Sustainable systems for the Age of AI

Green computing
Green computing

[This is the second part of the “Pure’s all-flash storage platforms: engineered for sustainability ” article.]

SPONSORED FEATURE: We can now turn to system-level, data centre electricity consumption issues, followed by eWaste and embedded carbon considerations.

The AI boom, led by generative AI relies on GPU servers for AI training, and they consume a lot of electricity.

McMullan tells us: ”In order to feed and sustain a large GPU cluster, you have to have investment in the right place. You’ve seen with the launch of Nvidia’s Blackwell’s systems that they’re offering a SuperPod that draws 125 kilowatts per rack, which is quite frankly terrifying for those of us who know how data centres operate and scale.”

He said: ”The Uptime Institute published a survey fairly recently in which they surveyed hundreds of their customers, asking “what is the power draw to sustain a rack and what is the trend over the last couple of years? Most people are in the 20 to 30 kilowatt per rack range with a really specialised few in the 50s.”

A SuperPod can run from 10 up to 44 or even more racks, meaning 1.25 to 5.5 megawatts per system.

McMullan thinks that: “In standard datacentres if you’re going to sustain 125 kilowatts per rack, you have to turn off pretty much everything else in the row. For most people in today’s data centres, when you have the most power-efficient system and the smallest footprint then saving on storage power helps you run more GPUs.”

“As the data centre gets bigger and bigger, your benefits with Pure become higher and higher.”

Customers can actually see this as “We actually report live power draw, CO2 footprint, and all of the green statistics you’d expect in terms of thermal output for each device. A customer can look at that, and see, this is what it is. And ask, how am I monitoring it, measuring it? How can I change that?”

He adds: “We also have other technologies in the pipeline, to reduce our power budget for Pure even further in terms of using AI in the way that we manage our workloads within the array.”

Embodied Carbon

A Pure Storage Life Cycle Assessment (LCA) of a FlashArray//X70 stated: “The total GHG emissions of one Pure Storage FlashArray//X70 622TB is 4.3 MTCO2e per PBe per year and 9.9 MTCO2e per array (with 2.29 PBe capacity) per year. The top contributor to GHG emissions is electricity consumption during the operation of the array, represented as the use phase (96%).”

That means just 4 percent is represented by the embedded carbon when the array is built, meaning the CO₂ emissions associated with materials and manufacturing processes: “What it takes to present Pure with a number of silicon chips, some bent sheet metal and some other boards and components that go around that.”

Talking about the overwhelming preponderance of array use-caused carbon emissions, McMullan said: “That’s really been one of the things that caused us to look and reflect quite deeply on how our products are used.” It’s driving the coming use of AI to monitor the power-consequences of workload placement and timing in a Pure system.

As Pure buys in storage system components: “The broader picture is; we want to make sure that, by encouraging our NAND partners to make bigger capacity chips, that cuts down the amount of embodied carbon per bit in each of those devices. That helps us as well as our CO2 goals.”

“It’s exactly the same concept as if you’ve got a system with five HDDs that are big, or a system with 50 HDDs that are small, there’s a big power draw difference. It’s exactly the same for chips as well. So, the fewer physical packages we have, the lower the embedded carbon for us.”

McMullan added: “We are committing this year to running that lifecycle assessment on a number of other devices as well, given their growth and popularity.”


Pure Storage thinks it has an advantage for its customers in the e-waste area, e-waste being discarded electrical devices. This can be hazardous when it contains toxic materials or can produce toxic chemicals when treated inappropriately.

In McMullan’s view Pure provides strong e-waste benefits to its customers. First of all, with its Evergreen business model which upgrades system components to lengthen a system’s working life. He said: “As we change controllers on systems within three years, to extend and preserve the life and utility of all those systems, we make them better for customers.” Pure’s DFMs have at least a 10 year life span compared to 4-5 years for HDDs and 6-7 years for SSDs.

“We also offer a lifetime warranty on the DFMs, so we don’t ever say, “That system is too old.””

There is a system weight perspective: “If you look at weight as a reasonable proxy for e-waste, in terms of one system versus another, nobody puts in more material in than they have to. They all put in the minimum including us. An HDD system will weigh heading on towards a metric tonne. The big Pure systems are much less; I can still lift up and put one in a rack.”

“They’re all nicely packaged around that. We’ve been very focussed on getting rid of plastics. We’ve changed our bezels from plastic to aluminium now, all recycled.”

“We’re very much in the same mindset as people like Sir Dave Brailsford, in terms of incremental gains, and 1 percent compounded over a year still gives you an awful big improvement. We don’t leave anything to chance to leave anything behind on that.”

“The big reason for moving towards the DFMs was reliability,  as well as all the benefits we’ve talked about. Right now, our annualised failure rate is way less than 1 percent.” Fewer failed devices means lower e-waste amounts.

