Security Cloud Privacy Tech
Peter DeSantis' Keynote at AWS re:Invent 2021

Peter DeSantis' Keynote at AWS re:Invent 2021

AWS re:Invent is always THE event in cloud. Petere DeSantis’ keynote provided a peek behind the curtain of the technology that drive AWS itself.

This πŸ‘‡ is the Twitter thread of my coverage of the keynote…

Tweet 1/104 πŸ‘‡ Next tweet

rolling now with the keynote from Peter DeSantis, SVP Utility Computing and Apps, @awscloud


Tweet 2/104 πŸ‘‡ Next tweet πŸ‘† Start

this used to be Monday or Tuesday Night Live and is always one of my favourite talks of the week


Tweet 3/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter giving his perspective on the start of the @awscloud


Tweet 4/104 πŸ‘‡ Next tweet πŸ‘† Start

big focus from the team before the launch was on the key word β€œElastic”

<< I’m willing to be it still is!


Tweet 5/104 πŸ‘‡ Next tweet πŸ‘† Start

early memories from the original EC2 team. 7/10 of the team are still @awscloud


Tweet 6/104 πŸ‘‡ Next tweet πŸ‘† Start

email from @JeffBezos to the EC2 team in the early days. love the focus on the user experience


Tweet 7/104 πŸ‘‡ Next tweet πŸ‘† Start

key areas of focus for EVERYTHING @awscloud


Tweet 8/104 πŸ‘‡ Next tweet πŸ‘† Start

looks like we’re starting the night by looking at storage


Tweet 9/104 πŸ‘‡ Next tweet πŸ‘† Start

remember, this is (usually) not a feature/service launch keynote. it’s typically a look behind the scenes @awscloud. lots of amazing technology and engineering challenges


Tweet 10/104 πŸ‘‡ Next tweet πŸ‘† Start

remember when @awscloud looked like this?


Tweet 11/104 πŸ‘‡ Next tweet πŸ‘† Start

lots of stuff build out from S3


Tweet 12/104 πŸ‘‡ Next tweet πŸ‘† Start

for a longer look back at @awscloud’s early days. be sure to check out @jeffbarr’s session, β€œ15 years of AWS with Jeff Barr”


Tweet 13/104 πŸ‘‡ Next tweet πŸ‘† Start

we’re going right down to the fundamentals. looking at old school(ish) hard drives


Tweet 14/104 πŸ‘‡ Next tweet πŸ‘† Start

lots of mechanical engineering packed into these types of drives. Peter used a crazy airplane analogy that I won’t even try to summarize

lots of improvements but the mechanics are pretty steady


Tweet 15/104 πŸ‘‡ Next tweet πŸ‘† Start

β€œHard drives remain the best way to store large amounts of data when you need to access it immediately”, Peter << he left out, β€œβ€¦even though we don’t want that to be the case”


Tweet 16/104 πŸ‘‡ Next tweet πŸ‘† Start

β€œLet’s do some quick math”, Peter << a key indicator as to why this is always my favourite keynote


(sorry @werner)

Tweet 17/104 πŸ‘‡ Next tweet πŸ‘† Start

here’s what the math says : hard drives suck for these type of big data workloads


Tweet 18/104 πŸ‘‡ Next tweet πŸ‘† Start

the opposite view for a large but low usage workload


Tweet 19/104 πŸ‘‡ Next tweet πŸ‘† Start

basically the individual workload doesn’t align to the hard drive mechanics. you need to aggregate these workloads in order to smooth things out


Tweet 20/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter’s currently walking through how @awscloud distributes customer S3 workloads across the physical backend of the service

it’s the only way the math adds up


Tweet 21/104 πŸ‘‡ Next tweet πŸ‘† Start

β€œIf you want to innovate at scale, you need to move quickly. AND you need to do it safely”, Peter DeSantis, @awscloud


Tweet 22/104 πŸ‘‡ Next tweet πŸ‘† Start

talking about how @awscloud approaches testing S3 now


Tweet 23/104 πŸ‘‡ Next tweet πŸ‘† Start

I love this type of deep dive. especially the contrast in the complexity behind the scenes of S3 for something as simple as a bucket/key #ux


