Related News

Details on IBM's Blue Gene Q available

Last Updated on Monday, 18 March 2013

Information about the latest Blue Gene/Q(TM) (BGQ) series of supercomputers by IBM is becoming increasingly available. Capitalizing on the experience gained by building the two earlier generations of supercomputers (the Blue Gene/L(TM) and Blue Gene/P(TM) series), in the BlueGene/Q series IBM has put together a number of innovations that promise to realize unprecedented performance, scalability, reliability and at the same time energy efficiency.

At the heart of the BGQ system there is the Blue Gene Compute chip, a System-on-a-Chip (SoC) design integrating 18 processor units (cores), cache memory and a powerful interconnect. Each processor unit has a Single Instruction Multiple Data (SIMD) Instruction-Set Architecture (ISA) model, with a 4-way multi-threading capability. Among the 18 processors (1.6 Ghz 64-bit PowerPC A2 core), 16 are targeted to actual applications, 1 is dedicated to specific Operating System services and the last one is there for redundancy.
Reliability is omni-present at all levels of the design, from the register file within each processor unit employing error detection and recovery, to the Compute chip design, embedding 1 redundant core, to the memory subsystem supporting ECC.

The cache memory design has noteworthy characteristics to mention: it supports an innovative intelligent memory pre-fetching mechanism that is able to learn complex memory access patterns; it supports a number of atomic operations; it embeds a “Wake-up” unit that monitors memory locations accessed during spin-lock operations on behalf of the corresponding hardware threads, which can suspend instead of keeping re-trying the memory access over and over again, with performance benefits for the other threads; it is able to implement a “multi-version” memory model, supporting hardware-assisted transactional memory and speculative execution. These features need to be coupled with proper usage of proper programming paradigms, such as transactional memory blocks and speculative for loops.

The new architecture poses even more challenges on the capabilities of the software stack to catch-up with, and properly exploit, the new hardware features. From a S(o)OS perspective, the dedication to Transactional Memory and Speculative Execution support within the Operating System are of particular interest for future software layers. Once these new programming paradigms are well-understood, deployed and used, applications, that already can take advantage from a software-only implementation of these features, will scale even better in presence of such hardware supports.

With an internal 5D torus interconnect supporting 563 GB/s of bisection bandwidth and a memory bandwidth of 32 GB/s, each BG/Q processor can deliver 204.8 GFlop/s of peak computing power. Also, BGQ gained top rankings in the Top Green 500 list of most energy efficient supercomputers, with 2 GFlop/s per Watt.

With 1024 16-core compute chips, a BGQ rack can deliver 209 TFlop/s of peak computing performance. The largest deployment of BGQ under realization, the “Sequoia” system in the Lawrence Livermore National Laboratory (LLNL), to be completed in 2012, will be composed of 96 BGQ racks containing 1,572,864 cores exhibiting an overall peak power of 20 PFlop/s.

LLNL, a premier multidisciplinary national security laboratory for DOE’s National Nuclear Security Administration (NNSA), applies some of the world’s most powerful supercomputers to maintaining the nation’s aging nuclear deterrent without testing, as well as such challenges as grid and network management, energy research and climate change.

Therefore, the Sequoia installation at the LLNL will readily benefit the human being by allowing for less real and more virtual nuclear-weapon explosions. Who knows if these big supercomputers will ever gain the intelligence to not even start a nuclear war, like in that famous science-fiction movie in which a supercomputer was learning that “the only winning move is not to play” from playing Tic-Tac-Toe, refraining itself from executing a program leading to the destruction of the world.

BlueGene is trademark by IBM. Read more on their website.