With Crytek CEO Cevat Yerli promising that
Crysis 3 will "melt your PC" due to its uncompromising next-generation
visuals, I recently upgraded to Nvidia's GTX 670. Truth be told, I was a little
astonished when I realized my GTX 570 would be blowing out its second candle in
a couple months. It's done so well in recent games like Borderlands 2,
Dishonored and Far Cry 3 that I tended to forget its age. But
it was the arrival this year of brand new benchmarks such as 3DMark 13
and Unigine Valley that emphasized just how long 20 months can be in
the technology industry. Herewith is a look at the generational improvements in
moving from the GTX 570 to the GTX 670.
YIN &
YANG
"Kepler", Nvidia's code name for the
600-series, says as much about the GTX 670 as did the "Fermi" moniker for the
Three-Mile Island GTX 480. In 1601, noted scientist Johannes Kepler wrote,
"Where there is matter, there is geometry." Unlike Nostradamus, there's no way
Kepler had any inclination that four centuries later he would be named after the
fastest line of consumer graphics cards the world has ever seen. And while his
quote was initially applied to astronomy, it is equally apropos for the 102,480 Million texels per second generated by the GTX 670. He
would also not have understood that his namesake video card represents more than
just a die shrink from 40nm to 28nm. In fact, this new architecture allowed
Nvidia to pack more circuits in less physical area. In the GTX 570, Stream
Processors (SM) are the basic building blocks of the graphics pipeline. Inside
each Fermi SM block is 32 cores, so with 15 SMs the 570 had a total of 480
Stream Processors. But with the 670, Nvidia redesigned the Stream Processors
enabling a whopping 192 cores to fit inside one Kepler SMX. And the extra
overhead didn't stop there. Each Kepler SMX also benefits from its own
tessellation engine known as Polymorph which contains 16 texture units. These
improvements are why the 670 has so many more Stream Processors and Texture
Units than the 570.
The only items (on paper)
not significantly higher in Kepler are the Raster Operations per second (ROPs)
and the Memory Bus Width. Kepler’s ROP count has been reduced from the 40 ROPs of Fermi down
to 32 (eight per memory controller). However, this reduction is offset by
Kepler’s increased core clock speed. And assuming the basic ROPs are unchanged
between the two GPU architectures (which there's nothing to indicate otherwise)
32 Kepler ROPs at 980MHz are still faster than 40 Fermi ROPs at 732MHz. The same
goes for the memory interfaces-- Fermi's 320-bit memory interface has been
replaced with Kepler's 256-bit. Impressively, Nvidia has managed to push this
GDDR5 to excessively high frequencies, with the GTX 670's 2GB’s memory running
at 6GHz. This means that despite the reduced memory interface, the 670's
resulting memory bandwidth is still higher than the GTX
570.
MORE THAN MEETS
THE EYE
Additionally, the GTX 670
sports a new feature known as Bindless Textures. While previous Nvidia GPUs were
restricted to 128 simultaneous textures, Kepler removes that limitation and
allows up to 1 million simultaneous textures. Another great Kepler feature is
TXAA which stands for Temporal Anti-Aliasing. This technology will be quite a
boon for games as it improves the image quality without sacrificing performance.
With TXAA, the sampling pattern is spread out over multiple frames. By doing
this, the effect is better image quality than even 8x MSAA, but with a
performance hit similar to just 2x MSAA. Lastly, Kepler introduces Adaptive
Vsync which is basically an intelligent form of vsync. Without vsync, screen
tearing can be very distracting, but enabling it locks the refresh rate and
hurts performance. Nvidia's Adaptive Vsync ties the frame rate to the refresh
rate like traditional vsync until it detects a drop below the standard refresh
rate. Then it temporarily disables it until the framerate reaches the standard
refresh rate again. Nvidia claims this creates a much smoother overall
experience. These hardware improvements make Kepler a compelling purchase, even
without the raw performance boost over a GTX 570.
Remarkably, the 670 is nearly an inch
longer than the 570 it replaces, and as such necessitates that the dual six-pin
power connectors be moved from the end of the card to remedy clearance issues.
With an Antec 1000 Server case, I've never had to worry about a video card not
fitting, but I can certainly see how it might present a problem for smaller
enclosures.
In my July 2011 review of the 570, I
complained that PNY's installation disk contained drivers that were five-months
old. Sadly, the 301.42 drivers included for the 670 were left over from May 24,
2012, nearly ten months ago! I stress this point because updated drivers can
make or break a game. For instance, Nvidia rolled out brand new 314.07 drivers
optimized specifically for Crysis 3 that improve performance by up to
65%. I can't imagine anyone who pays $400 for a GTX 670 would load the drivers
from the disc, but if they do, they are needlessly sacrificing performance. I
was hoping at least that the disc included the "A New Dawn" program designed to
showcase the abilities of the 670, but I was not so lucky. Instead, I was forced
to download the 772 MB demo. Regrettably, the disc looks to be a carbon copy
that has been circulating since 2007 with only new drivers added every year or
so. Based on my experience with the 570, and now my 670, PNY builds solid
hardware. However, the department that cranks out the installation discs is
another matter all together.
