You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Softmachines, a bay area start up claims 2-4X performance with its new VISC architecture?
If all it takes is 3 magic extra stages in a pipeline to double IPC, then why has nobody else come up with this? No memory wall, no power wall, no sir, just add some more cores, and enjoy ever faster single-threaded performance! That all sounds way too good to be true...
With $125M in funding, 250 people, and 7 years to bring a chip to market I'd have to say that they are masters at fund raising but kind of slow to get silicon. When I see a customer endorsement and a competitive benchmark, then the company starts to become real in my mind.
I think most of the breakthroughs going forward are going to require a lot of software support in order to get old code to run properly and faster, and you want to build that piece first - before burning cash on Silicon.
TransMeta ran into problems because people didn't use code the way they expected.
I have a way to do stuff that should give a 2x improvement; there's a lot of software work involved in that too.
I would say that the chances of success are better if everybody is stuck at 28nm rather than having to compete with Intel on the next node,
Might improve SMT, SIMT and also 'IPC' in future Intel processors... If it all was not just a hoax (existing CPUs with 2-4x more resources per core will achieve up to 2-4x better 'instruction per clock' too, if in software will be enough parallelization).
