Hi, thanks for turning into Singularity
Prosperity. This video is the sixth in a
multi-part series discussing computing.
In this video, we'll be discussing the
rise of GPU computing. While CPUs utilize
multiple cores and as of recently have
been increasing core counts, graphics
processing units, GPUs, essentially take
the concept of parallelization and
expand on it in a massive way. Observing
the trend from the mid-1990s, you can see
core counts were steadily increasing for
about a decade, from 1 core in 1995 to
24 cores by 2006, so a steady rate of
growth of about 2 cores per year. 2006
was an inflection point in GPU computing,
where we saw a massive jump from 24 cores
to over 5 times that, 128 by 2009.
Progressing to present day, you can see
that the shift in massive parallelism
has been focused on more and more, from a
rate of growth of about 2 cores per year
to now about 200. As of this year, 2018, we
are now reaching graphics architectures,
such as Nvidia Volta, with core counts
over 5000. To understand why GPUs are
becoming so widely adopted today, it is
best to first understand their original
purpose, graphics, more specifically for
gaming. I'm sure that's how the majority
of people, myself included, still view
companies such as Nvidia and AMD today.
Gaming is an extremely computationally
intensive task, and as the years have
progressed, it has only continued to grow
more and more complicated. There are so
many things to simulate such as:
materials, lighting, shadows, physics of
different materials and their
interactions with lighting as well as
other materials, various fluid
simulations, procedural generation - the
list can go on and on. To highlight the
level of complexity some games can reach.
I want to play a clip from one of the
most ambitious games and development
today, Star Citizen:
[Music]
[Music]
[Music]
As you can see and from the thousands of
other applications I didn't even mention,
games requiring the order of millions
even billions of computations per second,
and this is only set to increase as
technologies such as VR start getting
more traction. This is why GPUs are
focused and have increased so rapidly
in parallelism since the mid-2000s. This
massive parallelism is now bleeding into
other applications,
one of the most mainstream currently and
why such mass adoption of GPUs has grown
is, artificial intelligence, more
specifically deep learning. We'll
explore this topic much deeper in this
channels AI video series, but essentially,
machine learning models typically
require billions even trillions of
matrix operations, which are just simple
multiplies and additions. GPUs by default
excel at doing large amounts of
repetitive calculations due to their use
in gaming, thus deep learning was a
perfect fit. All of this is a positive
feedback loop, increased graphics and
simulations and increased AI performance
will propel each other forward, such as AI
in computer characters and increased
realism in simulations and physics. This
also translates to progressing
scientific innovation as well, both in
the micro and macro scale, with the use
of GPUs to model and research the small
things such as protein folding to
things as large as more accurate weather
simulations or even larger such as
gravitational wave simulations. Now a
word from Nvidias CEO and co-founder,
Jensen Huang, to highlight some of what
I've talked about further: In
order to train this neural net with huge
amounts of data, you need to do huge
amounts of computation. Trillions and
trillions and trillions of operations,
necessary to effectively for this neural
network brain to learn how to perform an
intelligent task, like image recognition
or voice recognition.
Well that handicap, that handicap was
overcome with the discovery of the GPU.
So these two factors, the end of
Moore's Law and the emergence of a new
software development method - a new
software technique happened at almost
exactly the same time, and these two
dynamics turbocharged the adoption of
GPU computing, this form of computing
we've been working on for literally a
decade and a half. In fact you could say
that ever since the founding of our
company, we've focused on only this form
of computing, no
company of our scale and size has ever
dedicated itself does this one singular
field of computer science: GPU
accelerated computing is all we do, now
as a company of 12,000 people
we've now dedicated and invested some
close to 30 billion dollars in the
pursuit of advancing GPU accelerated
computing. Easily observable by Nvidias
stock price, GPU computing is the new
industry standard and is now being
massively adopted across all fields that
require intensive computing. This is also
observable by the trend of compute
performance over the past 120 years, with
the last seven data points from 2006 all
being GPUs. The majority of GPUs are
still primarily at 16 nanometer node
sizes, able to deliver performance at
teraflop levels. Such as Nvidia's Pascal
and AMD's Vega high-end consumer cards,
which both clock in around 12
teraflops. This is just the general
performance for calculations run on
these GPUs, as discussed in a previous
video in this computing series, hardware
and software optimizations can increase
this performance much further. This can
be exemplified by Nvidia latest GPU
architecture, Volta, specifically designed
for artificial intelligence. Their
highest-end card, the Tesla V100,
boasts a core count of 5120 with a node
size of 12 nanometers, yielding general
performance of 15 teraflops.
