The FG-DRAM is one of the developments that NVIDIA has in the bedroom in the future, it is a new type of memory that comes to replace the HBM2 as memory for high performance computing. Will we see this type of RAM in the NVIDIA Tesla with NVIDIA Hooper chipset? We don’t know, but for the moment we give you a little introduction.
Usually the way in which conventional RAM has evolved has been by taking advantage of the new manufacturing nodes to get faster memory, in bandwidth, under the same level of consumption, but with quite mild results.
To explain to you what is the problem of the evolution of RAM with respect to its performance, bandwidth, imagine that we have an interface that transmits 1 GB / s, which has about 64 pins and we want to expand it to support 2 GB / s , for them we have two options:
- Double the clock speed.
- Double the number of pins.
Raising the clock speed has the problem that consumption increases exponentially with clock speed and voltage and the worst thing is that to reach certain clock speeds the voltage has to rise, so the increase in consumption would not be 4 times higher, but even higher so that we may run out of enough energy to fuel the memory.
The other option is to increase the number of pins, this makes it necessary less and less clock speed, thereby reducing the voltage and reducing energy consumption. But the problem that this entails is that it is to expand the perimeter of the chip, so it increases its cost, so while it has a better solution for consumption, it is not better in terms of cost.
But the solution to this dilemma came through silicon pathways, to connect the memory interfaces vertically on the interposer in which both the processor and the memory were connected, with this the HBM memory was born, which allowed huge bandwidths, but with much lower power consumption.
But time passes and the HBM memory is falling short in terms of its capabilities since a new type of memory is necessary with a much lower energy consumption. That is why all eyes are on other types of memories such as Fine Grained DRAM.
Table of Contents
What is the FG-DRAM?
In 2017 NVIDIA made a presentation to talk about a type of DRAM called FG-DRAM or Fine Grained DRAM as a replacement proposal for the HBM2, which has been used in NVIDIA Tesla cards since the Tesla P100
The FG-DRAM like the HBM2 is a 2.5DIC type memory with the processor (or processors) on top of an interposer next to the memory stacks. Its strong point? It has the particularity of going from the 3.9 Pj / bit of the HBM2 memory to 2 Pj / bit, so it can offer twice the bandwidth under the same consumption as the HBM2.
From HBM2 to FG-DRAM
One way to evolve the HBM2 is to simply accept configurations with a higher bandwidth, by placing a greater number of chips in the stack, the problem is that as we are placing more and more chips then the probabilities that all the battery goes bad they trip, this is the reason why we do not see HBM2 memory with more than 4 memory chips per battery.
But the Fine Grained DRAM takes a different path than what has been proposed with the HBM 3 in order to obtain a stack with a bandwidth of 1 TB / s. To do this, the idea is to increase the number of banks in each memory chip, up to a total of 512 simultaneous banks, which can be accessed at a speed of 2 GB / s each.
Thanks to that. The FG-DRAM memory stack is accessible at 1TB / s for twice the power consumption of the HBM2 at 256GB / s . But since you get four times the bandwidth, this means that the consumption per transmitted bit is cut in half .
Discussion about this post