Metaram's announcement of a chipset that enables vendors to build 16 GB DRAM DIMMs may have you wondering: who needs it?
It probably isn't you.
Stacking DRAM chips several deep gets the capacity, but the secret sauce resides in interface chips that performs several key functions:
- Only powering the DRAM when it is in use to reduce the power and cooling load on a systems not designed for so much memory
- Faking out the CPU's memory controller so Metaram's controller chip looks like DRAM, while also emulating a CPU controller chip to DRAM.
- Handling the latency issues with technique that looks like something disks do - tagged command queuing - that enables multiple command issues with out-of-order responses.
The chipsets will add about $200 to the cost of a DIMM.
Disk vs. RAM
The late, great Jim Gray commented in his most excellent paper Rules of thumb in data engineering on the relationship between RAM and disk:
The cost/MB of RAM declines with time: about 100x per decade. Since disk and RAM have a 1:100 price ratio, this price decline suggests that what is economical to put on disk today will be economical to put in RAM in about 10 years.
That actually works. About 10 years ago I bought a system with a 4 GB hard drive. Now my main system has 5 GB of RAM and a terabyte of disk.
Metaram isn't targeting this to home users, even gamers. They are focused on database and workstation apps.
Loading an entire database into RAM really improves performance. Avoiding database corruption due to power loss is the next question.
The Storage Bits take
DRAM is just fast storage. Stuffing more of it into servers will help speed them up to handle more jobs with less energy and rack space.
Us home users will have be content with cheaper, standard memory for now. For me, that's plenty.
Comments welcome, of course.