XP-100

The DSP solution for very high bandwidth real-time environments

PCIe bus-based DSP card processes 1 Million point FFT in under a millisecond!

Texas Memory Systems delivers high-performance DSP hardware and software.  In classical computer marketing, all computer manufacturers touted the speed of their hardware.  Now as hardware gets incredibly fast the marketing emphasis is put on achieving this speed in real world applications.  This usable performance is directly related to the hardware-software interface efficiency.

Across four generations of TMS DSP products that have raised the performance bar by a cumulative factor of 600, TMS has maintained a uniform architecture for the user.  This compatibility has allowed the user to reuse legacy code on each successive generation of hardware.  Each next generation of TMS DSP hardware is designed with the current generation software in mind.  TMS has always delivered high-performance hardware that has efficient, usable, and compatible software.

The TMS XP-100 PCIe Accelerator card is based on the TM-100 DSP chip which is about 12x faster than the previous generation TM-44 chip.  The TM-100 is rated at 100-GFLOPS and offers bandwidth of 50-GB/s.  The chip operates at 333 MHz with a dual core design.  Each core is programmed independently with separate data memories.  The XP-100 is 6x faster than the XP-30/35 while maintaining compatibility with existing applications. 

At TMS, we provide the total solution.  While our DSP hardware is ranked at the top of the high-performance list, our included DSP software has been well received because it is powerful and easy to use.  The XP-100 comes with a mature DSP math library and many valuable software development tools to make programming the XP-100 faster and easier.

The TMS DSP software suite has been designed to operate as a stand-alone application development environment, or to integrate into the XMIDAS application environment (via VPX).  The stand-alone environment offers features that are not supported by XMIDAS.  One such feature is dynamic configurability, allowing application changes without requiring a full restart.  The stand-alone environment has been optimized for latency and throughput, allowing the final operator to scrutinize signals-of-interest or run full speed analysis of vast amounts of data.

The XP-100 PCIe card integrates with existing computer systems, working as a back-end DSP number cruncher.  It offers these features:

  • 100 GFLOPS
  • 50 GB/s bandwidth
  • local & global RAM
  • supports PCIe bus (x8, 1.5 GB/s)
  • 500+-routine math library

The XP-100 is a single board DSP accelerator card that plugs into a PCIe (x16) slot.  DSP number crunching power is provided by the TM-100 ASIC chip.  The TM­100 has two cores, each with its own local RAM.  Its Harvard architecture and VLIW makes the TM-100 a very efficient processor, even with 320 floating-point units to control.  Also, special hardware units are included for FFT twiddle-generation, and CORDIC Elementary Function Units are included for quick transcendental math.

TM-100 program memory is internal to the TM-100 ASIC, while data is stored in external, tightly coupled local RAM.   With a larger local RAM than is usually available on other DSP chips, data can be processed in larger blocks with more complex algorithms.  Complementing local RAM is global RAM, which is connected to the XP­controller.  Global RAM is a staging area for data streams passing between the PCIe bus and the local RAM. 

Bandwidth-Bandwidth-Bandwidth
The XP-100 DSP card has extremely fast buses to prevent data starvation at the TM-100 chip.  The TM-100 chip has a front-side I/O bus that connects to the XP controller.  In addition, the TM­100 chip has a direct and fast connection to its own local RAM.  Each core's local RAM can be loaded quickly (5-GB/s) from global RAM while the other core is still running.

Performance

Interfaces

Mechanical