Elpida's ReRAM prototype will pave the way for volume production in 2013, the company claims.
Japanese memory specialist Elpida has announced the development of a resistance memory (ReRAM) prototype, which it claims paves the way for high-speed non-volatile storage that combines the best of DRAM and NAND flash technologies.
Jointly developed with the New Energy and Industrial Technology Development Organisation (NEDO) in Japan, the prototype boasts an overall capacity of 64Mb in a multi-cell array based on a 50nm process size. While that might not sound like much, it's one of the highest capacities yet achieved with ReRAM technology.
As the name suggests, resistance RAM uses a material which changes resistance in response to changes in applied electrical voltage. Because the changes are permanent - at least, until a new voltage is applied to reverse the change - the memory stores its data even when the power is cut, in the same way as the NAND flash cells used in solid-state storage devices.
Unlike NAND flash, however, ReRAM has performance characteristics around the same as dynamic RAM (DRAM), the normal RAM found in a PC which is extremely fast but which wipes its stored data following the loss of power.
According to Elpida's testing, the ReRAM prototype boasts a write speed of 10 nanoseconds and write endurance of one million times, a ten-fold improvement over the best NAND flash currently available.
While not yet ready for commercialisation, Elpida has promised that it will continue developing ReRAM technology in partnership with NEDO, Sharp, the National Institute of Advanced Industrial Science and Technology (AIST,) and the University of Tokyo with a view to producing gigabit-class ReRAM modules in volume using a 30nm process by 2013.
If the company manages to hit its goals, it claims that its ReRAM modules will be a serious competitor to standard NAND flash for smartphone, tablet and ultra-thin laptop storage thanks to both its improved performance and extremely low power draw.
ReRAM isn't the only memory technology which aims to combine the non-volatile nature of flash with the speed of DRAM, however: with IBM continuing its work on
magnetoresistive RAM (MRAM) and
racetrack memory, HP and Hynix teaming up on
memristors for their own ReRAM products and
DFG-FET offering yet another route to universal memory, Elpida could have a fight on its hands for a share of that oh-so-lucrative high-performance storage market.
Are you impressed with Elpida's efforts, or saving your applause until the first volume production begins? Share your thoughts over in the
forums.
9 Comments
Discuss in the forums ReplySadly, the technology is a fair way away from that level of performance and affordability... Still, it's a nice thought for the future.
I can see how ReRAM could be pretty useful, especially in embedded markets such as routers where you could cut out a chip, maybe some complexity and have extremely fast storage, so a lot of operations like restarting a router are near instantaneous as you are using the same general pool for main memory and storage of the OS. I could also see how it could benifit things like tablets and phones with a lower power standby mode, lower idle power (if nothing is being actively written to the memory, no idle memory power consumption instead of, I guess, needing to keep the main memory in a higher power state than "self refresh") and much faster boot up if you can just save the state in main memory, which wouldn't purge on "shutting down".
With storage that has the same throughput performance as DRAM, why would you want to replace it...?
It may be at some point, but as they said, it hasn't been commercialized yet. I don't think it will replace RAM, but I could see one or both of these happening: 1) it's used mainly in smartphones/tablets, etc, as replacement or augmentation of current storage, RAM and storage. However, it won't be marketed as such, just whatever BS name the manufacturers come up with, and so you might not even realize it's in the wild. 2) augmentation of current PC RAM/storage, where it's been given an equally silly name as in (1), but it's also pushed as ReRAM, as that will give nerds boners.
Obviously not both can be entirely true at the same time (not knowing it's out vs nerd boners, for example), but I think it's inevitable.
DDR 1600 CL7 memory has an access time of 8.75ns, which is only "a bit" faster than the 10ns that this stuff supposedly can handle. That is the equivelent to CL8 memory at DDR1600 (which has an access time of 10ns).
This does not take in to account the actual round trip time from the CPU to main memory, which is part of the reason why CL7 really doesn't translate to significantly higher performance than CL9 memory does, even though the access time is about 25% faster. On a Sandbridge processor there is a latency of about 150 clock cycles to retrieve data from main memory. At 3.3Ghz that translates to about 45ns of built in lag (not sure if that takes in to account the actual memory access time as part of the latency or not). So the memory chip latency accounts for at most 18% of the overall latency and possibly as little as 14%, so even a "big" change in main memory chip access times only changes overall main memory latency a small amount (25% of 14% leads to a 4% change in memory latency).
L1 cache speed is roughly .5ns and L2 is roughly 7ns AFAIK that takes in to account the clock cycles, with L1 cache is 2 clock cycles on SB and I think L2 is around 10-15 clock cycles (L3 I think is around 30-40 clock cycles). So as you can see, the memory access time is ridiculously low, well below a nanosecond for L1 cache. So if you use slower access memory on the CPU die, you are going to severly limit processor speed. You want the gosh darned fastest switching transistors you can manage on them (even for L3).