Sustainability in data centers or how to shrink your carbon footprint while saving money - Rich Kenny from Interact
Cloud CommuteMarch 06, 2024
1
x
1
00:24:4022.6 MB

Sustainability in data centers or how to shrink your carbon footprint while saving money - Rich Kenny from Interact

In this episode, host Chris Engelbert sits down with Rich Kenny, Managing Director of Interact and a visiting lecturer at London South Bank University. Rich shares his extensive experience in machine learning-based environmental consultancy and his work on the circular economy.

Rich dives into Interact's innovative product that optimizes the performance of enterprise IT, focusing on energy efficiency and carbon reduction. With the world's largest dataset on server performance, Interact offers advanced recommendations for reconfiguring IT assets to achieve significant energy savings—up to 70% in many cases.

Rich also discusses the surprising impact of over-provisioning in data centers, the benefits of rethinking server configuration, and the potential of tape storage as a carbon-efficient method. He explains how businesses can achieve cost savings and sustainability by eliminating unnecessary waste in their IT infrastructure.

Tune in to learn how cutting-edge technology can drive both environmental and financial benefits for enterprises. Whether you're a data center operator, an IT professional, or just curious about the intersection of technology and sustainability, this episode offers valuable insights.

New episodes will be released every two weeks, so stay tuned for more fascinating discussions on the future of sustainable IT.

For questions, you can reach Rich at:

You can learn more about Interact at:

The Cloud Commute Podcast is presented by simplyblock (https://www.simplyblock.io)


01:00:00
We've got a small server estate

01:00:01
that's very low utilized, but

01:00:04
massively over provisioned on RAM.

01:00:05
Because someone someday, you know,

01:00:07
10 years ago bought

01:00:08
a server and went,

01:00:09
"Stick two terabytes in it." And

01:00:10
it's like, how much

01:00:11
are you using? 200 gig.

01:00:12
And it's like, you've got 10 times

01:00:13
more RAM than you

01:00:14
need, even at peak.

01:00:16
So can you just take

01:00:17
out half your RAM, please?

01:00:21
You're listening to simplyblock's

01:00:23
Cloud Commute podcast, your weekly

01:00:25
20 minute podcast about

01:00:26
cloud technologies, Kubernetes,

01:00:27
security,

01:00:28
sustainability, and more.

01:00:30
Hello, folks. Great to be probably

01:00:34
the first or second episode. I

01:00:36
think it will be the first one.

01:00:38
We'll see what's going to happen

01:00:41
in the future. I'm really happy to

01:00:43
have our first guest,

01:00:45
Richard, who's really interesting.

01:00:48
He's done a lot of things and he's

01:00:49
going to talk about that

01:00:50
in a second. But apart from that,

01:00:53
you can expect a new episode from,

01:00:56
well, about every two weeks or

01:00:58
so. We're going to figure that

01:01:00
out. It's still a new concept. So

01:01:02
we'll see. So

01:01:04
with that, thank you,

01:01:05
Richard, for being here. Really

01:01:07
happy to have you on board. And

01:01:10
maybe just start with a short

01:01:11
introduction of yourself. Yeah,

01:01:13
cool. So my name is Rich Kenny.

01:01:15
I'm the Managing

01:01:15
Director of Interact.

01:01:17
We're a machine learning based

01:01:19
environmental consultancy that

01:01:20
specializes in circular economy.

01:01:22
And I'm also a visiting lecturer

01:01:24
and research fellow at London

01:01:26
South Bank University in the

01:01:28
School of Engineering. So a bit of

01:01:30
business, bit of academia, bit of

01:01:32
research. Know a few things

01:01:35
about a few things. You know a few

01:01:37
things about a few things. That's

01:01:39
always better than most people.

01:01:40
I think it's better than knowing

01:01:41
nothing about a lot of things.

01:01:44
That's fair. I think it's good to

01:01:47
know what you don't know. That's

01:01:50
the important thing, right?

01:01:51
So you said you're doing a little

01:01:53
bit of university work, but you

01:01:54
also have a company

01:01:55
doing sustainability AI

01:01:57
management. Can you elaborate

01:02:00
a little bit on that?

