Synchroscalar and Imagine

1
Synchroscalar and Imagine

• Chris Thomas
• Chris Chaney
• 10/5/05

2
Background - Synchroscalar

• ASICs can perform tasks with low power
• DSPs are much more flexible, lower performance
  and higher power/Op
• ASICs expensive to develop
• Synchroscalar seeks performance and power similar
  to ASIC with flexibility of DSP

3
Scynchroscalar - Architecture

• Array of processing elements arranged in a grid
• Elements grouped into columns
• Each column runs one stream of instructions
• High bandwidth within column (256 bit bus)
• Segmented bus static scheduling
• Allows for broadcast or localized communication
• Aimed at high parallelism with low clock rate
• Each column has its own clockrate/voltage

4
Synchroscalar - Architecture
5
Synchroscalar - Voltage/Freq

• Each column has independent voltage/freq
• Analysis of workload producer/consumer
  relationships
• Scale back column performance for less
  computationally intensive threads
• Power scales dramatically with frequency
• Up to 32 power savings on full applications
• Up to 83 power savings on kernels
• Power consumption only 8-30x worse than ASICs
• Number of tiles was varied for each application

6
Synchroscalar - Results
7
Background - Imagine

• Aims to provide very high performance on DLP
  workloads
• Media processing apps are a common workload
• High ratio of computation to memory bandwidth
• Ops are mostly independent, very latency tolerant
• Performance of an ASIC with more flexibility
• High percentage of area devoted to ALUs

8
Imagine - Architecture

• 48 ALUs organized into 8 clusters
• Each ALU has 2 ported local register file
• Outputs of ALUs go to all LRFs in cluster
• 544 GB/s
• Each cluster has 3 adders, 2 multipliers, and
  divide/sqrt unit
• Each cluster also has a scratch pad register
  file, and inter-cluster communication unit
• Clusters execute 576 bit VLIW instructions

9
Imagine - Architecture

• Large stream register file (128 KB)
• Single ported
• 22 logical ports
• Each buffer reads 2 blocks of 32 words
• Contend for single physical port
• Works because data accesses are streaming, can
  prefetch very effectively.

10
Imagine - Architecture
11
Imagine - Results

• Comparision to Nvidia Quadro
• 1/5 the speed in raster limited scenes
• Faster in geometry limited scenes