Tweet 24/104 πŸ‘‡ Next tweet πŸ‘† Start

for testing, @awscloud is leveraging automated reasoning more and more. it’s the only way to truly test things out

more at


Tweet 25/104 πŸ‘‡ Next tweet πŸ‘† Start

excellent timing by the @AmazonScience team, β€œA gentle introduction to automated reasoning” << just published


Tweet 26/104 πŸ‘‡ Next tweet πŸ‘† Start

formal tools are great. they have a very high bar for correctness

…but at the cost of speed and reducing the available talent pool


Tweet 27/104 πŸ‘‡ Next tweet πŸ‘† Start

this isn’t a great choice to face. at @awscloud’s scale, they deal with it all of the time

…so they developed a new approach, β€œLightweight Formal Methods”


Tweet 28/104 πŸ‘‡ Next tweet πŸ‘† Start

here’s a great related paper from @AmazonScience , β€œUsing lightweight formal methods to validate a key-value storage node in Amazon S3”


Tweet 29/104 πŸ‘‡ Next tweet πŸ‘† Start

lots of advantages to this approach…biggest challenge would be the initial design to allow for the application of these formal methods when required


Tweet 30/104 πŸ‘‡ Next tweet πŸ‘† Start

ohhh, I beat Peter to the punch πŸ‘Š



Tweet 31/104 πŸ‘‡ Next tweet πŸ‘† Start

first customer story up, this one’s from @Adobe


Tweet 32/104 πŸ‘‡ Next tweet πŸ‘† Start

. @adobe’s intro covered how they are massive scale in the cloud, not turning to storage


Tweet 33/104 πŸ‘‡ Next tweet πŸ‘† Start

talking about photos and how we stored physical photos


Tweet 34/104 πŸ‘‡ Next tweet πŸ‘† Start

digital is a similar challenge but at a larger scale

how many photos & videos do you have?


Tweet 35/104 πŸ‘‡ Next tweet πŸ‘† Start

next year β€œwe” (not sure who that is, all of us?), we’ll generate 1,500,000,000,000 photos


Tweet 36/104 πŸ‘‡ Next tweet πŸ‘† Start

. @adobe turning to the two primary storage use cases for storing photos and videos

β€œI want it now"

β€œI want it sometime, whenever really”


Tweet 37/104 πŸ‘‡ Next tweet πŸ‘† Start

the obvious questions about this story form @adobe is why do I only get 100GB of storage with Creative Cloud then?

<< feature request == let me use my own @awscloud account of Creative Cloud storage & sharing


Tweet 38/104 πŸ‘‡ Next tweet πŸ‘† Start

πŸ‘† s/of/for/

Tweet 39/104 πŸ‘‡ Next tweet πŸ‘† Start

learn more about @adobe Sensei at


Tweet 40/104 πŸ‘‡ Next tweet πŸ‘† Start

. @adobe is obviously the reference customer for Amazon S3 Glacier Instance Retrieval

more on the @awscloud feature at


Tweet 41/104 πŸ‘‡ Next tweet πŸ‘† Start

πŸ‘† it snuck under the radar for most people but it could be quite a game changer


Tweet 42/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter moving on to block storage types now


Tweet 43/104 πŸ‘‡ Next tweet πŸ‘† Start

now diving into SSD technology


Tweet 44/104 πŸ‘‡ Next tweet πŸ‘† Start

SSD remove most of the mechanical challenges but they have their own issues…like the limitations of flash storage


Tweet 45/104 πŸ‘‡ Next tweet πŸ‘† Start

primary the paging system & the lifetime of those pages


Tweet 46/104 πŸ‘‡ Next tweet πŸ‘† Start

as Peter is delivering this section of the talk, I realize I know way too much of the low level functionality of both HDD and SDD.

this is the downside of nerd πŸ€“ life. lots of obscure knowledge

…but I wouldn’t have it any other way


Tweet 47/104 πŸ‘‡ Next tweet πŸ‘† Start

of course there are different issues at hyper scale that @awscloud has discovered with low level SSD tech


Tweet 48/104 πŸ‘‡ Next tweet πŸ‘† Start

small inconsistencies add up fast at hyper scale. things we would never notice even with hundreds of systems, they see regularly


Tweet 49/104 πŸ‘‡ Next tweet πŸ‘† Start

. @awscloud Nitro smooths a lot of this out

more on Nitro at


Tweet 50/104 πŸ‘‡ Next tweet πŸ‘† Start

…of course they built their own Nitro SSD


Tweet 51/104 πŸ‘‡ Next tweet πŸ‘† Start

btw, @QuinnyPig it live tweeting this keynote as well. for his take (😈) head over to


Tweet 52/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter citing the performance delivered by the custom SSD approach via @awscloud Nitro


Tweet 53/104 πŸ‘‡ Next tweet πŸ‘† Start

Tweet 54/104 πŸ‘‡ Next tweet πŸ‘† Start

on to @awscloud Graviton now...