01:02:03
Yeah, so we've got a product that

01:02:05
looks at the performance of

01:02:08
enterprise IT, so servers,

01:02:10
storage, networking. It's got the

01:02:12
world's largest data set behind it

01:02:13
and some very advanced

01:02:14
mathematical models and energy

01:02:15
calculations. And basically allows

01:02:17
us to look at data set

01:02:19
IT hardware and make really,

01:02:21
really good recommendations for

01:02:23
lower carbon compute,

01:02:25
reconfiguration of assets, product

01:02:27
life extension. Basically it lets

01:02:28
us holistically look at the IT

01:02:30
performance of an estate and then

01:02:32
apply very advanced techniques to

01:02:35
reduce that output.

01:02:37
So saving cost, energy and carbon

01:02:39
to do the same work better. We've

01:02:41
done about 400 data centers now

01:02:43
in the last three years and we see

01:02:45
an average of about 70% energy

01:02:47
reduction, which is also quite

01:02:48
often a 70% carbon reduction in a

01:02:50
lot of cases as well from a scope

01:02:51
two point of view. Yeah,

01:02:54
there's nothing like it on the

01:02:55
market at the moment and we've

01:02:57
been doing this for

01:02:58
as a business probably three and a

01:03:00
half, four years as a research

01:03:01
project for the

01:03:02
best part of seven.

01:03:03
So how do I have to think about

01:03:06
that? It's about a web UI that

01:03:09
shows you how much

01:03:11
energy you consume right now at a

01:03:13
specific server and it gives you a

01:03:15
recommendation on

01:03:16
like, I don't know, exchange the

01:03:18
graphics card or this

01:03:20
storage or whatever.

01:03:22
So specifically it looks at the

01:03:23
configuration and what work it's

01:03:25
capable of doing.

01:03:26
So every time you

01:03:27
have a variation in the

01:03:28
configuration of a server, it is

01:03:30
more or less

01:03:31
efficient. It does more or less

01:03:33
work per watt. So what we do is we

01:03:35
apply a massive machine learning

01:03:37
data set to any make model

01:03:39
generation configuration of any

01:03:41
type of server and we tell you how

01:03:43
much work it can do, how

01:03:44
effectively can do it, what its

01:03:45
utilization pathway looks like. So

01:03:47
it's really great to

01:03:48
be able to apply that to existing

01:03:51
data center architecture. Once

01:03:52
you've got the utilization

01:03:53
and the config and say, you could

01:03:55
do the same work you're doing with

01:03:57
2000 servers in this way,

01:03:58
with 150 servers in this way. And

01:04:01
this is how much energy that would

01:04:02
use, how much carbon that

01:04:03
will generate and how much work it

01:04:04
will do. And we can do things like

01:04:06
carbon shifting scenarios.

01:04:08
So we can take a service

01:04:09
application, say a CRM that's in

01:04:12
20 data centers

01:04:13
across 1000 machines

01:04:15
using fractional parts of it and

01:04:17
say, this service is using X

01:04:19
amount of carbon costing this much

01:04:20
energy. So basically your CRM is

01:04:23
costing X to run from an energy

01:04:24
and carbon point of view.

01:04:26
And you could consolidate that to

01:04:28
Z for example. So the ability to

01:04:31
look at service level,

01:04:33
application level and system level

01:04:34
data and then serve that service

01:04:37
more efficiently.

01:04:38
So we're not talking about sort of

01:04:40
rewriting the application because

01:04:41
that's one step lower

01:04:42
down the stack. We're talking

01:04:43
about how do you do the same work

01:04:45
more efficiently

01:04:46
and more effectively

01:04:47
by looking at the hardware itself

01:04:50
from the actual the physical asset

01:04:51
value. And it's massive low

01:04:53
hanging fruit because no one's

01:04:55
ever done this before. So we don't

01:04:57
we don't see as unusual to see

01:05:00
consolidation options of 60 plus

01:05:02
percent of just waste material. A

01:05:06
waste, a lot of it is doing the

01:05:08
same work more effectively and

01:05:10
efficiently. And that drives huge

01:05:11
sustainability based outcomes.

01:05:13
Because you're just removing stuff

01:05:15
you don't need. The transparency

01:05:16
bit is really important because

01:05:18
quite often you don't know what

01:05:19
your server can do or how it does

01:05:21
it. You're like, I bought this,

01:05:22
it's great, it's new and it must

01:05:23
be really, really effective. But

01:05:25
the actual individual

01:05:26
configuration, the interplay

01:05:28
between the CPU, RAM and the

01:05:29
storage determines

01:05:30
actually how good it

01:05:31
is at doing its job and how much

01:05:32
buying you get for your buck. And

01:05:35
you can see, you know,

01:05:36
intergenerational variance of

01:05:38
300%. So you know, we've got

01:05:40
DL360, all the

01:05:42
DL360s are pretty much

01:05:43
the same of this generation. It's

01:05:45
like, no, there's like a 300%

01:05:46
variance depending on how you

01:05:48
actually build the build of

01:05:49
materials. Right. So that me,

01:05:53
sorry, sorry, go ahead.

