Datasets:
Shabdobhedi
/
TTS_English_Technical_Terms

Dataset card Data Studio Files
xet
 Community
1
audio
audioduration (s)
3.99
4
transcription
stringlengths
1
1.4k
⌀
has a ridiculously large number of flops. Right, so that's almost 10.
and tariff lobs of double precision performance on those 20 of single precision.
And that's not counting the tensor calls, which are way too complex to get into here.
more flopped on work. Let's discuss. It's easy to say. For the
more of those flops your application is getting, the better you're doing.
And there's no question about it, but very often it's not...
It's vital that we give access to every last bit of performance that's available.
actually the fact that it is limiting performance that your application is able to get.
There's usually more flops than you need. The 10-ter flops is a lot.
For reference, this single chip.
It's more powerful than the biggest supercomputer in the world from 2021.
years ago. The ASCII-Y Super computer was 200k.
it weighed over a hundred tons and it was about three quarters of a pound.
powerful look, there's this tiny little piece of silicon. It's if nothing else.
a testament to Mooslaw. It's truly amazing, but it means...
that this thing that you have here has all of
the capability, all of the power of that giant machine. And you've got it.
but at the same time were incredibly aware that programming a massive amount of time.
all of your fingertips in one step. And your by step.
I hope we've evolved much more now to get to a point where it's easier to program, easier to
access that kind of performance. But that 20 years ago.
the scale was an entire room.
I'd like to argue that Flops doesn't matter. Memory bandwidth.
is what matters. I mean, yeah, you want the computer performance.
That's not what limits you and that's almost never what ends up defining how
your CUDA program works.
Let's look at a schematic of the A100 GPU here on the right.
it has 100 nm of SMs connected to the memory system. And.
the paranormal seen as complex. So the single most important priority.
Each SM, which is sort of the core of the core of the machine.
Each SM can request 64 bytes of memory per clock.
And a clock speed of 1410 megahertz.
almost 10 terabytes of memory requested per second. By Con.
trust the incredibly fast HBM2 memory system can provide.
over one and a half terabytes of memory per second. That's an enormous
amount of memory bandwidth, but it's still over six times
less data than the SEM itself. And we're going to be able to see the other one to pull in.
So there's just no way to feed data into the GPU at the
rate that it can request it. And even the rate it can request it is not enough to
me as always program building. There's an interesting tension there.
all of the flops that it has available with that's another story, right?
means we're almost always limited by memory performance and not by the...
the threads being unable to keep up with the data.
keeping your threads busy, you're not getting the most done that you can.
So if I'm looking at it.
rate I can feed data into my system purely from a memory standpoint.
which is not completely representative because obviously I'm happy to reuse.
and there's other factors with going to Fox, but if I just look at...
what program that needs to suck in data.
and process that data as fast as it can. The limit is going to
strikes s of
and open ID Connect. But a lot of people don't really know what they are, or maybe they understand.
a little bit less confusing. And you can let me know at the end if it makes a little bit more sense.
So, let's go back in time. Let's take our time machine and go back in time.
time about maybe 10 years back to like 2006 2000
This is a long time ago, especially in internet years. This is a long time ago.
long time ago because back in say 2006
we didn't even really have smartphones. That was like brand new with the iPhone came out.
and like oh seven i think uh... so this is a long time ago back then
though just like today, applications and websites had a number of
of different use cases that they might have to deal with in terms of like.
Woggan authentication authorization. I'm just going to use the
in them a little bit but not really understand them fully. So just as a quick
term identity as kind of like a catch all term for like authentication authorization
log in permissions that type of thing. So if we look at the kind of common idea.
at NAD use cases from like 10 years ago, we have that simple login stuff.
like I said, which you could probably just fulfill with some form.
some cookies or something, just log in to use your name with a username and password, that's pretty simple.
We also have some apps that need to do single sign on or SSO.
And that was typically done with a protocol called SAML.
Today it's still very often done with a protocol called Samo as well.
Sample protocol is very good for this type of single sign on use case, which
is basically where you have one account in a system that lets you get log in or get access.
Who here has at least heard of Open ID Connect or OAuth?
as to a number of different systems. This is really common in businesses where you have a single...
ask your account or you have an active directory account or an employee account.
that lets you get into a number of different systems. And Samo is the product
all that kind of makes that work under the hood. The sample protocol
It kind of has a reputation of being a little bit hard to work with as a developer.
because it's pretty obscure. Even more obscure than Oath, it's very...
very dense, but it works and it was used back then just like it's
there's a couple of new use cases that were just starting to come up.
onto the scene about 10 years ago or so, that we're new.
though. So one of them was mobile apps. Like I said mobile apps smart
What about would you would you feel like you're an expert on it?
phones were brand new, the idea of mobile apps and app store was like brand new.
And so there is this whole new set of use cases where we want to be able to log into a mobile
app, but you want to stay logged in after you close the app most likely.
How do you keep a long live session alive on a device?
device. Do you use like a cookie that doesn't really work well on devices?
It was a new feel, it was brand new. So we didn't really have a good solution for that.
way. As well as something called a use case called delegated authorization.
which probably sounds like the most boring academic thing you can imagine.
But this is actually what I'm going to focus on here. Delegated authorization.
is an interesting use case which is also the genesis or where
or you really understand the protocol at a pretty good level. Not.
the O-Watt protocol came out of. So despite sounding very boring.
This is exactly what I'm going to focus on. So the delegated authorization problem.
is something that, despite sounding very boring, something that
that you probably interact with, maybe on a daily basis, even maybe today I'll-