Simultaneous multi-processor apparatus applicable to achieving exascale performance for algorithms and program systems
Apparatus adapted for exascale computers are disclosed. The apparatus includes, but is not limited to at least one of: a system, data processor chip (DPC), Landing module (LM), chips including LM, anticipator chips, simultaneous multi-processor (SMP) cores, SMP channel (SMPC) cores, channels, bundles of channels, printed circuit boards (PCB) including bundles, floating point adders, accumulation managers, QUAD Link Anticipating Memory (QUADLAM), communication networks extended by coupling links of QUADLAM, log 2 calculators, exp2 calculators, log ALU, Non-Linear Accelerator (NLA), and stairways. Methods of algorithm and program development, verification and debugging are also disclosed. Collectively, embodiments of these elements disclose a class of supercomputers that obsolete Amdahl's Law, providing cabinets of petaflop performance and systems that may meet or exceed an exaflop of performance for Block LU Decomposition (Linpack).
1. A apparatus, comprising:
A. an anticipator chip adapted to respond to a system performance requirement by said system performing non-additive term generation using at least one Non-Linear Accelerators (NLA) for computing an algorithm and an incremental state of said algorithm received by said anticipator;
B. said anticipator chip is adapted to respond to said incremental state by creating an anticipated requirement; and
C. said anticipator chip is adapted to respond to said anticipated requirement by directing at least part of said system to achieve said system performance requirement through the use of said at least one NLA.
2. The apparatus of claim 1 , wherein said anticipated requirement, includes
A. an anticipated future memory transfer requirement of at least one memory unit array as an associated large memory to said anticipator chip; and
B. an anticipated future transfer requirement of at least one Landing Module (LM) chip as at least one associated communication node chip to said anticipator chip.
3. The apparatus of claim 2 , wherein said anticipator adapted to respond to said anticipated requirement includes said anticipator configured to perform
A. said anticipator scheduling memory transfers of said associated memory unit array to fulfill said anticipated future memory transfer requirement;
B. said anticipator configuring at least one of said associated communication node chips to fulfill said anticipated future transfer requirement.
4. The apparatus of claim 2 , wherein said anticipated requirement further includes:
A. an anticipated internal transfer requirement for a Data Processor Chip (DPC) as an associated DPC to said anticipator chip; and
B. said anticipator configuring at most one of said associated DPC to respond to said anticipated internal transfer requirement of said associated DPC with any coupled said associated communication node chips so that said performance requirement is met in the average over said sustained runtime.
5. The apparatus of claim 2 , wherein said system performance requirement includes said system performing a Kperf for said sustained runtime directed by said algorithm;
wherein said Kpref is a member of a first group consisting of ¼ exaflop to 1 exaflop.
6. The apparatus of claim 1 ,
wherein said anticipator further includes a state table adapted for configuration to integrate said incremental states of said algorithm to update said state table to account for said anticipated requirement; and
said anticipator responds to a successor incremental state based upon said state table in order to generate a successor anticipated requirement.
7. The apparatus of claim 6 , wherein said state table is adapted to integrate said incremental states of said algorithm to update said state table to account for said anticipated requirement, for each of said incremental states.
8. The apparatus of claim 7 ,
wherein said algorithm includes a form of Block LU Decomposition with partial pivoting of a matrix A including at least N rows and at least N columns of double precision floating point numbers, where said N is a member of a second group consisting of a number multiplied by K of said rows and said columns resulting in said Kpref sustained performance, where said K is 1024;
wherein said number is a member of a group consisting of ¼ to 16;
wherein said sustained runtime is at least one and no more than 8 hours; and
wherein said incremental state includes a pivot decision for one of said columns of said matrix A;
wherein said at least one NLA is used to perform at least one of double precision floating point reciprocal and/or double precision floating point division based upon said pivot decision.