01:05:54
All right. So I think it sounds

01:05:59
like if it does things more

01:06:01
efficiently, it's not only about

01:06:03
carbon footprint, it's also about

01:06:05
cost savings, right? So I guess

01:06:06
that's something that is really

01:06:08
interesting for your customers for

01:06:10
the enterprises buying that or the

01:06:12
data center. Yeah. 100%.

01:06:14
It's the first time they're saving

01:06:15
money while working towards

01:06:17
sustainability outcomes,

01:06:18
other than what you would do in

01:06:19
cloud for like green

01:06:20
ops, where realistically,

01:06:22
you're doing financial operations

01:06:23
and saying, I'm going to reduce

01:06:25
carbon, but really see I'm

01:06:26
reducing compute or I'm reducing

01:06:27
wastage or removing stranded

01:06:28
applications. We're doing the

01:06:30
exact same thing on the hardware

01:06:31
level and going, how do you do the

01:06:33
same work efficiently, rather

01:06:36
than just doing it. And so you're

01:06:38
going to get huge cost savings in

01:06:39
millions, you get thousands

01:06:40
of tons of carbon reduction. And

01:06:43
none of it has an impact on your

01:06:44
business, because you're just

01:06:45
eradicating waste. Right. So that

01:06:48
means your customers, I think are

01:06:51
mostly the data center

01:06:55
providers or is that like no

01:06:57
primary enterprise, truth be told,

01:07:00
because majority of data center

01:07:01
operators are colo or hyperscale

01:07:03
really, like realistically, people

01:07:05
have got to in other

01:07:06
people's co located facilities,

01:07:08
the colos are facilities managers,

01:07:11
they're not IT specialists,

01:07:12
they're not experts, compute,

01:07:14
they're experts in providing a

01:07:17
good environment for that compute,

01:07:19
which is why all the efficiency

01:07:20
metrics geared towards data center

01:07:22
have historically been around

01:07:24
buildings. Because it's been like,

01:07:25
how do we build efficiently? How

01:07:27
do we cool efficiently? How do we

01:07:28
reduce, you know, heat, density,

01:07:30
all this sort of stuff? None of

01:07:32
that addresses of why is the

01:07:33
building there, the building's

01:07:34
there to serve storage and

01:07:37
compute. And they just want every

01:07:39
colo washes their hands of that

01:07:40
goes, well, it's not our service,

01:07:42
someone else is renting the space,

01:07:44
we're just providing the space. So

01:07:45
you have this real unusual gap

01:07:47
that you don't see in many

01:07:49
businesses where the supplier has

01:07:51
a much higher level of knowledge

01:07:53
than the owner. So when you're

01:07:55
talking to someone saying, you

01:07:56
know, I think you should buy this

01:07:57
server, the manufacturer tells you

01:07:59
what to buy and the colo tells you

01:08:00
where to put it. But in between

01:08:02
that is the IT professional,

01:08:03
who's like, I have really no

01:08:04
control over the situation, the IT

01:08:07
provider doesn't tell me how

01:08:07
good it is. And the colo doesn't

01:08:09
tell me how to effectively run it.

01:08:11
So what I get is my asset,

01:08:13
and I give it someone else to

01:08:14
manage. So when you get this

01:08:16
perfect storm of

01:08:17
nobody really trying to

01:08:18
serve it better. And that's what

01:08:20
we do, we come in and go, you

01:08:21
know, you there's

01:08:22
there's huge amounts

01:08:22
of waste here. Yeah, that makes

01:08:25
sense. So it's the people or the

01:08:30
companies co-locating their

01:08:32
hardware into the data center.

