Non-Volatile Memory Arrives
Previously I've talked about RRAM, and how non-volatile memory is going to come in and cause lots disruption in computing. The non-volatile part still looks to be happening but it appears I might be wrong about it being RRAM that does it, though, since now samples of the first standard memory sticks of non-volatile RAM are actually being sent out (PDF), and it's not RRAM like I expected.
Rather, its MRAM or magnetic RAM. That stuff has actually been around for a while, I used some back in '08 when I needed a bit of non-volatile memory I could write to very fast but didn't need a large amount of storage. That last was the reason it wasn't in wide use, though. MRAM was fast, and low power, and many other wonderful things. But each individual MRAM cell was also very big, which meant that you couldn't fit very many of them on a chip. And that meant that on a bit-by-bit basis it was very expensive. But recently people have figured out a way to make MRAM cells much smaller using a technology called Spin Torque (ST) MRAM. And that's what the memory sticks going out now are using.
ST-MRAM does have one disadvantage compared to regular MRAM. Whereas standard memroy and plain MRAM could be written to all day forever and never wear out but it looks like the new ST-MRAM will have a finite write endurance. Luckily what I've been able to gather says that that write endurance is on the same order of magnitude as RRAM - stupendously larger than that of Flash memory - and so something you could reasonably put in your computer if you have a few levels of cache sitting between your processor and main memory, like all modern computers do.
This is still early stages, though. While Everspin is sending out samples of a standard DDR3-1600 memory that you could plug into your PC and use, this first one is only 64 megabits. Which is much better than a memory module made from the older style of MRAM would have been, but still not horribly impressive. You see, they decided to make the first batch on an old, cheap, and conservative 90 nm process. These days, state of the art components come out on 20nm or 22nm processes. Especially memory, which due to being fairly simple and regular is often the first thing foundries are able to produce in bulk on a new process.
Was this because it's much less expensive to get things fabbed on older processes and they weren't sure of market demand yet? Is it just that they're getting the kinks worked out as they scale MRAM down from humongous to reasonable sizes? Probably both, I'd guess but if they are able to scale down the chips to take advantage of modern processes then you'll be seeing 1 or 2 GB of memory per stick which is pretty reasonable. Not quite as dense as our current DRAM, but if they're already able to achieve current DRAM speeds at 90nms I imagine that later generations will end up much faster. And much more power efficient too, since the cells don't have to be refreshed regularly.
Will ST-MRAM dominate future main memory? Or will other technologies come into their own before it scales down far enough to be competitive from a storage perspective? I don't know but I bet things will be interesting.
Here are a couple of other places that have covered this: SemiAccurate, XBitLabs.
Rather, its MRAM or magnetic RAM. That stuff has actually been around for a while, I used some back in '08 when I needed a bit of non-volatile memory I could write to very fast but didn't need a large amount of storage. That last was the reason it wasn't in wide use, though. MRAM was fast, and low power, and many other wonderful things. But each individual MRAM cell was also very big, which meant that you couldn't fit very many of them on a chip. And that meant that on a bit-by-bit basis it was very expensive. But recently people have figured out a way to make MRAM cells much smaller using a technology called Spin Torque (ST) MRAM. And that's what the memory sticks going out now are using.
ST-MRAM does have one disadvantage compared to regular MRAM. Whereas standard memroy and plain MRAM could be written to all day forever and never wear out but it looks like the new ST-MRAM will have a finite write endurance. Luckily what I've been able to gather says that that write endurance is on the same order of magnitude as RRAM - stupendously larger than that of Flash memory - and so something you could reasonably put in your computer if you have a few levels of cache sitting between your processor and main memory, like all modern computers do.
This is still early stages, though. While Everspin is sending out samples of a standard DDR3-1600 memory that you could plug into your PC and use, this first one is only 64 megabits. Which is much better than a memory module made from the older style of MRAM would have been, but still not horribly impressive. You see, they decided to make the first batch on an old, cheap, and conservative 90 nm process. These days, state of the art components come out on 20nm or 22nm processes. Especially memory, which due to being fairly simple and regular is often the first thing foundries are able to produce in bulk on a new process.
Was this because it's much less expensive to get things fabbed on older processes and they weren't sure of market demand yet? Is it just that they're getting the kinks worked out as they scale MRAM down from humongous to reasonable sizes? Probably both, I'd guess but if they are able to scale down the chips to take advantage of modern processes then you'll be seeing 1 or 2 GB of memory per stick which is pretty reasonable. Not quite as dense as our current DRAM, but if they're already able to achieve current DRAM speeds at 90nms I imagine that later generations will end up much faster. And much more power efficient too, since the cells don't have to be refreshed regularly.
Will ST-MRAM dominate future main memory? Or will other technologies come into their own before it scales down far enough to be competitive from a storage perspective? I don't know but I bet things will be interesting.
Here are a couple of other places that have covered this: SemiAccurate, XBitLabs.
Comments
Post a Comment