Home Products Technology Description
Technology Description
Introducing Xilinx "Spartan" FPGAs

Todays general purpose processing units are designed to provide a processor that is optimized for general purpose applications. Which means that you have to optimize your code for the processor as it is physical predefined silicon, which can not be modified to fit to your application. The Field Programmable Gate Array, in short "FPGA", is different. This type of chip allows you to modify the silicon via software configuration to be the perfect Application Specific Circuit or short "ASIC" for your personal application, while remaining reconfigurable - in case that you frequently change the code. SciEngines oftentimes uses Xilinx Spartan series FPGAs (but also others) due to a highly optimal price-performance ratio. Using 16 to 128 of such FPGAs per computer, SciEngines achieves cluster or even super-computer performance at a fraction of the cost.

 

Faster per clock.

Usually, multiple parallel processing elements are synthesized into an FPGA. Thus, each Field Programmable Gate Array typically provides more computational performance than a higher clocked general purpose CPU or GPUs. Furthermore, SciEngines' unique architecture allows to scale this computational performance to a massively parallel level. Allowing automatic algorithm distribution and concurrent applications of multiple users on different FPGAs. Programmable clock generators allow high level clock optimization to maximize computational performance. All this combined results in being significantly faster per clock.

 


RIVYERA Architecture

 

Benefits of two worlds

The Hex-Core Intel "Gulftown" or Quad-Core Intel “Nehalam” processor has been combined with a high performance FPGA Super Computer. Joining the benefits of Intels single-die, 64-bit architecture makes 8MB of fully shared L3 cache readily available to each of the processor cores. This fully scalable FPGA computing setup allows to accelerate computationally expesive applications dramatically. The result is fast access to cache data via a SciEngines unique Interface or open standard interface structures like PCI Express. Combine that with the other technological advances and you get a supercomputer which is up to four orders of magnitude faster than current general purpose architecture.

 

More I/O bandwidth.

The new RIVYERA Family delivers up to twice the overall I/O bandwidth of the previous generation. Users can choose from one up to seven 1-lane (x1), 4-lane (x4) or even 16-lane (x16) wide PCI Express interface cards as well as from SciEngines unique ultra high bandwidth memory coupling. Because we provide mainboards with fully independant PCIe-slots, the bandwidth isn’t shared. So you get all the bandwidth you need by use of the latest I/O cards, e.g. 10Gb Ethernet cards, multiport 4Gb Fibre Channel cards, or InfiniBand interface cards for connecting your FPGA super computer.


RIVYERA Architecture

 

Increased performance per watt.

SciEngines products provide significantly better performance per watt than regular hardware. No matter if it is CPUs or GPUs, SciEngines hardware can provide the same performance at only a fration of the power consumption - and therefore also at a fraction of the heat. A peak power consumption of  only 650 Watts allows you to save enough on the energy-bill to easily pay for the whole investment. Compared to an equally powerful standard PC cluster, you can save up to 98 percent energy - depending on the application. How do you get more performance with less energy? Primarily, the FPGA is built on an industry-leading technology process allowing your algorithm to utilize almost all of the silicon area for your application. Therefore no idle units have to be powered and in a fully pipelined and optimized design it delivers a result every clock cycle. Thus RIVYERA delivers an outstanding energy efficiency and performance. A smart system design, combined with modern power management, further reduces energy consumption and results in a  literally "cool" supercomputer.


RIVYERA Architecture