01:08:34
Yeah, or running their own data

01:08:37
centers on premise, server rooms,

01:08:39
cabinets, you know, we do work

01:08:41
sometimes with people that got as

01:08:42
few as eight servers. And we

01:08:44
you know, we might make

01:08:45
recommendations like make

01:08:47
recommendations about

01:08:47
reconfiguring to change the

01:08:48
RAM set up, switch out CPUs,

01:08:51
things like that, that can have,

01:08:52
you know, 20, 30, 40% benefits,

01:08:55
but cost almost nothing. So it

01:08:57
could be the we see this client

01:08:58
where we've got

01:08:59
a small server estate that's very

01:09:01
low utilized, but massively over

01:09:03
provisioned on RAM because someone

01:09:05
someday, you know, 10 years ago,

01:09:07
bought a server and went stick two

01:09:08
terabytes in it. And it's like

01:09:09
how much you're using 200 gig and

01:09:11
it's like you've got 10 times more

01:09:13
RAM than you need, even at peak.

01:09:15
So can you just take out half your

01:09:16
RAM, please. And it sounds really

01:09:18
counterintuitive. Like just

01:09:19
just take out that RAM and put it

01:09:20
to one side. And if you scale up,

01:09:22
you can just plug it back in

01:09:22
again next week. But you know,

01:09:24
you've been using this for 8

01:09:25
to 10 years, and

01:09:26
you haven't needed

01:09:27
anywhere near that. But it's

01:09:28
sitting there drawing energy,

01:09:30
doing nothing,

01:09:30
providing no benefit, no

01:09:32
speed, no improvement, no

01:09:33
performance, just hogging energy.

01:09:35
And we look at that and go, that's

01:09:37
unnecessary. Yeah, and I think

01:09:40
because you brought up the

01:09:41
example of RAM, most people will

01:09:43
probably think, ah, that

01:09:45
little bit of RAM, that can't be a

01:09:47
lot of energy. But I mean,

01:09:49
accumulator of a whole year, it

01:09:51
comes down to something.

01:09:53
Especially when you have multiple

01:09:54
servers. Right? Yeah, absolutely.

01:09:56
Like RAM can be as much as 20 or

01:09:57
30% of the energy use of a server

01:09:59
sometimes from a configuration

01:10:00
level CPU is the main driver up to

01:10:02
65% of the energy use of a server

01:10:04
CPU. I mean, we're talking non GPU

01:10:06
servers, when it gets GPU, you get

01:10:08
an order of magnitude. But the

01:10:09
RAM can be using, you know, 30% of

01:10:11
the power on some of these

01:10:12
servers. And if you're only using

01:10:14
10% of that, you can literally

01:10:16
eradicate, you know, almost 20% of

01:10:18
the combined energy,

01:10:20
just by decommissioning either

01:10:21
certain aspects of that RAM or

01:10:23
just removing it

01:10:24
and putting on the

01:10:24
shelf until you know, you need it

01:10:26
next year or the year after. But

01:10:28
the industry is rife with

01:10:29
over provisioning at day one, to

01:10:31
give it scale at year five. But

01:10:33
actually, what would be more

01:10:34
sensible is provision for year one

01:10:36
and two with an ability to upgrade

01:10:38
to grow with the organization.

01:10:40
And what you'll save is you'll

01:10:41
decrease your carbon energy

01:10:42
footprint year on year,

01:10:44
you won't overpay month one for

01:10:46
the asset. And then in year two,

01:10:48
you can buy some more RAM in

01:10:49
year three, you can buy some more

01:10:50
RAM in year four, you can change

01:10:51
out the CPUs. And the CPU

01:10:53
buying in year four, by the time

01:10:55
you need to use it, you haven't

01:10:57
paid a 300% premium for buying the

01:10:59
latest and greatest. So it's about

01:11:01
effective procurement as well. You

01:11:03
know, you want 20 servers,

01:11:05
fine, but by the servers you want

01:11:06
for year one and year two, and

01:11:08
then year three, buy the upgrade,

01:11:11
like upgrade the components year

01:11:12
four, upgrade year

01:11:13
five upgrade do, you know,

01:11:15
incremental improvement, and then

01:11:16
you're not paying a really high

01:11:17
sunk energy cost at year one.

01:11:19
But also your procurement costs

01:11:20
really high, because the second

01:11:21
you buy it when it's new,

01:11:23
two years later, it's half the

01:11:24
price. If you haven't

01:11:25
used it to its, you know,

01:11:27
fullest potential in years one and

01:11:29
two, you fundamentally get 50%

01:11:30
saving if you buy it in

01:11:31
year three. But no one thinks I

01:11:33
want to buy and forget. Do you

01:11:35
know what I mean?

