> Keep quoting ten-year old articles! Ok, will do. Until IBM says something else.
> As for POWER7, it is powering the new Blue Waters supercomputer at NCSA and > is funded by the > NSF and DARPA. It will hook together 16,384 Power7 chips! > The transfer rate between nodes will > be 192GB per second! Cool, but that is not a commodity server. It is a special tailored machine and is not interesting. Just because F1 cars are very fast, it doesnt say something about the commodity cars. Or, the POWER 595 IBM used to get the old TPC-C record. It had 2TB of RAM and several 1000s of short stroked hard drives. That is a pathological computer and no company uses such a machine for database work. As for the transfer rate of 192GB/sec I dont care too much about. I remember when IBM said the POWER6 has 220GB/sec bandwidth - and it turned out that IBM had added all L1+L2+L3 cache bandwidth and what not. If there is a bottleneck of 1GB/sec, then the chip will not be faster than 1GB/sec, no matter how much bandwidth the different parts of the chip has. You can not add up bandwidth. This follows from the max-flow=min-cut theorem from graph theory. Or, I remember when IBM said that one IBM Mainframe can virtualize 1500 of x86 servers. It turned out that all those servers idle and the Mainframe is 100% loaded! In that case, my laptop virtualizes 10 Mainframes (if they all idle) - this is hardly a true statement? And as we all know, the Mainframe cpus are dog slow. You need 10 Intel Nehalem-EX cpus to match the biggest IBM z10 Mainframe with 64 cpus - in terms of computing power. So, 192GB/sec is probably not true and I am very sceptical to that figure. > I'm not sure what you mean by "2013 there will be 8 socket T3+ machines with > 1024 beefy > threads." The IBM Power 795 already supports 256 cores and 1,024 threads. > That was in Feb > of 2010. So if you're saying in the year 2013 the T3 will support that, then > it is already 3 years > behind IBM. Yes, I hope that the biggest machine IBM with 32 cpus, will win over a T3+ machine with 8 cpus. Otherwise IBM did something wrong. But wait until a 32 socket Niagara machine comes out and compare to the biggest IBM machine, that would be fair. But I know IBM does not like fair comparisons. Typically, IBM compares a server with 16 POWER6 cpus, to a 4 socket Niagara machine - and concludes the "POWER6 is a faster cpu than Niagara" because the 16 socket server is faster. Only IBM thinks it is fair to compare 16 cpus to 4 cpus. IBM says: "both servers have equally many cores - and therefore it is fair comparison and the benchmark shows that POWER6 is a faster cpu". Well, if we talk about which cpu is fastest, then we should compare cpu vs cpu. Not core vs core which IBM does. > x86 cannot match that and isn't close, so I don't know why you Linux/x86 fans > continue to > think it is better than AIX/POWER or Solaris/SPARC. Clearly, Linux/x86 is > behind and the gap > is wide. I know x86 can not match that. But the gap is decreasing fast. Look at the trend and extrapolate. -- This message posted from opensolaris.org _______________________________________________ opensolaris-discuss mailing list opensolaris-discuss@opensolaris.org
