|
For the IT
professional, it pays to peek into possible futures on a regular basis.
With that in mind, we invite you to peek into the CPU looking glass at
what will be a radical change to the CPU as we know it. In fact, in a
few years time we may no longer recognize the CPU for what it used to
be - in fact, and at this very moment the very definition of a
"CPU" is being redefined. We may soon replace the term
"CPU" with the term "GPGPU", "HPU"
(Heterogeneous Processing Unit) or something similar.
 So
what is the GPGPU, and what could it possibly mean for the IT industry?
The answer to the first question seems simple enough: Take a CPU, and a
Graphics chip, and put them together. Of course things are never that
simple, so to answer that question properly we need to review some of
the recent history between GPU rivals, ATI© and Nvidia©.
We start off with the acquisition of ATI Technologies by AMD back in
October of 2006 for a notable $5.4 billion. With the move came the
announcement of a new initiative called "Fusion" that
promises the marriage of AMD's CPU technology with ATI's GPU offerings.
Soon, Intel© had responded to this move by AMD with an announcement of
a similar offering down the road. Intel is already the leading producer
of low-end graphics chips, but that won't make this new arena any
easier for them. With respect to performance, the graphics technologies
that ATI and Nvidia offer completely eclipse that which Intel currently
produces.
But this is more than the seemingly simple concept of taking a CPU and
putting it in the same chip as a GPU. Nvidia has been developing
extensions to the C programming language that it names CUDA (Compute
Unified Device Architecture), which allows any C programmer to write
programs for Nvidia GPU's in the same way that they do with the common
CPU. Thus far, CUDA has shown that a GPU can significantly outperform a
CPU in a variety of ways, particularly in the one area where current
CPU technology fails us - parallelism. Graphics technology, by its very
nature, is focused on solving the issue of parallelism at the silicon
level, whereas CPUs are serial in nature and focus on general purpose
processing. CPUs are jacks of all trades, with parallelism as an
Achilles heel. That is why mainstream processors have taken to using
multiple, yet primarily independent cores; the opposite cannot be
strictly said of the GPU. Most code can in one form or another be
parallelized and thus accelerated on a graphics card.
In past years, there has been much grumbling over a fundamental misunderstanding
of Moore's Law. Some have, incorrectly, claimed that Moore's Law had
been broken - and to the misinformed it certainly might have seemed
that way. After all, from the early 90's to around 2003, clock speeds
on mainstream processors were on a steady rise, but from 2003 to 2005
speeds only increased from 3 to ~3.8GHz. It seemed that the laws of
physics had put an end to advancement in this area - but we know that
Moore's Law is really all about how many transistors you can fit on a
die. So while clock speeds have not improved much, the number of
transistors on a chip has. Instead of increasing clock speeds,
engineers have been hard at work going sideways by adding additional
cores and in just a few years quad-core systems have become common and octet-core
systems are moving into the market.
Would it then surprise you to know that the AMD's latest and greatest
midrange GPU, the RV770 Radeon HD 4850 has, get this: 800, parallelized
and fully pipelined mini-cores called stream processors. The chip has
956 Million transistors, compared to the latest quad-core CPUs at 820
Million, and the latest offering from Nvidia which contains an amazing
1.4 Billion. Performance wise, AMD's midrange Radeon cards are far more
efficient and cost effective than the best offerings from Nvidia, which
is only able to offer marginal performance gains at a steep premium.
This puts AMD in the prime position to forge the way for HPU.
Operating systems are moving towards 3D navigation and touch screen
technology, and in order to keep up even increasing demand for
computational resources, computer systems will have to move away from
the bottom of the barrel graphics solutions offered by Intel. The
future will see the CPU swallowed up by the GPU as the processing load
on the typical computer becomes graphics centric and thus parallel
processing dependant. At first, AMD's Fusion project will see a general
purpose CPU fit into the same chip with a basic GPU processor, and is
being targeted for laptops to start - the benefits there being the
greatest. But as mentioned above, transistor counts for midrange GPU's
are already higher. It is only a matter of time before the majority of
any given CPU, is actually the GPU. Who would have imagined that the
red-headed step child of technology, the graphics chip, would one day
absorb the modern day workhorse? And of course, there is a tremendous
performance benefit to giving birth to the GPGPU - and that is where
the affect on the IT industry comes in.
And it's good news - very good news. This new breed of processor will
offer more bang for the buck than ever, and will ensure that desktop
performance will continue to advance. It will end the need to purchase
additional video hardware when users complain about their puny
integrated solutions. Just think of what you'll save in shipping costs
alone. And when graphics subsystems advance, you won't need to
purchase entirely new machines/motherboards because the graphics bus
will be on the chip itself. Developers will have a new and very accessible
toy to play with. Motherboard designs will simplify, become cheaper and
smaller - further reducing footprint. And if you think today's laptops
are small now, just wait and see. Over all, you'll see reduced overhead
for your helpdesk, and even though you might think this won't help your
IT backend (servers don't need graphics) you'd be wrong. Software
solutions will soon take advantage of the newly accessible raw
horsepower to accelerate every aspect of multitasking, because the
truly important thing to note is that with this move - the fancy
parallel graphics hardware we see today will finally be useful to more
than just gamers and art departments.
The impact on virtualization alone is worth noting. The one major flaw
in the area of virtualization is that while a Host server may have 8 or
more CPUs, each virtual machine has to share what graphic resources are
available - and usually that's just one graphics card for multiple
Guest machines. Never mind the fact that most Host machines come with
bare bones graphics, and that virtualization relies heavily on CPU
resources for the rendering of graphics. With HPUs, each processor will
contain its own graphics core, eliminating the current bottleneck and
allowing Guest machines to run more than just business apps. For post
video production, having a graphics core in each processor will be a
boon to the entire industry. All the efforts made over the last 30
years to advance silicon technology for the purpose of entertainment
(Pong, Super Mario and Halo) will finally bare the kind of fruit that
the business world can absorb.
With the advent of the HPU, GPGPU, or whatever you choose to name it,
the phrase "system on a chip" comes one step closer to
reality for the desktop and server market. And that is certainly worth watching
unfold.
|