01:11:36
Yeah, especially for CPUs, I think

01:11:37
in three years time, you have

01:11:39
quite some leap,

01:11:40
maybe a new generation, same

01:11:42
socket, lower TDP, something like

01:11:44
that. But you shocked me with a

01:11:47
30%. I think I have to look at my

01:11:48
server in the in the basement.

01:11:51
Just shocked me.

01:11:53
Crazy. And like, we're seeing some

01:11:55
stuff now that we

01:11:55
can get persistent RAM,

01:11:57
which actually doesn't act like

01:11:58
RAM actually stores some aspects

01:12:00
in the in the memory.

01:12:01
You know, that's that's fairly

01:12:02
energy intensive, because it's

01:12:04
sitting there constantly using you

01:12:05
when the system's not up. But

01:12:08
realistically, yeah, your RAM is,

01:12:10
it's a relatively big energy user,

01:12:12
we know for every sort of, you

01:12:14
know, degree of gigabytes, you've

01:12:17
got an actual wattage figure for

01:12:18
that. You know, so it's not

01:12:21
inconsequential. And that's a

01:12:23
really easy one.

01:12:24
That's not exactly

01:12:24
everything we look at. But there's

01:12:25
aspects of that. So we had CPUs

01:12:30
and we had RAM, I think CPUs are

01:12:32
obvious. You also mentioned like

01:12:34
graphics cards, I think if you

01:12:36
have like a server with a lot of

01:12:38
graphic cards, it's obvious, but

01:12:40
it's, it's gonna use a lot of

01:12:42
energy. You had RAM, any anything

01:12:45
else that comes to mind? Like,

01:12:48
this, I think hard disk drives are

01:12:50
probably worse than than SSDs and

01:12:52
NVMe. Yeah, it's an interesting

01:12:54
one. So storage is really, a

01:12:56
really fascinating

01:12:56
one for me, because I

01:12:57
think we're moving back towards

01:12:59
tape storage as a carbon

01:13:02
efficient method of storage. And

01:13:03
people look at me and go, why

01:13:05
would you say that? And it's like,

01:13:06
well, if you're if you accept

01:13:08
the fact that 60-70% of data is

01:13:10
worthless, as in like you use it

01:13:12
once you never use it again.

01:13:13
And that's a that's a pretty

01:13:14
standard metric. I think it's I

01:13:16
think it's as high as 90% of data

01:13:17
doesn't get used, but 65% will

01:13:19
never get used. And what we have

01:13:22
is loads of people moving that to

01:13:23
cloud that storage and going,

01:13:24
right, I can now immediately

01:13:26
access data that I don't want, and

01:13:28
we'll never use and we'll never

01:13:29
look at. And it sits there on

01:13:31
really high available SSDs going,

01:13:33
I can retrieve this information I

01:13:34
never want. Instantly. Well, the

01:13:38
SSD wears over time, every

01:13:39
time you read, write every time

01:13:40
you pass information through it

01:13:42
wears right like that's

01:13:43
that's how that's how flash memory

01:13:44
works. HDDs have a much longer

01:13:47
lifecycle than SSDs, but lower

01:13:50
performance. And you know, your

01:13:51
average say your average hard

01:13:52
drive, say it uses

01:13:54
six watts an hour

01:13:55
and an SSD uses four. So you go,

01:13:57
well, okay, that's, you know, 34%

01:13:59
more efficient to use SSD.

01:14:01
And it's like, well, it is except

01:14:03
for there's an embodied cost of

01:14:04
the SSD, the creation of the SSD

01:14:06
is 10-15 times higher than the hard

01:14:09
drive. So if you're storing data

01:14:11
that you never use, so no

01:14:13
one's ever using that six watt

01:14:14
read read and write, it just sits

01:14:17
there with a really high sunk

01:14:18
environmental cost until it runs

01:14:20
out. And then you kind of might be

01:14:22
able to reuse it, you might not.

01:14:24
But realistically, you're going to

01:14:25
get through two or three

01:14:26
lifecycles of SSD

01:14:28
for every hard drive.

