The NVIDIA card is a runaway leader with 41% better performance double precision and 6.7% better performance single precision per watt compared to the AMD card.