AS/400 chief architect Frank Soltis told our sister publication IBM System User he’s convinced that Very Long Instruction Word techniques will replace RISC technology, which he believes is reaching its limits. RISC performance is not enough. It will quit by the end of the decade. What is needed is a different processing structure, says Soltis. […]
AS/400 chief architect Frank Soltis told our sister publication IBM System User he’s convinced that Very Long Instruction Word techniques will replace RISC technology, which he believes is reaching its limits. RISC performance is not enough. It will quit by the end of the decade. What is needed is a different processing structure, says Soltis. This is where Very Long Instruction Word fits in. Very Long Instruction Word is a set of compiler techniques that moves scheduling off the processor. With RISC architecture, part of the processor space has to be given up to schedule the best order in which to execute instructions. Because scheduling is carried out at compilation, Very Long Instruction Word should to enable greater parallelism in instruction execution. Each Very Long Instruction Word consists of any number of instructions between one and 20, which can be executed concurrently. Intel Corp and Hewlett-Packard Co are jointly developing Very Long Instruction Word technology for the P7, but Soltis believes the AS/400 developers have the lead.
VLIW versus RISC
Hewlett-Packard is saying it is getting three times the performance using VLIW. We are getting improvement of five times, he says. However, Soltis admits that his enthusiasm for Very Long Instruction Word is not shared throughout IBM. Some believe RISC has a much longer future. It is a debate under way now, he says. In the traditions of the AS/400 division, Soltis is convinced that the AS/400 will lead IBM into implementing Very Long Instruction Word throughout its machines. He is confident of the outcome of the debate. The AS/400 wants VLIW and it won’t go it alone. It will pull other divisions along to keep compatibility, says Soltis. The need for all the different divisions to agree a long-term strategy fits well with another of Soltis’s ambitions for the AS/400 and IBM – a single server, enabling components to be used across different IBM operating systems, from OS/2 Warp through to MVS. At some point there should be an IBM single server – one box where the only differentiator will be the operating system, he says. But at the moment, he says, there is still resistance within IBM. I believe customers would back it. It is just selling and marketing people at IBM that don’t want it, he says. We already share a lot of components with PC Server, RS/6000, AS/400 and the mainframe, says Soltis. There are already plans for a new low-end AS/400 server using common components with IBM personal computer servers that will be priced at the same level as the iAPX-86 servers. They will blow away the theory that PC Servers are cheaper. Alongside development of Very Long Instruction Word, Soltis claims the AS/400 division is also leading IBM’s development of NUMA Non-Uniform Memory Architecture. This is partly due to the architecture of the AS/400. NUMA fits AS/400 much better than other types. The RS/6000 division is not going with NUMA as, with all traditional Unix systems, there is no address space sharing. For Soltis one of the key strengths of the AS/400 is in the capacity to build in new technologies. Since launch the machines have been able to support 128-bit architecture, proving, says Soltis, that in integrating new technologies IBM can change the internal structure of the AS/400 and not lose customer base. There are also plans to run AIX on the AS/400 input-output co- processors but the facility will only be offered, in the short term at least, on a very limited basis.