Tweet 55/104 πŸ‘‡ Next tweet πŸ‘† Start

more on Graviton at


Tweet 56/104 πŸ‘‡ Next tweet πŸ‘† Start

. @awslcoud is pushing Graviton everywhere. and why not? more bang for your buck in most cases w/cloud native designs


Tweet 57/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter touches on the Graviton3 update


Tweet 58/104 πŸ‘‡ Next tweet πŸ‘† Start

have we seen these stats for Graviton3 yet?

…not that we should focus on them according to Peter (and I agree)


Tweet 59/104 πŸ‘‡ Next tweet πŸ‘† Start

I like Apple’s performance per watt approach. which @awscloud has followed as well


Tweet 60/104 πŸ‘‡ Next tweet πŸ‘† Start

more on Dennard scaling at


Tweet 61/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter talking about power consumption now. it’s a critical attribute for all of us, mind boggling at @awscloud’s hyper scale


Tweet 62/104 πŸ‘‡ Next tweet πŸ‘† Start

How do we efficiently increase performance of a Graviton core?

Make the core WIDER


Tweet 63/104 πŸ‘‡ Next tweet πŸ‘† Start

there have been lots of improvements in silicon design over the past decade. the easy way to think about it is for year all we worried about was the size of the engine

we finally realized that the tires, chassis, fuel, and other attributes matter as well…


Tweet 64/104 πŸ‘‡ Next tweet πŸ‘† Start

…that means more cores, wider memory lanes, direct access to storage, etc.

this is why we’ve seen massive boosts in performance in things like @apple’s M1 and @awscloud Graviton3


Tweet 65/104 πŸ‘‡ Next tweet πŸ‘† Start

another customer story up now, this time it’s Fannie Mae


Tweet 66/104 πŸ‘‡ Next tweet πŸ‘† Start

Kimberly Johnson, EVP & COO of @FannieMae given the scale and scope of their operations

tl:dr = it’s big


Tweet 67/104 πŸ‘‡ Next tweet πŸ‘† Start

ooohhh, @FannieMae built out an HPC solution @awscloud Lambda. would love to get a deep dive on that one…


Tweet 68/104 πŸ‘‡ Next tweet πŸ‘† Start

very interesting solution from @fanniemae. they leverage some key @awscloud data services to expand credit score analysis to include rent payments…regardless of how they were made


Tweet 69/104 πŸ‘‡ Next tweet πŸ‘† Start

now talking about evaluating risk to housing due to climate change. again another massive data problem, taking unstructured data into structure data for analysis and prediction


Tweet 70/104 πŸ‘‡ Next tweet πŸ‘† Start

good customer story from @fanniemae. I think they struck the right depth to solve solutions to unique problems at a scale that’s only really possible in the cloud


Tweet 71/104 πŸ‘‡ Next tweet πŸ‘† Start

on to #ml now. I swear this is slide is even more packed than @SwamiSivasubram from earlier today


Tweet 72/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter’s breaking #ml down into two big stages:
  • training
  • inference

you need different infrastructure & tools for both


Tweet 73/104 πŸ‘‡ Next tweet πŸ‘† Start

diving into inference now...


