audio
audioduration (s) 3.99
4
| transcription
stringlengths 1
1.4k
⌀ |
|---|---|
Hi, I'm Stephen Jones and I'm one of the
|
|
And one of the ways that that comes out is that there are a key performance elements.
|
|
be this one and a half terabytes per second data, right?
|
|
do the division, that's 194 billion double versus values per second.
|
|
giving me a peak performance based on memory of just 190.
|
|
for gigaflops. Right now I'm only beating the 1996
|
|
computer in the world.
|
|
So let's have a look at this memory thing. Let's have a closer look at.
|
|
how it works because it's so important in the performance.
|
|
machine. A single bit of memory is a capacitor.
|
|
And now the holes are fucked for one bit on the left or it's empty for a server.
|
|
The memory is read by switching on the transistor which can be...
|
|
you need to take into account. Even while the bulk of your program can be pretty naive, see, plus,
|
|
exit to a wire, the bit line. The wire then carries a
|
|
based on the charge in that capacitor so it's the wired record either.
|
|
and honor and offer one or a certain.
|
|
DRAM chip consists of millions of these cells all connected together.
|
|
a big 2D matrix. This matrix layout lets me access any row, any call.
|
|
And this is why it's called random access memory. That's the random access.
|
|
as again say a magnetic tape which has not in your access.
|
|
Data is addressed by a row and a column index.
|
|
We are taken from the requested address.
|
|
First, the row is accessed. All the cells in the row...
|
|
plus I'm not going to teach you could today there's not enough time for that.
|
|
activated and their state is copied up to these things called the sensor.
|
|
and the sense amplifiers read the tiny chart
|
|
on each of the capacitors in the cells and turn them into well-defined bolts.
|
|
to do that can much more easily be read in the next set. The problem is...
|
|
the charge and the capacitor is drained as this happens right, I'm connecting a wide.
|
|
to the capacitor's drainable the electrons out. And so the data in the row is destroyed.
|
|
I'll come to that in a moment. Next.
|
|
the column access takes place. Instead of reading from memory cells,
|
|
row is already in the amplifier so it reads the data held in the amplifier.
|
|
much quicker, much easier to read than a row because the amplifiers will produce a strong clear
|
|
But I'll teach you a few things that I think are vital to think.
|
|
signal and so I can read more quickly.
|
|
You can read repeatedly from the amplifiers because they hold that voltage. You can read as many
|
|
times you like from the Festrow. So if you can open a row and use it repeatedly.
|
|
then you will not have it deal with the capacitor.
|
|
Because it's so common in fact to read adjacent memory locations in a
|
|
There's this thing called burst mode, where a single request returned multiple weather.
|
|
date. This is a huge deal because it means I don't have to pay for the individual request.
|
|
over again and pretty much every processor in the world uses this because the
|
|
system of the processor is always going to go and read multiple bites at a time.
|
|
And then GPU the Cassistum is 128 bytes a time. I'll talk about the Cassistum.
|
|
about when you're programming the GPU. I think the most
|
|
of that. The problem is the way
|
|
I need to read another row. I first have to write back the
|
|
data which was held in the amplifiers. If you remember, the row was drained when it was...
|
|
it was copied into the amplifiers because the capacitor is discharged. So we now have to...
|
|
rewrite it to avoid memory corruption. So this makes a page
|
|
which expensive could involve both the right back and then a new road low.
|
|
up the new road load into the amplifier.
|
|
Hardware, course things, roads or pages pretty much interchangeably, so if you hear the term.
|
|
page, then this is what they mean. I mean a row of your memory switching page.
|
|
It was about three times as expensive as switching column within a page because of this look.
|
|
important thing when doing any engineering is to have an accurate mental
|
|
store operation, sorry store and load operation. So, put in your couple of
|
|
model of the system they are using.
|
|
So today, a mental model, I really think the best way to understand the how of something is to know.
|
|
why it's that way. So this talk is really about why could
|
|
is the way that it is, not just how.
|
|
architects of CUDA. I've been working on the CUDA programming model in DB.
|
|
That's a good question. Why is Coup de la Weyrt is?
|
|
Right. It's the way it is because.
|
|
of the laws of physics, quite literally. So, what
|
|
I mean by that. Well, if you're using a GPU
|
|
because you want performance of some kind. Curivest designed in part to allow.
|
|
I would get maximum performance on the GPU. Right, it's obviously, as I said, also designed to make it.
|
|
programmable. Performance is limited by the laws.
|
|
physics and I'll get to that in a moment. And so Kudra is designed.
|
|
do it best to help you work with the hardware.
|
|
than loads of physics to get good performance.
|
|
computing since Scotch 2008 and why
|
|
So this is actually a really interesting point to make. You see what's special about
|
|
is that we make both programming language for the hardware and the hardware
|
|
programming language. This means what only do we get to adjust the programming
|
|
language to match what the hard way can do. But we also get to adjust the hardware.
|
|
So there's more programmable the hardware designers come up with really
|
|
have a staff to overcome limitations like speed of electricity and silicon.
|
|
and could have evolved to allow this clever stuff to be programmable. Literally speaking.
|
|
could it is shaped by the laws of physics?
|
|
So I made another possible
|
|
contentious statement that I want to look at more closely for women.
|
|
One of the best things about this job is that it could really be a co-design between hardware.
|
|
whole talk basically about the CTC last year and I put the link below.
|
|
low as a seamless plunge for my talk, but also because if you're interested.
|
|
gives you a lot more detail and then I'm going to get into right here about the hardware.
|
|
and overcoming physical constraints. Anyway, I won't repeat the whole thing.
|
|
but I will bring up the main points. Let's start with
|
|
system, though, because presumably you paid money and...
|
|
investing time in GPU computing because you want performance.
|
|
from it. So let's look at what that means.
|
|
make up, I hope his lung controversial statement that's getting the best performance.
|
|
is about using all the GPU resources that you can.
|
|
software. Since CUDA is the way you program the DPU directly,
|
|
In other words, the more threads I'm running, the more memory I'm moving,
|
|
calculations I'm making, the better I'm probably doing.
|
|
So these are the feeds and speeds of the Ampede view.
|
|
and the obvious performance metric to look at is flops.
|
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 2