Stage 3: The Ultimate Data Center in Space
In the 1st blog in this series we discussed how Medium density MRAM modules acts as a buffer or level 4 cache to enable Data Storage Satellites to capture data from 100s of LEO satellites without a real time link to earth. Such systems with MRAM modules with densities of 1-10TB guarantee secure and safe intermediate storage before transmission either down to earth or true data centers in Space. In the 3rd and final part of this blog series, we cover how the invention of the selector enables high enough density MRAMs solution that change the system architectures to pivot and enable standalone space data centers. This enables the storage of data for machine learning/deep learning model generation.
Ultimate space data centers evolution:
- Today & Near Future
- The current space data center evolution looks more like our terrestrial on-premise data centers, with both the specialized processing and storage entities residing as boxes on the satellite that is capturing the data. However, the satellite data centers are evolving right before our very eyes to leverage the data from the countless sensors orbiting our planet and aggregate it in a more cost effective and focused approach. This enables established companies and platforms to better stay within their swim lanes and allow other entrepreneurs to help grow out the rest of the new ecosystem. At this stage of the evolution, the memory aspect is more of the traditional one constrained by the space environment and sized in the 10s of Terabytes (TB) for getting the data down to the ground at a deterministic rate with minimized risk of data loss, when in view of ground stations. The data center is growing in this phase by adding modules to the satellite boxes, with MRAM storing the high-speed streaming transient data from the sensors before it is committed to NAND or always-on DRAM.
- These Medium density MRAM modules ~4TB are built using monolithic MRAM devices that use a transistor to switch the MRAM cell. The gating element in the ability to shrink the device has not been the size of the MRAM pMTJ cell, but the transistor capable of delivering a switching current (Learn more HERE). Avalanche’s Space Grade Gen 3 monolithic high reliability 1Gb devices are at the limit of what is possible at 22nm. These devices have been optimized for performance and reliability.
- Further Out Future…
- The next step in the evolution of MRAM (increasing density) is not going to come from the move to the next geometry 14-12nm or 7-5nm. Such moves need a transition from standard transistors to either vertical transistors or FinFets capable of delivering the required switching currents. The next evolution and increased density will come by replacing the transistor with a Selector that operates between an “AP” and “P” resistance states (Learn more HERE). The “Selector” can be reduced in size compared to a transistor by a factor of 4, and such circuits can be stacked to increase the density by orders of magnitude giving us the densities in 100s of TeraBytes of pure high reliability storage for Space.
- This would allow the storage of sensor data in space to tune their AI/ML models, and continuously evolve the autonomous space systems to more accurately and timely adjust their tasking and generation of actionable data. This enables the data centers to evolve into the more Space-IoT evolution of beaming the data from the sensors to the specialized ultimate data centers that store and process the data.
These advancements enable Space to become even more autonomous and independent of Terrestrial support, eliminating the need for a link to Earth. This is imperative if we are to successfully deploy a similar model around the Moon, Mars and Beyond. Backhaul to earth will not be an option.
In the next Blog series we will discuss how Compute environments in space will change as a result of introduction of DDR MRAM, which brings a level of resilience to space computing unknown to date. Such simplified system architectures speed up compute systems that enable a new generation of machine learning/deep learning models in space.