Tweet 74/104 πŸ‘‡ Next tweet πŸ‘† Start

more on @awscloud’s custom silicon, Inferentia, for #ml inference at


Tweet 75/104 πŸ‘‡ Next tweet πŸ‘† Start

we’ll get there in a minute (I’m sure) but here’s more on @awscloud Trainium, the custom chip for #ml training


Tweet 76/104 πŸ‘‡ Next tweet πŸ‘† Start

back to inference, @awscloud is working to improve inference work on general purpose CPUs as well. sometimes you just need to run the process where the workload is w/out the custom chips


Tweet 77/104 πŸ‘‡ Next tweet πŸ‘† Start

…or at least _those_ custom chips


Tweet 78/104 πŸ‘‡ Next tweet πŸ‘† Start

sticker stats for Trainium…


Tweet 79/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter says the name β€œTrainium” implies training #ml models. for most, it’s probably πŸ‘‡


Tweet 80/104 πŸ‘‡ Next tweet πŸ‘† Start

β€œ is normal in a math party” << f--k yeah, Peter


Tweet 81/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter diving into the math of #ml training runs. you need it to figure out how to scale out/up a workload to get the job done


Tweet 82/104 πŸ‘‡ Next tweet πŸ‘† Start

GPT-3 is slightly larger than BERT-Large #nlp #ml models


Tweet 83/104 πŸ‘‡ Next tweet πŸ‘† Start

these models are massive. they create their own computer science problems


Tweet 84/104 πŸ‘‡ Next tweet πŸ‘† Start

more on how to tackle these challenges the @awscloud Well-Architected Framework, Machine Learning Lens

that’s available at


Tweet 85/104 πŸ‘‡ Next tweet πŸ‘† Start

more math party analogy/reference!


Tweet 86/104 πŸ‘‡ Next tweet πŸ‘† Start

…lots of room left in that graph...


Tweet 87/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter talking about the biggest @awscloud EC2 instance types for #ml training


Tweet 88/104 πŸ‘‡ Next tweet πŸ‘† Start

the network optimized version of the Trn1 instance type will have 1600 Gbps network bandwidth << fantastic!


Tweet 89/104 πŸ‘‡ Next tweet πŸ‘† Start

sizing your training pool is a critical decision. not only will it impact how LONG you need to wait but also how MUCH you pay

read that Well-Architected Lens, please!


Tweet 90/104 πŸ‘‡ Next tweet πŸ‘† Start

Tranium uses a technique called β€œSystolic Array Manipulation”

going to have to dig into this one later on:


Tweet 91/104 πŸ‘‡ Next tweet πŸ‘† Start

Tranium provide 16 fully programmable inline data processors. this allows further tuning for each training workload

<< still doesn’t fix the name though 🀦


Tweet 92/104 πŸ‘‡ Next tweet πŸ‘† Start

now we’re talking about how to round numbers. I ❀️ this stuff

super niche but fascinating


Tweet 93/104 πŸ‘‡ Next tweet πŸ‘† Start

of course Trainium supports stochastic rounding directly in hardware. because, why not?


Tweet 94/104 πŸ‘‡ Next tweet πŸ‘† Start

more on the @awscloud Neuron SDK at

<< it abstracts a lot of what Peter just covered away, so you just get the benefits


Tweet 95/104 πŸ‘‡ Next tweet πŸ‘† Start

Peter is going to wrap up with sustainability and the climate pledge

more on the pledge (now over 200 signatories) at


Tweet 96/104 πŸ‘‡ Next tweet πŸ‘† Start

. @awscloud is always looking to improve efficiency in every aspect of our infrastructure << we’ve heard various achievements and efforts year over year from them. it’s nice to see


Tweet 97/104 πŸ‘‡ Next tweet πŸ‘† Start

that’s a big commitment…and they are on a path to achieve it 5 years early (originally target was 2030)


Tweet 98/104 πŸ‘‡ Next tweet πŸ‘† Start


Tweet 99/104 πŸ‘‡ Next tweet πŸ‘† Start

storage isn’t the only problem. moving that power around is a big problem as well


Tweet 100/104 πŸ‘‡ Next tweet πŸ‘† Start

here’s a great research piece from @voxdotcom


Tweet 101/104 πŸ‘‡ Next tweet πŸ‘† Start

. @mikegchambers called this the other night in our day one recap!


Tweet 102/104 πŸ‘‡ Next tweet πŸ‘† Start

you can watch that recap on @mikegchambers excellent YouTube channel at


Tweet 103/104 πŸ‘‡ Next tweet πŸ‘† Start

here are some other sessions at @awscloud #reinvent that focus on sustainability. be sure to check them out


Tweet 104/104 πŸ‘‡ Next tweet πŸ‘† Start

..and that’s a wrap from Peter. not nearly as broad as previous years but still a great deep dive into things like @awscloud custom silicon

/🧡 #reinvent

More Content