For those, like me, who wondered how much data was written in 400 picoseconds, the answer is a single bit.
If I’m doing the math correctly, that’s write speeds in the 10s-100s GBps range.
If it’s sustainable.
1 bit / 400 picoseconds is 2.5Gbit/s, or 10x slower than a 1-bit GDDR7 bus (which the 5090 runs at 28Gbit/s * 512 bits).
To be fair this is non-volatile memory though, so the closest real comparison might be Intel Optame. The speeds actually seem somewhat comparable to DDR5, though even that is starting to run in to physical distance and timing issues. The real questions will be around density, cost, and reliability.
You can always parallelize, this would be more beneficial for latency.
Still about 100 picoseconds too slow for my taste.
400 for my use case, we’re trying to violate causality
The human eye can’t even perceive faster than 1000 picoseconds, so…
Really? I would have guessed the eye was 6 orders of magnitude slower than that.
What, you can’t measure the size of a room by timing the bounces of light hitting the walls?
No! I didn’t know that’s how you guys were doing it. I feel silly for using perspective and the slight differences from my right and left eyes to judge distance this whole time!
Sheesh, you’re living in the stone age my dude.
Other than just making everything generally faster, what would be a use-case that really benefits the most from something like this? My first thought is something like high-speed cameras; some Phantom cameras can capture hundreds, even thousands of gigabytes of data per second, so I think this tech could probably find some great applications there.
There’s some servers using SSDs as a direct extension of RAM. It doesn’t currently have the write endurance or the latency to fully replace RAM. This solves one of those.
Imagine, though, if we could unify RAM and mass storage. That’s a major assumption in the memory heirarchy that goes away.
This was actually the main market for Intel Optame. It’s got great write endurance, and better latency than Flash. I think they ended up stopping making it because it wasn’t cost effective. I’m actually using some old Optame drives in my server for the OS boot drive.
The article highlights on device AI processing. Could be game changing in a lot of ways.
I doubt it would work for the buffer memory in a high speed camera. That needs to be overwritten very frequently until the camera is triggered. They didn’t say what the erase time or write endurance is. It could work for quickly dumping the RAM after triggering, but you don’t need low latency for that. A large number of normal flash chips written in parallel will work just fine.
The speed of many machine learning models is bound by the speed of the memory they’re loaded on so that’s probably the biggest one.
Unfortunately this 1 bit / 400 picoseconds metric is 10x slower than GDDR7. The applications for this will be limited to things that need non-volatile memory.
It’s using graphene so we’ll see this as soon as the 100s of graphene innovations come too in who knows when?
So… How many cycles can it withstand?
At least 1
