3 ms·

Few years ago HP made sensational claims about its memristor-based Machine project. One can doubt this will ever become real, but their assessment that memory,

Few years ago HP made sensational claims about its memristor-based Machine project. One can doubt this will ever become real, but their assessment that memory, not logic, is the main bottleneck in computing seems to make sense. Yet it seems that tech news keep being unreasonably excited about all research about new transistors designs, no matter how low they are on the technological readiness level. That's deplorable, IMHO.

nine_k 10y ago
Well, this means that we have a potentially great technology that possibly can make a difference 10 years from now (or more). It's good to know, but it's not something to expect in new consumer products by the coming Xmas. Many things were invented way before their mass adoption. For instance, LED effect was known since 1907, first prototypes built in 1920s, but commercial LED indicators only took up in 1970s, and LED lighting, in 2010s. So we just have to wait — unless you're a solid state physics researcher, then you can actually work on making it practical.
WorldMaker 10y ago
HPE is still making sensational claims about the Machine project and its commercial viability. The impression I got the last time I talked to an HPE person on the subject (so this is third-hand subjective, take it as you will) is that the hard technology has been ready and practical to some some sort of productization for a while now, but the entire project got boondoggled in trying to "truly revolutionize" and rebuild computing from first principles including brand new OSes and applications... It's sounds like an interesting lesson in a sort of reverse Sagan Effect where "we have a new whipped cream" turns into "well, to really sell it we first must produce an apple pie from scratch... and do that well, how about we try a new big bang?"
Dylan16807 10y ago
We have the ability to build eDRAM with latency under 2ns and directly attach mass quantities to a CPU. But it's more cost-effective to have a fast L2 and let your L3 and main memory be slower. In a multi-CPU machine you have to wait around anyway for a lot of memory accesses. It's not the memory technology that bottlenecks memory. It's locating and routing data. And that cares a lot about better transistors.

nine_k 10y ago

Well, this means that we have a potentially great technology that possibly can make a difference 10 years from now (or more). It's good to know, but it's not something to expect in new consumer products by the coming Xmas. Many things were invented way before their mass adoption. For instance, LED effect was known since 1907, first prototypes built in 1920s, but commercial LED indicators only took up in 1970s, and LED lighting, in 2010s. So we just have to wait — unless you're a solid state physics researcher, then you can actually work on making it practical.

WorldMaker 10y ago

HPE is still making sensational claims about the Machine project and its commercial viability. The impression I got the last time I talked to an HPE person on the subject (so this is third-hand subjective, take it as you will) is that the hard technology has been ready and practical to some some sort of productization for a while now, but the entire project got boondoggled in trying to "truly revolutionize" and rebuild computing from first principles including brand new OSes and applications... It's sounds like an interesting lesson in a sort of reverse Sagan Effect where "we have a new whipped cream" turns into "well, to really sell it we first must produce an apple pie from scratch... and do that well, how about we try a new big bang?"

Dylan16807 10y ago

We have the ability to build eDRAM with latency under 2ns and directly attach mass quantities to a CPU. But it's more cost-effective to have a fast L2 and let your L3 and main memory be slower. In a multi-CPU machine you have to wait around anyway for a lot of memory accesses. It's not the memory technology that bottlenecks memory. It's locating and routing data. And that cares a lot about better transistors.