BlueLight design approach addresses all five of the ASCI challenges
Attack the distance to memory problem
- Large on-chip Embedded DRAM (EDRAM)
- DDR-2 SDRAM on small node
- Low latency, enormous sustainable bandwidth
Architecture optimized for cost/performance
- Address a large, but limited set of applications
- $100/GF vs $9,000/GF for White and
$7,000/GF for 30 TeraOPs
- Cost of system also determined, to a lesser extent,
by memory and disk scaling
Optimized for power,cooling and floor space
- New generation of embedded processors
- Low power, low cost, high performance
- Low risk, easy to implement with standard system-on-chip offerings
Architecture optimized to provide high efficiency for modern applications
- Modern applications have complicated code and data structures which result in random memory access patterns spanning a vast amount of memory
- Enormous bandwidth to memory at low latency with small cache line ? good delivered performance for modern applications of national interest
- Sustainable performance previously restricted to high end processors
- Efficient implementation of MPI global operations
Proven method
- BlueLight builds on 4 generations of previous QCD machines