However, when running machine learning
computational tasks, yields a performance
of 120 teraflops, 8 times greater
than its general performance and well
exceeding the prediction of Moore's Law!
This insane performance is achieved due to
tensor cores and tensors sort. As
mentioned earlier, deep learning involves
many matrix operations, tensor cores are
specialized hardware on Volta GPUs
that allow for asynchronous computation
for the multiplications and additions
involved in these matrix operations. This
yields 12 times greater throughput
than the same operations on a Pascal
card. Tensor sort simply refered to as
TensorRT is an optimized compiler for
machine learning code.
So essentially, Nvidia Volta cards
support various machine learning
languages such as: Tensorflow,
PyTorch and others, then TensorRT
compiles and optimizes the code to be
run on the tensor cores.
This is done through various processes
such as reduced precision, layer fusion,
multistream execution and much more that
we'll explore in a future AI video. This
optimization gives 40 times boosts in
performance and 30 times boosts in
efficiency and this design methodology
is thr pinnacle of hardware and software
optimization and how 1^14
flop performance is achieved for these
machine learning computations. In terms
of practical applications for machine
learning models, a 40 times boost in
image recognition and a 140 times boost
in natural language processing is seen,
as well as a halved latency for both.
This is a huge deal, when talking to a
virtual assistant for example, 3/10
of a second versus 1/10 of a
second make a vast difference in the
experience. It is also worth noting, this
is just the first iteration of Volta
architecture. The Tesla V100 is a data
center chip and only buyable in a DGX unit,
this is quite expensive, in the range of
$70,000+, but its purpose isn't for the
average consumer and mainly made for
startups and other such companies. As of
recently, in December 2017, the Titan V
was announced. This card is the first foray
into consumer level machine learning
with a price tag of $3,000. As time
progresses, industry cards like the V100
and consumer cards will only get better and
cheaper, opening up AI for everyone!
In terms of general GPU computing, also
refer to as GPGPU computing, there are
various optimizations both in hardware
and software being made. GPUs have
built-in memory called VRAM which is
essentially like DRAM, new GPU memory
architectures are reducing the latency
calls to memory induce, this is called HBM,
high bandwidth memory and like 3D NAND
utilizes more z-space, which allows for
more memory to be added to GPUs as well
as bringing the memory closer to the
processing die to reduce latency. Also
evolving PCIe standards with 4.0 and 5.0
will also assist with data bandwidth and
GPU power consumption. If you want to
learn more about memory and data, be sure
to watch the previous video in this
series, if you haven't already. In terms
of software optimization, Nvidia's
compute unified device architecture, CUDA
platform is beginning to grow massively.
CUDA is essentially a parallel computing
platform which allows developers to
utilize the mass parallelism in GPUs
for general-purpose computing. There is
also improvement in the graphics and
simulation front with DirectX 12 and
Vulkan APIs. These allow better GPU
multi-threading, CPU communication and
much much more when it comes to gaming
and other simulation and graphically
intensive tasks.
All in all, GPUs are just beginning
to hit their prime, with more adoption
and exponentially increasing performance,
due to major innovations of both
hardware and software and they still
have a long ways to scale with general
performance expected to continue growing
into the 2040s. Performance is currently
growing at a rate of 10 times every five
years, so following that logic by 2040
we'll have GPUs 100,000 times better than
today! St this point the video has come
to a conclusion, I'd like to thank you
for taking the time to watch it. If you
enjoyed it consider supporting me on
Patreon to keep this channel growing and
if you want me to elaborate on any of
topics discussed or have any topic
suggestions, please leave them in the
comments below.
Consider subscribing for more content,
follow my Medium publication for
accompanying blogs and like my Facebook
page for more bite-sized chunks of
content. This has been Ankur, you've
watching Singularity Prosperity and I'll
see you again soon!
[Music]
For more infomation >> 一番手軽で話題の外付けマイクを買いました ソニー ECM C10 - Duration: 5:33. 
For more infomation >> Rap về Pennywise ( IT - Chú Hề Ma Quái ) | 360hot REN Music - Duration: 2:09. 
For more infomation >> How to Create a Cat-Friendly Indoor Garden - Way to Grow - HGTV - Duration: 1:31. 
For more infomation >> How to embrace: Contemporary music - Duration: 1:40. 
For more infomation >> Advanced English 🇺🇸 🇬🇧 - Jack & Kate Ask Each Other Questions in the 2nd, 3rd, Mixed Conditional - Duration: 11:31. 
For more infomation >> Как получать подарки от мужчин, даже если их не просить - Duration: 8:27. 
For more infomation >> The Young Offenders Go On A Road Trip | The Young Offenders - Duration: 2:38. 
Không có nhận xét nào:
Đăng nhận xét