Xilinx Logo  
HomeProductsSupport
                                                        Devices
 
Home : Products : Devices : Virtex Series : System Timing
Virtex Series FPGAs - System Timing

  Virtex & Virtex-E Overview
  Virtex & Virtex-E Product Tables
  Virtex-EM Overview
  System Interface
  System IO
  System Timing
  System Design Tools
  Demo Boards

Wireless Connection
Memory Corner
Speeds Files
IBIS Models
Press Releases

 

DLL Application Notes and Reference Designs
Using the Virtex Delay-Locked Loop Using the Virtex Delay-Locked Loop

Eight High Performance DLLs - Drop-in Bandwidth Optimization

Supporting the highest bandwidth data rates between devices requires advanced clock management technology such as DLL. The DLL circuitry allows for very precise synchronization of external and internal clocks. Xilinx was the first to deliver DLLs in programmable logic by offering four 200 MHz DLLs in every Virtex device. The Virtex-E family takes this technology to the next level with devices containing eight DLLs capable of over 311 MHz. As a fully digital implementation, the Virtex and Virtex-E DLLs do not have the typical problems of an analog phase locked loop (PLL) including board isolation and decoupling of power and ground. Virtex-E DLLs provide precise clock edges through phase shifting, frequency multiplication, and frequency division. The precise duty cycle generation is critical for high performance applications (like Double Data Rate, or DDR) in which a slight shift in duty cycle can dramatically decrease overall system performance.

Bandwidth Critical Specifications of the Virtex-E DLL
Parameter Value
Maximum Output Frequency 320 MHz*
Maximum Output Jitter 100 ps
Output Frequency Duty Cycle 50%+/- 100 Ps

* Based on Virtex-E 7 speed grade product

A key technique for increasing the bandwidth of a particular data port is to have signals change on both edges of a clock, commonly referred to as the Double Data Rate technique. Memory suppliers have already started to support this type of high performance technique to increase the memory bandwidth of their devices. At high frequencies, signal integrity limits the clock performance, which limits the bandwidth of the data. Bandwidth for the port is immediately doubled if the architecture can change data at each edge of a system clock. It is critical that a clock duty cycle is very precise to 50 percent for this technique. Since Virtex-E DLLs can generate clocks with a duty cycle guaranteed to be within 100 picoseconds (Ps) of 50 percent; system designers can achieve the maximum memory bandwidth in the DDR application. The following diagram demonstrates how Virtex-E DLLs help achieve maximum bandwidth in a 266 MHz DDR application.

For more details regarding DLL, please see the DLL Tech Topic.

System Timing Solution (Virtex)

Highlights

The Virtex family solves clock skew issues by allowing internal and external clock synchronization. With four Delay Locked Loops, your system's clock skew will be removed. You can get clock-to-output delays of less than three nanoseconds and very substantial increases in system performance.
Fully Digital Delay Locked Loops (DLL)
  • Four independent DLL circuits for internal and external clock synchronization
  • 200+ MHz chip-to-chip communication
  • Less than 3 ns clock-to-output time across all devices
  • Clock doubling and clock division
    0o, 90o, 180o, and 270o phase clocks
Zero Delay Clock Management
Zero Delay Clock Management
Multiple DLLs facilitate precise generation of zero-delay clocks both inside and outside the FPGA for highest chip-to-chip speeds.

Overview of Delay-Locked Loops

As FPGAs grow in size, quality on-chip clock distribution becomes increasingly important. Clock skew and clock delay impact device performance and the task of managing clock skew and clock delay with conventional clock trees becomes more difficult in large devices. The Virtex series of devices resolve this potential problem by providing four fully digital dedicated on-chip Delay-Locked Loop (DLL) circuits which provide zero propagation delay and zero clock skew between output clock signals distributed throughout the device.

Each DLL can drive up to two global clock routing networks within the device. The global clock distribution network minimizes clock skews due to loading differences. By monitoring a sample of the DLL output clock, the DLL can compensate for the delay on the routing network, effectively eliminating the delay from the external input port to the individual clock loads within the device.

In addition to providing zero clock skew and zero delay with respect to a user source clock, the DLL can provide multiple phases of the source clock. The DLL can also act as a clock doubler or it can divide the user source clock by up to 16.

Clock multiplication gives you a number of design alternatives. For instance, a 50 MHz source clock doubled by the DLL can drive an FPGA design operating at 100 MHz. This technique can simplify board design because the clock path on the board no longer distributes such a high-speed signal. A multiplied clock also provides designers the option of time-domain-multiplexing, using one circuit twice per clock cycle, consuming less area than two copies of the same circuit. To increase the effective clock multiplication factor to four, you can connect two DLLs in series.

The DLL can also act as a clock mirror. By driving the DLL output off-chip and then back in again, you can use the DLL to deskew a board level clock between multiple devices.

In order to guarantee the system clock establishes prior to the device waking up,the DLL can delay the completion of the device configuration process until after the DLL achieves lock.

By taking advantage of the DLL to remove on-chip clock delay, you can greatly simplify and improve system level design involving high-fanout, high-performance clocks.

Customer Comments:

"Virtex FPGAs have allowed us to implement our next generation digital TV broadcast systems in record time," said John Simmons, project manager, of NDS, a world leader in digital broadcasting solutions. "A key time saver was the availability of multiple DLLs that allowed us to synchronize a 74 MHz clock to more than 30 devices including multiple FPGAs, SDRAMs, and other components. Designing a no-skew clock system from scratch would take months. Xilinx delivered a ready-made solution to us with Virtex FPGAs."

"We did an exhaustive search and found that only the Xilinx Virtex FPGAs could provide us with the performance and density necessary to add Gigabit capability to the Nebula switch family. John Peters, vice president of development at Performance Technologies.We are very impressed with the system level capabilities of the Virtex FPGAs, particularly the digital delay locked loops and support for multiple I/O standards."

 

Back to Top  
  Trademarks and Patents
Legal Information

Privacy Policy
| Home | Products | Support |