01:14:29
If you never look at the data,

01:14:31
it's worthless, you've got no

01:14:32
benefit there. But there's a huge

01:14:33
environmental cost from a credit

01:14:34
for all materials and from a

01:14:36
storage point of view,

01:14:37
consequently take. So the great

01:14:39
example, if you've got loads of

01:14:40
storage on cloud,

01:14:42
and you never need

01:14:44
it, but you've got to store it

01:14:45
like medical data for 100 years,

01:14:47
why are you storing data that you

01:14:49
need for 100 years on SSD in cloud

01:14:51
and paying per gig, when you could

01:14:54
literally pack, you know,

01:14:56
a million pounds worth of storage

01:14:57
onto one tape and have someone

01:14:59
like Iron Mountain run archival

01:15:01
as a service for you, where you

01:15:03
can say if you need any data, we

01:15:04
can retrieve it and pass it into

01:15:05
your cloud instance. And there's a

01:15:07
really good company called called

01:15:08
TEZ in the UK. And TEZ

01:15:12
basically have this great archival

01:15:13
system. And when I was talking to

01:15:14
them, I was like, it really made

01:15:15
sense of how we position system of

01:15:17
systems thinking, where they go,

01:15:19
well, we run tape. So we take all

01:15:22
your long term storage and put it

01:15:23
on tape, but we give you an RTO of

01:15:25
six hours, and you just raise

01:15:28
a ticket going, you know, I want I

01:15:29
want information on this patient,

01:15:31
and they retrieve it and put it

01:15:32
into cloud instance. So you have

01:15:33
it immediately. No one needs that

01:15:35
data instantaneously, but you're

01:15:36
sitting it on NVMe storage, which

01:15:38
has a really high environmental

01:15:40
energy cost and financial cost

01:15:42
to basically be readily available

01:15:44
when you never need it.

01:15:46
Consequently, stick

01:15:46
it in a vault on tape

01:15:47
for 30 years and have someone

01:15:48
bring it when you need it. You

01:15:50
know, you drop your

01:15:51
costs by 99 times.

01:15:53
And you're environmental input is

01:15:54
that makes a lot of sense. Yeah,

01:15:56
that makes a lot of sense,

01:15:56
especially with all data that

01:15:58
needs to be stored for regulatory

01:15:59
reasons or stuff like that. And I

01:16:02
think some people kind of try to

01:16:06
solve that or mitigate it a little

01:16:08
bit by by employing some

01:16:10
tiering or using some tiering

01:16:11
technologies going from NVMe down

01:16:13
to HDD and eventually maybe to

01:16:16
something like S3 or what is

01:16:18
it S3 Glacier? Glacier storage

01:16:21
for AWS. Yeah, yeah. Right. So

01:16:23
but I think the tape thing is

01:16:25
still one step below that. Yeah, I

01:16:29
mean, you think it's just

01:16:30
already writing off and I heard a

01:16:32
horror story of guys moving from

01:16:34
Glacier storage as an energy and

01:16:35
cost saving mechanism, but not

01:16:37
understanding that you pay per

01:16:38
file not per per watt or not

01:16:41
per terabyte or per gig. And it

01:16:42
costing like six figures to move

01:16:45
this data over to Glacier and

01:16:47
going, it's going to save you

01:16:48
three grand a year, but it's now

01:16:49
your payback point

01:16:51
is like 50 decades.

01:16:53
And it's like you don't realize

01:16:54
when you make these decisions that

01:16:55
you go well, actually,

01:16:56
there's a huge egress cost there.

01:16:58
Whereas how much would it cost to

01:17:00
take that data and stick it into

01:17:01
onto a tape? 100 quid? 200 quid?

01:17:04
You know, you're talking about

01:17:06
significant cost savings and

01:17:07
environmentally, you're not

01:17:08
looking after the systems, you're

01:17:10
not looking after the storage,

01:17:11
you're using an MSP to hold that

01:17:13
storage for you and then guarantee

01:17:15
your retrieval within

01:17:16
timescales you want. It's a very

01:17:17
clever business model that I think

01:17:19
we need to revisit of when is

01:17:21
tape the best option. And for long

01:17:23
term storage, archival storage,

01:17:25
from an energy point of view and

01:17:26
a cost point of view, it's

01:17:27
very clever and sustainability

01:17:28
wise, it's a real win. So yeah,

01:17:33
tape as a service, it's a thing

01:17:35
you've heard it here first. It's

01:17:37
what people are going to be doing.

01:17:39
I like that. TaaS, tape as a

01:17:41
service. So going from super

01:17:46
old technology to a little bit

01:17:48
newer stuff, like, what

01:17:51
would drive sustainability in

01:17:53
terms of new technologies?

01:17:55
I hinted at lower TPS for new

01:17:58
CPUs, probably the same goes for

01:18:00
RAM, I think the

01:18:00
chips get lower in

01:18:03
wattage or watt usage over time.

