The Environmental Shadow Of Our Digital Footprint
Globally, data centers consume over 190 TWh of electrical power, more than the entirety of the United Kingdom, which uses only 166 TWh, and equates to 90,8 kt/h of CO2 produced. Data centers are one of the few specific things that consume a measurable amount of global electrical power. Data centers use all that power for one thing, data. And if you excuse the unintended pun, we have a lot of data to draw on about data.
As it turns out, every gigabyte of data is estimated to use about 60 Wh of electrical power on average. This means the 190 TWh of power used by data centers should roughly equal 3 Exabytes, which checks out with the estimated amount of data on the Internet and private networks. This, of course, does not include the billions of gallons of water per hour used to cool data centers or any of the other environmental and health impacts posed by data centers. These are just a few reasons why regulators have targeted data centers across the globe with new regulations like 2030 Data Center Carbon Neutrality.
As it stands, it would be impossible to directly power all of the world’s data centers using solely renewable energy sources, let alone carbon-neutral ones. Only the smallest edge and regional data centers could consider this as a viable option. Ironically for reasons we will talk about later, it’s a density issue; how much reliably continuous power can we produce from within an economical range to transport it to the data center?
So this leaves us with one alternative, reduce or at least slow the growth of the demands of data centers. This does not just help us meet new carbon goals, mandates, and regulations but will also dramatically lower the TCO of running the data center and, in turn, any services hosted out of it. As an added benefit, as we have all come to know, even “Green” tech has a major environmental impact reinforcing the reduction of power consumption as the only real solution.
But to do this, we must first understand what the data we are creating, storing, and using is. A hint to this answer is the rate at which it is growing. Based on how much storage is being purchased, we know data is growing at a rate of 10 times a year. While humans produce the majority of individual data objects like documents, spreadsheets, and art, if this were all human-generated data, we would need a computer using a population growth rate of 2x and a production increase per user of 5x. Something we know is not the case, even with all the duplication of cute cat videos across social media sites.
So this leaves us with machine-generated data. This would be everything from IoT telemetry to event data like stock trades or credit card swipes. When we consider people adding multiple “smart” devices in their lives each year combined with new “smart” grids, buildings, transit, and cities feeding data-hungry AI trainers and ML models, we can easily see a ten times growth pattern. And a significant amount of this type of data is what we refer to as Hyperscale Datasets, each often thousands, if not millions, of times larger than a two-hour high-definition move.
This is the type of persistent data FermiHDI focuses on with major impacts on the power and cooling costs of Hyperscale Datasets. Specifically designed for Hyperscale Datasets, FermiHDI can easily reach over 75 million operations per second on low-end hardware or cloud instances. Such unheard-of performance on persistent data dramatically lowers the total number of servers required while reducing the power draw and heat generation of the remaining ones.
But with FermiHDI, we can take this a step further. FermiHDI’s patented logic has been demonstrated to be portable to Application Specific Compute (ASC) systems and is the only system capable of doing so. This opens an entirely new realm. Running on an ASC platform, FermiHDI can reach 344 million operations per second per node, exceeding the performance of not only every current persistent datastore but almost every in-memory datastore too.
Moreover, ASC platforms are far more compact and draw a fraction of the power of traditional general compute x86 and ARM-based systems. In fact, FermiHDI running on ASC uses less than 2% of the space, power, and cooling at 1.3% of the five-year TCO of leading legacy solutions. With the elimination of an extreme number of servers or cloud instances and the drastic reduction in resource consumption by the ones in use, FermiHDI achieves an energy cost of less than 1 Watt per gigabyte.
While we are sure most of the displaced servers will end up back in the pool ready to run some other application, using FermiHD will have a major environmental and bottom-line impact as it slows data center growth. If you are interested in learning more about FermiHDI, please contact us.
