Computers for lattice field theories

Abstract

Parallel computers dedicated to lattice field theories are reviewed with emphasis on the three recent projects, the Teraflops project in the US, the CP-PACS project in Japan and the 0.5-Teraflops project in the US. Some new commercial parallel computers are also discussed. Recent development of semiconductor technologies is briefly surveyed in relation to possible approaches toward Teraflops computers.

The CP-PACS project

“…The CP-PACS project [6] aims at developing a massively parallel computer designed to achieve high performance for numerical research of the major problems of computational physics. It further aims at significant progress in the solution of these problems through the application of the computer upon completion of its development.…”

QPACE 2 and Domain Decomposition on the Intel Xeon Phi

“…As before, (a) occurs after (1), but in the z-direction we now send boundary data when the computation on half of the boundary has finished, i.e., (b) occurs after (2). The other half of the boundary is sent at the end, i.e., (c) occurs after (5). The first half is needed at the start of the next iteration, which means that (b) overlaps with (3)- (5).…”

“…The nodes are connected by an IB-based network with a hyper-crossbar topology. This topology was introduced by the CP-PACS collaboration [5]. For QPACE 2 we use a two-dimensional version as indicated in Fig.…”

“…If the domains are small enough to fit in cache, the ensuing inversion on the domains (i.e., the block-diagonal parts) can be done without off-chip data movement. 5 Computation of the remainder does require communication but occurs less frequently. As a result, the overall algorithm needs less communication bandwidth, is more latency tolerant, and gives more cache reuse than the conjugate-gradient algorithm.…”

“…The other half of the boundary is sent at the end, i.e., (c) occurs after (5). The first half is needed at the start of the next iteration, which means that (b) overlaps with (3)- (5). The second half is needed before (4) of the next iteration so that (c) then overlaps with (1)-(3).…”