Some of you may remember when Folding@Home was still in its infancy, the
first GPU client was introduced. Believe it or not that first (and
even second) GPU client only supported ATI graphics cards with
compatibility for NVIDIA’s CUDA being added much later.
Unfortunately, due to the Brooke programming language used in previous
generations of AMD GPUs, they consistently took a back seat to NVIDIA’s
CUDA supporting clients. GeForce graphics cards soon became the de
facto kings of the Folding world but with the introduction of AMD
graphics cores which support OpenCL, things may be about to change.
A few weeks ago, Stanford introduced the new Core 16 Project 11293 work
units which are specifically tailored towards OpenCL-supporting AMD
graphics cards and are available through the
GPU3 V7 Client open beta
program. In theory, this new project should represent just the tip of
the iceberg when it comes to the future of AMD’s OpenCL GPU Folding@Home
performance but being supported by a beta client means there are a few
small hoops to jump through. As such, the any GPUs prior to the HD 5000
series can't process these new WUs and there is a limitation whereby
the AMD video decoder can’t be used alongside this new client.
In this short article, we take this new client and its associated Core
16 project out for a spin while showing some comparative results between
past and present WUs. By no means is this meant as an end-all for AMD
Folding since the Stanford team is just getting started with OpenCL
support but it should give you a good idea about where things are going.
The beta V7 GPU3 client isn’t something many of you will be accustomed
to since it discards the usual interface for a design that’s a bit more
icon-filled. On the positive side it brings all of the console commands
to a user interface that’s user friendly but for many F@H users,
certain commands can’t be utilized like they used to. Most importantly,
this applies to the important –advmethods flag that must be used to
receive AMD’s new P11293 WUs.
In order to actually start receiving these new Work Units, before the
first project is downloaded (we recommend disabling network access until
this is done) ensure Expert is selected for the client, click the
Configure button and navigate to the Slots tab. Here you will need to
select the appropriate GPU (in our case gpu-0) and use the Edit button
to open an advanced configuration dialog box. In the last section, you
will Add an new slot with a name of “client-type” and a value of
“advanced” as in the picture above. This is used in place of the
–advmethods flag and will ensure that you can begin downloading the new
AMD WUs. However, since the new Core 16 projects seem to be in short
supply, you may not receive one right away.
One of the most important things to remember here is that since their
11.2 release, AMD has been including their APP SDK within their standard
driver installation package. This means once installation is complete,
any HD 5000-series and prior GPU should be ready to begin folding.
Just make sure you download and install the necessary APP components
from within the chosen Catalyst package.
So now that we’ve done that, let’s jump straight to performance numbers.
We used a very basic system with a stock i5 750 and 4GB of memory for
these results and took the average of four completed WUs for each
average result.
Before we go on it should be noted that our older Core 11 results were
really all over the place so averages were a bit hard to glean from the
numbers received. With that being said, with or without absolutely
spot-on PPD numbers for the older WUs the results speak for themselves:
the beta OpenCL project has done absolute wonders for AMD’s PPD across
the HD 5000 and HD 6000 series of graphics cards. In every case we are
seeing a near doubling of overall folding performance which is
impressive to say the least.
While things look quite rosy all around, performance seems to scale in a
linear fashion based on the number of Stream Processors on a given card
and the graphics engine’s clock speeds. For example the HD 6870 has
1120 cores and a clock speed of 900Mhz so it can hang with a 1440 SP,
725Mhz HD 5850. Meanwhile, the 1600 core HD 5870 edges out the 1536 SP
HD 6970. There also wasn’t much of a difference between identical cards
sporting different memory configurations. As the new Core 16 projects
mature, we’ll surely see things change a bit but for the time being it
seems like the older AMD cards are more than holding their own.
There are some additional points which we should mention that may have a
bearing on how some people will look at this new offering from the
Stanford team. Much like other AMD clients and projects, this one seems
to eat up excess CPU cycles and eats up a good amount of memory as
well. The toll on system resources amounts to about
25% CPU usage on a quad core processor and
a memory footprint of about 400MB.
Hopefully this will be addressed with future releases but most modern
systems should be able to take this kind of hit in stride and still be
able to offer a lag-free multimedia experience.
When it comes to power consumption, if you absolutely have to fold with
an AMD card the older HD 5870 with its 1600 cores currently looks like a
good bet from a PPD / watt perspective. However, due to their sky-high
PPD numbers when compared to the AMD cards, there are still several
lower end NVIDIA products that will be much more appealing as dedicated
folders.
Parting Thoughts
So there you have it; the new AMD Core 16 project seems to do wonders
for AMD’s Folding@Home performance. Currently there is only a single
project making the rounds but there are surely more in the pipeline and
what we’ve seen so far is only the first step down a very long road.
The overall PPD increase we saw was certainly impressive but when put
into context against NVIDIA’s current crop of GPUs (the relatively
inexpensive GTX 560 Ti sports PPD values of 14,000 to 16,000) there’s
obviously a long way to go before parity is achieved. Nonetheless,
with a single project the Stanford team has breathed new life into AMD
folding rigs. Let’s hope that more OpenCL-based projects continue to
tap the once-hidden power of AMD’s GPUs because our first taste of the
potential lying behind the scenes promises big things in the future.
Source:
Hardware Canucks
