|
S4 |
4
Gflops (2 Nodes)
Four TM-66 Chips
60 Instr. per Clock
32-bit FPU's
VLIW Architecture
64 MB Static RAM
1.6 GB/sec DMA |
|
|
|
S4
Vector Processor |
The S4 vector processor is specially
designed to execute intense scientific algorithms efficiently and quickly. The S4
processor board is the basic scientific processing node in the SAM system. Each S4
combines two vector processor nodes on one board and each S4 node delivers GFLOPs of
actual sustained processing power. No other vector processor on the market
delivers a higher percentage of its peak theoretical GFLOPs on real applications.
The vector processor node architecture is well suited to scientific and DSP
applications, and specifically to FFT, vector, and matrix operations. User application
programs are quickly written as sequences of function calls for execution on the S4 vector
processor. Hundreds of optimized scientific and DSP function routines are available
in the DSP Math library.
|
|
S4
Architecture |
Each S4 node features a sophisticated very
long instruction word (VLIW) architecture which allows multiple hardware resources to work
in parallel and for a continuous flow of data to be processed. TM-66 DSP chips, in
conjunction with a very high bandwidth memory, supply the number crunching power of the
S4. Two 800 Mbytes/sec DMA ports keep the data flowing between the S4 nodes and SAM system
memory. This combination of processing power and memory bandwidth makes for a very
powerful vector processor board well suited for repetitive DSP processing on extremely
large data sets.
|
|
TM-66
DSP Chip |
The TM-66 chip was designed by Texas Memory
Systems to provide the number crunching power for the SAM system. Each TM-66 chip has 20
floating point execution units, six fast data I/O ports, and a separate instruction port.
Internally, twelve adders and eight multipliers are arranged into two parallel pipelines.
While this chip has many execution units, it has a simple architecture. Since all
data management and processing are deterministic, efficient programming algorithms are
easy to implement.
|
|

|