Analog Computing is GENIUS - Here's Why!

in my 40 years on Earth I've seen almost everything go digital I remember watching Terminator 2 on VHS in the 9s that was analog the 4K screen you're watching this YouTube video on today that's digital it's the same with vinyl records and cassette Walkman both analog and now digital smartphones and MP3s most of our world has completely gone digital so when I heard that there may be a analog Computing revolution in the making I was pretty confused and skeptical digital is so vastly Superior it just didn't make any sense so I started Di and I was blown away with what I discovered it's a tale of compute power efficiency AI neural networks betrayal and revenge okay maybe not the Betrayal and revenge but this story is pretty amazing so let's figure this out together I'm Ricky and this is Tu bit to Vinci this video is brought to you by Robo Rock let's start with what these terms actually mean analog or analog as it's been shortened just means having agreement or relation with something else like an analogy in story storytelling take for example old film cameras these were analog because they were an analog representation of the real world in fact early cameras were nothing more than a box with a circular cutout allowing light in and leaving an etched image that was analogous to the real world maybe a bit blurry and fuzzy black and white or upside down but analogous nonetheless the most famous Analog Devices that we romanticized about today are vinyl record plays this analog device stores sound analogously in grooves on a vinyl record many love the rich character of records for this reason because any attempt to digitize sound comes with a trade-off imagine this Soundwave which can pretty much be represented as grooves on vinyl as it is but a digital recording has to convert this into ones and zeros early digital recordings were low quality as low as 8 bit or a string of eight ones and zeros with a maximum combination of 256 values CDs are 16bit with 65,536 possible values the sample rate is how many many samples are recorded per second 44 khz is common and high quality audio recording can be 96 khz so with the higher sampling rate and higher bit depth you can see that the digital approximation of the analog Soundwave gets closer and closer it's why today's high quality digital music on platforms like Spotify or apple music sound so much better but some still prefer good old vinyl s off in the comments below where do you come down on the great audio debate digital or vinyl and this is a great example of how digital computers work everything is a series of ones and zeros and grouped together it can make some beautiful music sure digital is an approximation but it's very precise in fact if you download the same song Twice you're going to get identical carbon copies whereas vinyl records are all going to be a little bit different close of course but no two will be exactly the same digital computers are general purpose this MacBook can play solitire or run blender or Final Cut Pro generally analog computers are much more bespoke single use case devices made to just do one thing digital computers are exact and precise with high repeatability analog computers use voltages resistors and Hardware that will all vary slightly leading to close but not exactly accurate or repeatable results but analog computers do have some benefits otherwise we probably wouldn't be making this video right analog computers are actually quite powerful for a specific task very fast and very energy efficient and I think I'm starting to see where this is headed I made a video recently on how the internet and all the data centers in the world consume roughly 1% of all the world's electricity and with the rise of AI and neural network training that's about to explode and if you're thinking AI probably has something to do with this you'd be correct artificial neuron networks are one of the hottest things in Tech right now they're actually nothing new but they've matured to such a level recently as to be pretty damn mindblowing AI is slowly making so many things possible from better recommendations on Netflix to detecting burglars on on your camera doorbell but my favorite use case has to be smart machines like our sponsor this week Robo Rock and this the S8 Pro Ultra you guys know I value my time above all I don't go grocery shopping I have it delivered and now I don't even vacuum or mop my floors the S8 Pro Ultra has 6,000 pascals of suction power that's more powerful than my regular vacuum one of the biggest advancements thanks to AI is the reactive 3D obstacle avoidance which identifies lots of common household items and avoids them its Duo roller Riser brush do does a remarkable job of picking up everything from dust and pet hair to food waste and cereal and of course with its amazing Robo dock it automatically empties the dust bin so you never have to bother but the amazing part is it'll even mop your floors with its vibrin 2.0 mopping Tech it'll vibrate up to 3,000 times per minute for amazing mopping performance when it encounters a rugg or carpet it can lift the mopping head and keep vacuuming automatically I cannot tell you how much I love the S8 Pro Ultra let it get to work keeping your house clean so you can get back back to doing the things you love use our links in description to check out the robo Rock S8 Pro Ultra and win back your time today huge thanks to Robo Rock and you for supporting the show let's do a quick rundown of what a neural network is artificial just means that the architecture underpinnings are designed to round the idea of the neurons in our brain