On Tue, Nov 30, 2004 at 01:00:40AM +0800, Yan King Yin wrote: > It doesn't sound plausible that ordinary people will buy clusters
I didn't realize this was your premise. I was thinking about a current developer looking for affordable crunch. The Infiniband is becoming a very interesting option, due to the impending commodization. 10 GBit Ethernet isn't nearly there yet. As to commodity, most hosting places have seas of boxes connected to Level 2/3 switches. > in a popular way. Clusters of course are a good solution for > developers, but I'm talking about "bringing parallel power to the > masses". Where are the killer applications for the masses? You think many game developers grok MPI? The most likely place would seem an online game server. The Cell processor has just been shipped to selected game developers. The hardware is chronically overhyped, but it's definitely going towards affordable parallelism for the end user (adolescent and young adult gamer). > > FPGA accelerators are good, but you can't feed them fast enough. FPGA with > > onboard memory might be pretty optimal. > > This sounds promising, especially because of lower R&D costs. Are you talking about individual developers? You can get a very decent cluster for a price of a single FPGA developer kit. It is most emphatically not cost effective. > > Because there's no market. > > A market can be created. It's probably because of the risks A man can walk on the Moon. Just do it. > of dealing with highly complex architectures that people have not > figured out how to do it. If you're talking about "people" I'm no longer interested. "People" are mythical. > Thanks for the info. I found this article (2003): > http://www.computer.org/computer/homepage/1003/entertainment/ > "The GPU Enters Computing's Mainstream" > and it says that current GPUs do not support branching and pointers, 2003 is not current. You should check out http://gpgpu.org/ In general http://www.google.com/search?hl=en&q=GPU+scientific+computing&btnG=Google+Search will give you lots of juicy bits. > which makes them limited as general purpose platforms. But they are > interesting and certainly have the potential to evolve into future > parallel platforms. They're not interesting. GPU specialists are rare, and innovation cycles will break your code. It's not worth the effort. Ditto DSPs, this is only relevant for specialist niches. COTS clusters are the way to go. If you're latency-starved, platinum-plated interconnect might be worthwhile. That's the status quo, in a nutshell. -- Eugen* Leitl <a href="http://leitl.org";>leitl</a> ______________________________________________________________ ICBM: 48.07078, 11.61144 http://www.leitl.org 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE http://moleculardevices.org http://nanomachines.net ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
pgpo3nlibqEeM.pgp
Description: PGP signature
