Generative SSDs Coming Soon

Memory hierarchy pyramid standing on its apex, with tape at the point on the bottom, and L1 cache on the widest part at the top. At the sides are arrows pointing from bottom to top marked "Increasing Cost" and "Increasing speed"Now that Generative AI has hit its stride, the industry is beginning to see new applications of this technology that have previously been out of reach.  One such application is the development of the Generative SSD by an Andorran company by the name of SSDystopia, who took the time to brief The SSD Guy about its development.

The Generative SSD comes to the market with a completely different approach to data management.  Rather than store significant volumes of data, it might be better to create that data from stored Tokens in an Attention Matrix that are capable of recreating information that normally would have been stored in the SSD.  This eliminates issues that have plagued storage management for decades, issues like duplicate data, and like large data sets that differ in some nearly-insignificant way to create reams of redundant storage locations.

“The Generative SSD promises to create not only CapEx savings in the storage media itself, but also it will lead to significant savings in energy and cooling, to dramatically reduce OpEx and TCO,” says Dr. Oihane Intxausti-Etxeberria, creator of the device.

Generative SSDs represent a complete shift in the thinking behind storage, since it no longer requires for the media to expand to match the volume of the stored data.  Unlike compression, where the media still needs to expand in proportion to the size of the data, the generative approach uses its AI-based architecture to generate the required data no matter how large the data sets needs to be.  “We envision a future,” says Intxausti-Etxeberria, “where improvements to the architecture actually allow the capacity of this device to shrink as the data set expands.”  Intxausti-Etxeberria explains that this becomes possible as the dataset’s entropy increases, since the only data that must be accounted for is the shrinking amount of data that is not included in the dataset.  “It is quite feasible,” she continues, “to fit the entire contents of the Internet into a single microSD card.  The cost savings of such an approach are difficult to even comprehend.”

The secret behind this interesting device is the large language model (LLM) itself.  These models now boast one trillion parameters, which sounds enormous until one considers that terabyte SSDs are common these days.  Pair that fact with recent developments to produce quite accurate LLMs based on modest token resolutions of a single byte or even less, plus algorithms that suppress the unused elements of sparse data, and you lay the groundwork for a quite sophisticated generative device that appears to hold significantly more data than its terabyte capacity is capable of containing.

Memory hierarchy pyramid standing on its apex, with tape at the point on the bottom, and L1 cache on the widest part at the top. At the sides are arrows pointing from bottom to top marked "Increasing Cost" and "Increasing speed"Dr. Intxausti-Etxeberria postulates: “People look at the memory/storage hierarchy the completely wrong way by expressing it with a pyramid.  Cost and speed increase as you go from the bottom to the top of the pyramid, so it makes sense to give the fastest tiers the widest elements.  That turns the whole thing upside-down, and like any upside-down pyramid, it is teetering on its apex, and will fall with the slightest push.  Our Generative SSD will provide that push.”  In the long run she anticipates the development of Generative CXL DRAM modules, and eventually generative DRAM chips.  It could extend even to cache memories which, over the long run, might eliminate the need for any off-processor data storage, provided that the persistent emerging memory technologies that are just beginning to bloom replace SRAM caches, as they are expected to eventually do.

Moore’s Law is helping to support this departure from standard storage methods.  Since compute is increasing in complexity as well as NAND flash density, the data stored can be reduced through the use of increasingly complex algorithms.  Data compression is an evolutionary approach to the use of this greater computational capability, but generational AI is now providing a revolutionary approach which will completely change data storage.

Weird pyramid found on the back of the US dollar bill. The top is an eye. Four levels are marked on the pyramid as they are in the memory/storage normal hierarchy pyramid, going from top to bottom: Cache, Memory, SSD, and HDD.“When you look at storage with a fresh approach,” says Intxausti-Etxeberria, “you find that compute has become cheaper than storage.  Organizations can save money by generating rather than storing their data.  This approach also consumes less energy than standard methods, so it’s more green.  The introduction of the Generative SSD is a point of no return.  The more data we wish to store, the more compute we should use to reduce our media requirements.  This puts an enormous dollar focus on storage.”  She emphasized this point with a diagram based on the pyramid that is printed on the back of every US one-dollar bill.

SSDystopia is currently sampling the water-cooled Generative SSD with some of Andorra’s largest ISPs, and is working with investors to bring mass production to the market by the first of next April.  The SSD Guy will be keeping our clients abreast of the company’s progress in that direction.

 

One thought on “Generative SSDs Coming Soon”

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.