our brain has over 86 billion neurons that connect through synapses analogously in an artificial neural network the nodes are interconnected to other nodes to form a network keeping things pretty high level you can think of a neural network as a series of layers read from left to right the First Column is the input layer this can be as simple as one input for Simplicity or hundreds or thousands imagine each input corresponding to the pixel of an image for a self-driving computer for example the next columns are hidden layers if you've ever heard the term deep neural network that just means a lot of hidden layers where a lot of manipulation is happening finally you have the output layer where the nodes finally give answers the magic happens when you break down the operations at play ultimately it's matrix multiplication so we can assign a value to each input how many inputs you need is a function of what you're trying to do but I think the easiest way to visualize this is an image recognition AI for example what's so genius about neural networks is how you train them if you want a neural network to identify shapes for example you start with a training set of circles and triangles and squares the key is at the weight value you assign to each node the value for the next node is equal to each input node times the weight plus the node times the weight of every other node this is the matrix multiplication operation you start with a random series of numbers for the weight and wait for an incorrect answer like the model claiming a square as a circle when this happens you assign the values to the weight influencing the future output do this over and over again with more and more data and models start getting better and better at predicting what actually is a rectangle and what's a circle as you can imagine the number of inputs can get really insane and so can the number of internal hidden layers much like the sampling rate or bit depth for higher quality audio recordings this is how you get more robust neural networks but all comes at the cost of a really high level of compute power I've given really broad Strokes here but if you want a deeper dive on neural networks or if you're curious about more sign off in the comments below and we can make future episodes for digital computers the fastest means of storing data is memory via the memory bus so larger complex neural networks can have massive amounts of data particularly the weight we discussed in memory this has some challenges but how does an analog computer perform the same task let's think about solar panels something I think about a lot connect two solar panels in parallel and the current adds up so addition is as simple as combining different currents which is something anog computers are very good at but remember that weight value that's multiplied to each node before adding them up so how could an analog computer perform multiplication while keeping our electrical engineering hats on the easiest way would be be adding resistance Ohm's law tells us that voltage equals IR or you can reorganize it in a different way and think of it as current equals voltage divided by resistance so if we set the node's value to current and pick a resistor to match the weight value we can measure the current and figure out the multiplied value doing math with an analog computer it's pretty amazing right but we have some problems it wouldn't be practical to change thousands of resistors right imagine if you wanted to update your neural network with new weight data that you just trained you'd have to change out thousands of resistors and that's not practical and this is why digital computers are so amazing as you can imagine AI companies are constantly training their models and getting better and better data prediction so it's critical that these machines can be updated more easily so to solve this problem analog trip startup Mythic has some pretty clever Solutions I just came across a video from veritasium one of my favorite YouTube channels that did an onsite with Mythic I'll put a link in the description but the key is how data or variables like weight values is stored in a digital computer it's stored in memory each weight value can be stored in arrays or arrays of arrays and when the calculation is needed the digital CPU pulls up that value in memory the problem with memory though is that it needs to be powered up to retain the value think about your smartphone versus an eink reader screen the smartphone needs power constantly to keep powered on if the battery dies your screen goes black but with the Inc electricity is used to permanently store values on the screen and even if the device dies it'll keep showing the last created image and that's the key so how does Mythic solve this now this is pure genius from an engineering perspective they use the same SSD flash drivve that you have with your SD cards or hard drives like in this MacBook except they use it in an entirely different way hard drives are digital devices so each storage cell usually holds either a one or a zero that makes up a larger data packet typically in flash storage a positive voltage is applied to the control gate which attracts electrons through an isolating barrier into the floating gate unlike memory this is permanent so when the voltage is removed these electrons stay in the floating gate this prevents current from flowing and registers as a zero on the digital domain this is why your data on your laptop or smartphone persists even if you power it off or the battery dies what's genius about mythics approach is that instead of treating this as a switch either on or off one or