01:18:05
Are there any other

01:18:07
like, specific factors?

01:18:10
Yeah, I think the big one for me

01:18:11
is that the new DDR5 RAM is

01:18:15
really good. Like it unlocks a lot

01:18:16
of potential the CPU level as in

01:18:18
like, the actual most recent jump

01:18:20
in efficiency is not coming from

01:18:22
CPUs, you know, Moore's law slowed

01:18:23
down in 2015, I still think it's

01:18:25
not hitting the level it was,

01:18:27
but the next generation for us is

01:18:29
ASIC based, you know, as in

01:18:31
application

01:18:32
specific interface chips,

01:18:34
basically, because I think

01:18:35
realistically, there's not much

01:18:36
further the CPU can

01:18:37
go, it can still go,

01:18:39
we can still get some more juice

01:18:39
out of it, but it's, we're not

01:18:41
doubling every two years. So the

01:18:42
CPU is not where it's at, whereas

01:18:44
the ASIC is very much where it's

01:18:45
at now, like specific chips

01:18:47
built very specific function, like

01:18:49
Google's VSOs, for example, you

01:18:51
know, that they're entirely geared

01:18:53
towards encoding for YouTube, 100

01:18:55
times more efficient than a CPU or

01:18:57
GPU at doing that task,

01:18:58
you know, we saw the rise of the

01:18:59
ASIC through Bitcoin, right? Like

01:19:01
specific mining ASICs.

01:19:02
So I think specificity around

01:19:04
chips is really, really good. New,

01:19:06
like I said, new RAM is decent.

01:19:07
Very, very good. The GPU wars is

01:19:11
an interesting one for

01:19:11
me, because we've got

01:19:13
GPUs, but there's no really

01:19:15
definable benchmark for comparison

01:19:17
of how good a GPU

01:19:18
is other than total

01:19:19
work. So we have this thing where

01:19:21
it's like how much total grunt do

01:19:22
you have, but we don't really

01:19:23
have metrics of how much grunt per

01:19:25
watt. Because GPUs have always

01:19:28
been one of those things we power

01:19:29
supercomputers with so you know,

01:19:30
it does a million flops and this

01:19:31
many mips and all the rest of it,

01:19:33
and we go, but how good does it do

01:19:34
it? How good is it at doing its

01:19:35
job? It's like that's irrelevant.

01:19:37
It's how much total work it can

01:19:38
do. Yeah, like right. So we need a

01:19:40
rebalancing of that. And I

01:19:41
don't, you know, that's not there

01:19:42
yet. But I think it'll come soon.

01:19:43
So we can understand what GPU

01:19:45
specific functions are. And I

01:19:48
think the real big change for us

01:19:49
is behavioral change. Now,

01:19:50
I don't think it's technology. You

01:19:52
know, I think understanding how we

01:19:54
use our assets, visualizing

01:19:56
use in terms of non economic

01:19:58
measures. So big basically being

01:20:02
decent digital citizens, I think

01:20:03
is the next step. I don't think

01:20:05
it's a technological

01:20:06
revolution. I

01:20:07
think it's an ethical

01:20:08
revolution, where people are going

01:20:10
to apply grown up thinking to

01:20:13
technology problems rather than

01:20:15
expecting technology to solve

01:20:16
every problem. So yeah, I mean,

01:20:19
there are there are incremental

01:20:20
changes, we've got some good

01:20:20
stuff. But realistically, the next

01:20:23
the next step

01:20:24
evolution is how we apply

01:20:26
our human brains to solve

01:20:27
technological problems, rather

01:20:29
than throw technology

01:20:31
as problems and hope

01:20:32
for the solution. I like that. I

01:20:34
think it's it's a general or a

01:20:38
really important thing in general

01:20:39
that we try not to just throw

01:20:42
technology at problems or

01:20:45
even worse, creating

01:20:47
technology in search for a

01:20:49
problem, right? Yeah, all the

01:20:53
time, number of

01:20:54
times I like, you know,

01:20:55
we're a scale up business. In

01:20:56
throat, we're doing we're doing

01:20:57
really, really

01:20:58
well. We don't act like

01:20:58
a scale up. But you know, last

01:21:00
year, I was I was sort of

01:21:02
mentoring some

01:21:03
startup guys and some

01:21:04
projects that were been done in

01:21:05
the UK. And 90% of people were

01:21:08
applying technology

01:21:09
to solve a problem

01:21:10
that didn't need solving. Like

01:21:11
every question I would ask

01:21:13
these people is,

01:21:14
what does this do?

