We have talked a lot about RAM, its new versions and speeds as well as the industry’s problems with it. Currently, processors and graphics cards are advancing at a much higher rate than RAM, which is generating bottlenecks and many companies are beginning to see further, especially in Exascale systems. Does DRAM have the hours counted? Will DDR5 be the latest version? Who will replace it?
AI has opened up a new field in computing, a new sector that is eager for resources and is devouring them with unusual speed. Bandwidth is experiencing its lowest hours as a concept, as the disparity between components is causing bottlenecks.
With MCM the processors have scaled in full power, the GPUs have HBM2E as a novelty and expand their power every two years, the SSDs raise their performance again and the DRAMs cannot keep up.
DRAM scaling and stacking may not suffice
The industry response to such bottlenecks is clear: DRAM needs to be scaled to smaller dimensions to increase capacity and increase speed. Although the forecasts are always negative with it for obvious reasons, it has always been the great one pointed out in the hardware world for going with the greatest performance delay compared to the other components.
Its end was predicted with 90 nm and instead we go for 10 nm, but the controversy is again on the table again for all the comments.
Except that now, DDR5 will be a momentary relief for the next step, which could end DRAM as we know it.
The 3D stacks will introduce HBM as we have already seen in Intel‘s Lakefield and also a new concept arises that puts the veteran DRAM in check, what are your real options to stay on the market?
Although limited, the market cannot get rid of it that easily
DRAM is always on the wire and manufacturers in the eye of the hurricane. But the market itself is aware that although it is the main bottleneck (we leave storage aside for obvious reasons) it has a number of benefits that other technology currently cannot offer:
- Easy byte level access.
- Relatively fast access times.
- Symmetrical writing and reading.
- Infinite data retention.
- Infinite resistance”.
- Mature technology and with low costs for its volume.
Seen in perspective, the dilemma of manufacturers and systems is understood, but what then is the solution for its non-disappearance? The increase in capacitance by area . Denser bit cell capacitors are needed to accommodate higher speeds, sizes and ultimately increase overall performance.
RAMBus claims that the speed doubles every 5 or 6 years, an insufficient time in the industry without a doubt and that determines an increasingly delicate cabling to design to support without errors the amount of data that is established.
What technology can supplant DRAM ?: NRAM
Apparently, the technology that can degrade DRAM to a new better life will be somewhere between DRAM itself and NAND Flash as such, something like the Intel and Micron XPoint but taken to a new level.
The new technology seems to be that it will include carbon nanotubes, also called CNTs and will take NRAM, where the N stands for Nanotubes. The technology has begun to be developed by Fijitsu, although its application is not being as NRAM as such, but as NVM devices.
The basis of the NRAMs are the Van der Waals forces, where the CNTs come together in such a way that they become a random mass of carbon tubes based on an electrode conductor. The bond is therefore voltage based and can be broken by thermal vibrations using an opposite voltage.
This effect renders NRAM inconsistent and that attempts by IBM and Samsung always failed, but where they did, Fujitsu hit the key: add one more layer of CNT arranged horizontally, which protects the cell as if it were a metal barrier.
The optimizations after this step have focused on the resistances and the width of the cells, thus achieving a switching speed of only 5 ns without depending on the size of the memory. It is said that they will be able to obtain 16 times the current density of the DRAM at the same speed as this.
Speed will be the second section to improve, but if it were to exceed DDR4 and later DDR5, only two large companies would be missing that would bet on it to break the current DRAM market. It is true that as with CMOS , the death of DRAM has been predicted several times, perhaps this is the best attempt that the industry will have to make the leap to a new concept that implies substantial improvements in the short and long term.