BENCHMARKS
All tests were run on my system which
consists of Windows 7 Ultimate 64-bit with a Core i7 3770 overclocked to 4.0
GHz, GIGABYTE GA-Z77-D3H mainboard, OCZ Fatal1ty 750W power supply, OCZ Vertex 4
SSD and G.Skill 8GB PC1600 DDR3 memory. For both
cards, I used the newest 314.07 WHQL drivers from February 18th. Save for
Passion Leads Army, all benchmarks were released in February 2013
making them the most current and comprehensive available.
3DMARK 13: It wouldn't be
a complete video card test without throwing in 3DMark 13, the newest
installment of the world's most popular graphics benchmark. However, the testing
methodology is a little different than in years past with a three-tier
system that covers everything from mobile devices to laptops to dedicated gaming
machines. On the bottom rung is Ice Storm, a DirectX 9 test intended
for the most basic hardware such as integrated graphics. Cloud Gate is
more advanced with a DirectX 10 feature set designed for a discrete video
card. And Fire Strike utilizes next-generation graphics and physics
applicable for world-class PCs. In short, Fire Strike is as
"Future-Proof" as currently possible and absolutely crippled my 570 with
single-digit frame rates. Fortunately, my 670 fared much better with the
pyrotechnic benchmark scoring nearly 51% higher.
PASSION LEADS
ARMY: Released last summer, this is a DirectX 11 Unreal Engine
3 benchmark from Chinese developer Giant IronHorse. It takes the viewer on
a whirlwind tour of an ancient oriental city, complete with heavily-tessellated
cobblestone walkways and glass storefronts that seemingly explode at random. If
there's a message to this demo, it was clearly lost in translation. But like a
Godzilla movie, the plot is secondary and the real star is the DX11 eye
candy. In fact, the limited benchmark is so good it looks like it came from a
North American Triple-A studio and not some communist country situated on the
Pacific Rim. Unfortunately, I'm at a loss to explain the slim performance
discrepancy between the cards. Both the average and maximum frames per second
(FPS) reflect only a meager 20% improvement. As the other benchmarks illustrate,
this divide should be closer to 50%. Visually, PLA looks better on the
670 with more realistic lens flare and smoke, but whether this accounts for the
missing performance is still unknown.
RESIDENT EVIL 6
BENCHMARK: This benchmark is an anomaly in that while I appreciate the
complexity of the MT Framework engine, it's a game that I will never play.
Gruesome horror shooters such as Dead Island, Dead Space and
the Resident Evil series don't appeal to me. That notwithstanding, the
graphics are impressive given that Capcom designed this as a straight DX9 title,
an obvious nod to its console roots. However, the stress test comes from
rendering over 100 zombies on screen at one time in a scene reminiscent of the
World War Z trailer. Despite the older API architecture,
RE6 at 1920x1080 with FXAA 3HQ drags even the 670 down. Judging by
the proprietary score, the 570 is 56% slower here.
UNIGINE VALLEY: Released
within a week of chief competitor 3DMARK, this new iteration builds on
the popularity of its sibling product, Unigine Heaven, with an all-new
benchmark. Unlike Heaven's utopia in the clouds, Valley is
literally down to earth as it focuses on grasslands and heavily wooded forests
where environmental effects such as rain and wind tax even the most powerful
video cards. A few times my 570 briefly climbed above the 30 FPS threshold, but
overall it recorded just 25.5 FPS at the Extreme HD Preset (1920x1080
8xAA). With the 670 at the same settings, it averaged 38.1 FPS. As a whole,
the performance delta between both cards was 56%.
CONCLUSION
A couple weeks after I bought it, I was
casually checking some websites when the entire display suddenly blacked out. My
first thought was that we'd lost power, but after a few seconds it popped back
up with the message "The Nvidia driver driver stopped responding and has
recovered". I didn't think much about it until it happened again about five
minutes later. At this point, I was beginning to be concerned that I had a
faulty new graphics card. Thinking maybe it was a corrupted display driver, I
downloaded the latest Beta drivers released the day before. Initially, this
seemed to stop the problem until it returned thirty minutes later. Then, as I
was deleting the temporary folder created by the beta Nvidia drivers, I noticed
a new folder created right around the time I started having problems. I
discovered the folder was tied to a running process called IEHighutil.exe. This
process was causing my video card to overheat and when it did, it would stop and
restart. Unbelievably, I determined it was stealing my GPU cycles for Bitcoin
mining using Poclbm, a python-based OpenCL framework that quickly performs the
hashing computations. Fortunately, it was simple to remove, but more troubling
was how easily it snuck past my three anti-virus programs. The only place I can
believe I might have picked it up was when I downloaded the new Tomb
Raider game. Incidentally, the Tomb Raider reboot is fantastic,
but regrettably arrived too late to be included in the
testing.
Coming
from a GTX 570, I was concerned that the 670 might be more of a lateral move
than a true upgrade. Originally, I had wanted to wait until the new 700 Series
was available from Nvidia, but Crysis 3 changed all that. After
shelling out $65 for it, I was determined to enjoy it in all its CGI glory. That
meant buying a new video card and the PNY GTX 670 fit the bill. Unfortunately,
as the benchmarks illustrate, the 670 is not the leap forward I was looking for.
The continuing improvements in graphics and image quality have helped nullify
the corresponding hardware advances. As such, the perceived progress between
video cards becomes less than what we're accustomed to. As with any high-end
toy, it all boils down to having to pay if you want to play.