01:21:15
What does this? I'm like, and does

01:21:16
the world need that? Well, it's a

01:21:18
problem. It's like I feel like

01:21:19
you've created a problem. Because

01:21:21
you have the solution to a

01:21:21
problem. It's a

01:21:22
bit like, you know,

01:21:23
it's a bit like an automatic tin

01:21:24
opener. Do we need a diesel

01:21:26
powered chainsaw

01:21:27
tin opener to open

01:21:29
open sins? Or do we already kind

01:21:31
of have tin openers? Like how far

01:21:33
do we need to innovate

01:21:34
before it's fundamentally useless?

01:21:37
Do you know what I mean? And I

01:21:38
think a lot of problems are,

01:21:39
we've got AI and we've got

01:21:40
technology. So we know we've got

01:21:42
an app for that. And

01:21:43
it's like, maybe we

01:21:44
don't need an app for that. Maybe

01:21:45
we need to just look at the

01:21:47
problem and go, is

01:21:47
it really a problem?

01:21:48
Or have you solved something that

01:21:50
didn't need solving? And a lot of

01:21:52
ingenuity and waste goes

01:21:54
into solving problems that don't

01:21:55
exist. And then conversely,

01:21:58
there's loads of

01:21:59
stuff out there that

01:22:00
solves really important problems,

01:22:01
but they get lost in the middle.

01:22:03
So I've seen some really,

01:22:04
really great products when I was

01:22:05
working with some of these startup

01:22:06
businesses, where they've

01:22:07
got really, really good solutions,

01:22:09
but they can't articulate the

01:22:12
problem it's solving. And in some

01:22:14
cases, the ones that are winning

01:22:16
are the ones that sound very

01:22:18
attractive. I remember there was

01:22:19
like a med tech one that was

01:22:20
talking about stress management.

01:22:21
And it was like

01:22:22
providing all these

01:22:23
data points on what levels of

01:22:26
stress you're dealing with. And

01:22:27
it's kind of like, that's really

01:22:28
useful to know that I'm very

01:22:30
stressed. But other than telling

01:22:33
me all the psychological factors

01:22:35
that I am feeling stressed, what

01:22:36
is the solution on the product

01:22:37
other than to give me data telling

01:22:38
me that I'm really stressed? It's

01:22:40
like, well, there isn't anything

01:22:41
that just tells you that data.

01:22:42
It's like, right, and then what,

01:22:43
and then we can take that data,

01:22:44
it'll solve the problem later on.

01:22:45
It's like, no, you're just

01:22:46
creating a load of data to tell me

01:22:47
things that you

01:22:49
don't really think has

01:22:49
a benefit. If you've then got the

01:22:51
solution of now with this data, we

01:22:53
can make this inference,

01:22:54
we can, we can solve this problem.

01:22:56
That's really useful. But

01:22:57
actually, you're

01:22:57
just creating a load

01:22:58
of data and going, and what do I

01:22:59
do with that? And you go, don't

01:23:00
know, it's up to you. Okay,

01:23:02
well, I'm just gonna look at it

01:23:02
and say, it looks like I'm

01:23:03
struggling today. Do

01:23:06
you know what I mean?

01:23:07
Yeah, yeah. Unfortunately, we're

01:23:09
out of time. I could chat about

01:23:12
that for about an hour. You must

01:23:14
be so happy, or you must have been

01:23:16
so happy when the proof of work

01:23:18
finally got removed from all the

01:23:22
blockchain kind of stuff. Anyway,

01:23:25
thank you very much. I was very

01:23:27
delightful. I love chatting and

01:23:32
just laughing because you just

01:23:34
hear all the stories from people,

01:23:37
and especially about things

01:23:38
you normally are not part of,

01:23:40
right? As I said, you completely

01:23:41
shocked me with 30%. Obviously,

01:23:43
RAM takes some energy, but I

01:23:46
didn't know that much. So I hope

01:23:48
that some other folks actually

01:23:51
also learned something and apply

01:23:53
the little bit of ethical thinking

01:23:56
in the future whenever we create

01:24:00
new startups, whenever we build

01:24:02
new data centers, apply a new

01:24:05
hardware or anything like that.

01:24:07
Thank you very much.

01:24:09
Appreciate it.