zero they apply only a certain amount of voltage to only move a certain number of electrons the higher the number of electrons the higher the resistance kind of like the physical resistors in our analog computers earlier but now you can change its value by resetting and applying a different number of electrons by changing the values on the gates back to M's law the current I is equal to V / R which is a division operation but it's a same as V * 1 / r 1 / R so essentially by setting the resistance distance of the gate and measuring the current we'd get the desired multiplication operation add the values from nodes with different current values times resistance the weight value and you just did neural network math using an analog computer and this is seriously genius because now you can write all the models weight data physically into flash storage then you don't need a power ram or anything else just measure the voltage in an analog way get the resistance and get your weight data and in this way an analog computer can do matrix multiplication almost as well as a digital computer with a fraction of the energy required Mythic claims their analog CPUs can do the same workload as a 100 watt digital CPU with as little as 3 watts of power now it's not all good news though because remember the analog computers aren't as accurate as digital the zero or one in a digital flash gate makes up a larger block of data often with a check sum to ensure the value of 7654 is exactly 7654 each and every time with the resistance approach and storing electrons and Flash gate it's not going to always be as precise but for the purposes of making an accurate identification for a visual model for example 90 to 95% accuracy might be just fine one more challenge worth noting is Runaway errors like I mentioned the current value will be close but not exact and so will the resistance value just one calculation isn't such a big deal but add these up and you can have a runaway effect imagine trying to measure your room by measuring a plank of your hardwood floor you have 100 planks and each measures to be 1T or 1 meter for our non-americans out there you might be tempted to think for saving time that if you have 100 planks and each plank is a foot you'd have a 100t room but each plank can be off by about 5% so that means after 10 planks you can be off by about half a foot or half a meter and by 100 planks you can be off by as much as 5 ft or 5 m now that's compounding runaway errors and it's a problem with the more calculations being performed Mythic solves this by converting from analog to digital along the way which adds complexity and consumes more power but it's a necessary evil the bottom line is this analog computers are really fast efficient and low power consumption but they're not as accurate and they aren't going to replace digital computers instead think about this a small bit of plastic made out of a very complex form the forms are made of very high tolerance High strength steel that way you can stamp out this part time and time again for years but the end product is this simple part that costs pennies analog computers won't cost pennies especially early on but when they reach economies of scale it's likely they will be much cheaper and abundant than digital processors the steel in this analogy would be the Super high-end GPU based supercomputer that runs on digital and handles all the training data and a slew of other programs but from that training we could load this data onto analog computers to actually run the simulations and models here are just a few potential use cases imagine an AI model running on a cheap low power analog computer that runs a camera and detects the presence of children you can mount this all around your pool and power it with a small solar cell like from a solar walkway light and it could run all day and send an alert to your smartphone if your children ever get too close to the pool being Wireless it wouldn't need to send all this data out to get processed it could do it all on site on the chip and it only has to send a notification if children are detected or imagine AI image detection sensors placed all around assembly lines looking for misalignment or other problems we did a video on the future of farming and those robots that are looking at weeds and shooting lasers to kill weeds imagine if all of that could be done on analog computers as the scale for our hunger for these sensors increase energy consumption starts to play a bigger role not to mention cost so yeah not replacing digital computers but by being the backbone that neural networks are run on we might just be looking at the beginning of an analog Computing Renaissance so what do you think are analog computers going to be a big part of the fabric of the future if you want to see more videos about any of these topics I've talked about here don't be shy sound off in the comments below and while you're down there hit the like button and subscribe all right I'm Ricky Tu D Vinci thank you so much for watching until next week check out this video next

Channel: Two Bit da Vinci

Views: 470,088

Keywords: two bit da vinci, The UNREAL Rise of Analog Computers!, analog computer, analog computer revolution, analog computing, analog computer chip, analog computers, analog cpu, machine learning with analog computers, cpu analog, analog computer processor, analog processor, analog mythic, mythic analog, Analog Computers Are Going to CHANGE The World!, Analog Computers Are About to Take OVER - Here's Why!, Analog Computing is GENIUS - Here's Why!

Id: VWn6Ixh2eDg

Length: 15min 28sec (928 seconds)

Published: Sat Dec 23 2023