“If you look at our controller strategy, we always ship the latest and greatest Intel CPU. We’re currently shipping Sapphire Rapids, in all of our FlashArray systems. And that’s probably one or in many cases, three generations ahead of our competitors.”

Because these CPUs are so powerful they can be re-purposed as they age for less mission-critical systems: ”The subscription business at Pure is one of the fastest growing parts of it. We, at this point, have built an incredibly sophisticated circular economy in that many of the controllers that we take from customers after the three year interval, are being broken down, recycled, repurposed, reused and going into our Evergreen//One public or private cloud capability instead.”

“The metrics we have thus far show we can at least double if not triple the useful lifetime of the controllers that we ship to customers.”

How does this compare to controllers from alternate suppliers’ systems, with 3- or 5-year warranty? When their warranty period ends?

“Mostly, it goes for reprocessing broken down for precious metals. Most of it goes to landfill.”

Some of it ends up in the second-hand market: “There is a genuine grey market on eBay for secondhand controllers. But most of it gets broken down and goes in landfill.”

The e-waste aspect of array provision does not typically appear on RFPs. McMullan said: “It’s something we’ve been advocating for with many of our bigger customers who are considered leaders in the sustainability space. And the same thing applies to obvious elements like clauses in RFPs around green credentials or ESG scoring from external third parties. But right now we’re not quite there.”

“So we’ve taken a conscious decision ourselves in that, when we’re offering customers a subscription offering, and as a service offering, we are offering them an SLA (Service Level Agreement), not a specific product. What you will find is that customers who are wanting a medium tier subscription, will have a high performance controller from two years ago, for example, being used to deliver that subscription.”

That’s an advantage to Pure financially, but it’s also an advantage in the environmental sense, because a portion of landfill isn’t being used up by that recycled, re-purposed controller.

McMullan again: “It’s a massive structural advantage for us as a company. But there are overheads and impacts. Because we have to maintain support for all the controllers, with the right features and quality and security, patching and all of those things. So it’s not like free controllers; we have to do a lot of engineering work around that. There are fluctuations in this, but we have seen quarters where up to 60 percent of the controller ships in our subscription model have been recycled.”

“That’s what happens in the public cloud. When you lease a server, you don’t get a new one, you get what’s available on that day. And you’re leasing an amount of server performance capacity or storage capacity. You’re not leasing a specific device or a specific CPU.”

“If you are entering into a cloud consumption model with Pure, you’re buying a capability. Quite honestly, the two year old controllers that we’ve taken from another customer are still faster and better than most of our competitors’ new thing anyway. So it’s not a concern from our perspective at all.”

Have other suppliers chosen to go down this route? “it’s a structural engineering decision, based on ambitious ESG goals as well as an engineering capability, which is second to none – without hyperbole.”

Drive reprogrammability

“Evergreen customers have the ability to do a capacity trade-in. If they’ve got a chassis that’s full, we’ll allow them to trade in a proportion of that capacity for bigger drives, and offset the costs against that. When we do that, those drives are erased and go back to our subscription cloud as well.”

There’s more that could come: “Because we make our own drives, we are acutely aware of exactly how each drive functions. And we monitor the wear rate on every drive down to a forensic level. If, for example, we were to take let’s say a 75 TB drive which is approaching the end of its wear life in a couple of years time. We have a technology to take that 75 TB drive and reprogram it as a 36 TB one.”

“It becomes a 36 TB drive with rather than more over provisioning than normal to manage the wear rate. That’s something that we have the code ready for. But we’re not going to do that yet. When it’s worth it, then it’s something we will do.”

He thinks the IT industry should adopt more recycling aims: “I want to try and break the mindset of where we sit. Our industry is  kind of unusual in terms of this. If you look across industry ecosystems, you will see clothing manufacturers, car manufacturers, shoe manufacturers, etc., telling you how much recycled componentry is in a new product that you’ve just bought. Whether that’s recycled steel on a car, whether that’s cloth which has gone into your new Nike trainers or whatever. Why shouldn’t we do that?”

Overall, McMullan says: “Looking back over where we spend our sustainability efforts, we’ve consciously designed our way into not just an incredibly innovative platform, but also one which has genuine sustainability credentials, structural benefits across every dimension, from power draw, to cooling to e-waste. We’re also showing I would say, pretty strong leadership in this part of the industry as to how we could actually do something  aligned with the Paris Agreement goals and which is good for the planet, rather than buying carbon offsets or committing to plant some trees. The growth in AI systems is driving huge capacity demands and we have a collective responsibility to push innovation in every dimension to cope with the increased pressure on sustainability that AI brings in the next decade.”

Note. This article was sponsored by Pure Storage.