 |
|
|
| ASProCore Overview |
At the heart of Aspex’s Linedancer processors is the ASProCore – Aspex’s unique parallel processing fabric.
Each Linedancer chip contains thousands of ASProCore Processing Elements, all working in parallel to achieve ultra-high performance for complex, data parallel operations such as image processing.
ASProCore - Fine Grain Parallel Processing Fabric
With 7 generations of silicon development over the last 15 years, we’ve learnt that
“small is beautiful”.
We consistently get more performance per square millimetre of silicon by keeping our processors small, and having lots more of them.
Each ASProCore Processing Element contains several elements:
-
A bit-serial ALU;
-
Content-addressable memory (CAM);
-
Extended (standard) memory for use as scratchpad memory;
-
A dual-ported Primary Data Store memory, allowing data to be streamed into and out of the ASProCore array at the same time that it is processing data;
-
An ultra-high speed communications network, allowing each PE to communicate with any other PE in the string; | ASProCore - Scalable “Smart Memory” implementation
Just like memory, ASProCore is built in a standard CMOS logic technology, using a “cookie cutter” approach. This allows us to quickly instantiate any number of processors on a single chip, giving us the ultimate flexibility in making cost/performance tradeoffs at a silicon level.
With the latest silicon technology, we can squeeze hundreds of Processing Elements into each square millimeter of silicon, allowing us to put several thousand processors on a single chip.
Twin benefits of Moore’s Law – area & performance
The ASProCore array benefits directly from Moore’s law, which states that the number of transistors on a silicon chip can double approximately every 18 months. Moore’s law has held true for the last 30 years, especially when applied to memory technology.
As two-thirds of ASProCore is memory, this has allowed us to consistently build more processors on a single chip, making ASProCore a futureproof architecture which inherently benefits from future silicon shrinks.
A side effect of the shrinking transistors has been an increase in performance – our ASProCore now clocks at more than 400 MHz, giving us more than 800 billion operations per second. And unlike fixed-function hardware accelerators, ASProCore can generally be used at the highest possible clock rate to maximize the performance per square millimeter.
Shorter wiring = higher performance
Until the mid-1990s, silicon performance was limited by the speed of the transistors on the silicon. As silicon process geometry shrinks below 130nm, delays in the on-chip wiring start to dominate the performance instead. Managing and balancing the timing across a chip, especially with complex 2D arrays of processors, can be a major headache.
With ASProCore, the 1-dimensional string topology means that each Processing Element communicates with its nearest neighbours, keeping the wiring short, and the performance high.
| |
|
|
