How Recommendation Systems Leverage Intel’s Storage Solution
Intel presents new storage solutions providing huge amount of RAM with faster I/O, etc.
Hey hey everyone, welcome to another new article from me! Today, I’m exploring an innovative technology from Intel that will affect every part of the devices you use today; smartphones, computers, the servers running at Data Centers of cloud providers that streaming companies like HBO Netflix, and Disney+ use when you stream movies.
Source: Tech Field Day
The image above describes new storage products from intel.
Let us briefly look at the buyers of these products.
Buyers
Cloud Service Providers running Data Centers whose servers require fast SSDs, RAM, and Processors. These computes will be resold to customers as cloud services.
Computer Manufacturers whose devices require RAM, SSD, and Processors
Mobile devices playing heavy graphic demanding games like Modern Combat 5
Internet service providers who provide Edge computing solutions for Industrial Robots, Streaming, etc
Source: Bitesize
RAM
With more RAM, you have more apps or browser tabs (that you probably will often never go back to read) running smoothly in the background without the fear of it shutting down. Your operating system which manages your software and hardware decides how much RAM memory your browser needs to run those many tabs you have open.
Also, just to be clear the operating system is stored in the Hard Disk but is copied into the RAM when the computer comes on.
Hard Drive
All those apps or browser tabs will load faster as you open them. Your operating system also decides how the browser accesses it when you download files into it or try to upload your favourite media file.
Processor
It works on the apps running on the RAM, calling every resource it needs to make sure the app runs, it does this by leveraging the Operating System. A resource the CPU can call could be a device driver for example. Say you are playing a game and need to talk to other players, the CPU calls the audio device driver which processes the audio you speak into your headset and
Use-Case of Intel Optane Persistent Memory for Gaming
If your device shuts down, the Intel Optane has the ability to store that particular state where the game froze.
The Intel Optane Persistent Memory product comprises a DRAM Memory + Persistent Memory.
The Persistent Memory is just your regular SSD capability but with a fast access memory speed of your regular DRAM. Which is so cool!
That Persistent Memory can possess up to 512GB of RAM. That means more space to store files that require fast access by the processor. Hence, the processor does not need to access the hard disk often.
This was from Intel’s previous presentation stating the storage space. Not the recent one from TFD25
Use-Case of the Intel Memory on AI Workload for Social Media Recommendation
During the TFD25 session, Meta, one of Intel’s storage solution consumers had a presentation describing their implementation of Intel Optane SSD storage for their AI workloads.
I have distinct taste, and there is this Instagram account which I have been admiring. He creates collages out of his favourite musicians.
Trust me, it’ll be so cool to find maybe one or two pages that do the same thing.
The above image describes a new page that does the same thing as @ian__woods. You see a notification stating “because you liked a post from @ian__woods”
Edited on October 8, 2022
A glimpse of how Tik Toks recommendations system works
Deep Learning for Recommendation (DLRM)
The above use-case leverages the DLRM algorithm and is being applied by Meta, to build their Recommendation System.
The algorithm leverages a mathematical concept called Matrix Factorization.
For example:
A = An Instagram account
B = Another Instagram account it interacts with
A * B = C
C = A similar page to the one an Instagram account interacts with very often
In reality, the data, are actually more in-depth than I mentioned above. The data are classified into two features called Dense Feature (age, no of minutes spent online, etc.) and Sparse Feature (Gender, Locations, etc.). I won’t be going into it.
The Recommendation System can achieve the above using this method:
Look at the similarities of my past behaviour on Instagram
Liking many posts from @ian__woods
Liking other graphic designs post, I guess
Make recommendations to me based on my last activities
Achieving the above requires a storage medium that can provide low latency and bandwidth.
Source: Tech Field Day
It also requires a lot of storage space and power to train the model. And I’m not joking when I tell you it takes about 1.2 Terabytes of datasets.
Source: Tech Field Day
This is because the model has to produce Matrices of results, with continuous feedback from the user, until the user finds the perfect match.
I found my perfect match when I came across the @synnstudios account.
According to a research paper presented by Ehsan Ardestani from Meta, it was observed that there was a 29% of performance/power saved using Optane SSDs.
It’s pretty neat how Intel’s new solution comes in pretty neat for this algorithm. I’m certain many Enterprises like Meta had huge issues like power and performance from using the old type of SSDs in their Data Center.
Recommendation Systems have a huge high accuracy requirement because better accuracy equals an increase in revenue for the organization. This new storage solution from Intel is just lovely.