|
Vrzone.com Article: 64 Shader Processors: NVIDIA Geforce 9600GT
Sixty Four in a Sixty Five Shaking off the old fats of the G80, NVIDIA has successfully shrunk the G80 into the G92, the latter of which spawned the 8800GT, 8800GS and 8800GTS 512MB graphics accelerators. With us today is the Geforce 9600GT, based on the G94 graphics acceleration chipset (specifically the G94-300). G94 is wafer-ed by Taiwan Semiconductor Manufacturing Company (TSMC) using the 65nm process. (TSMC currently offers fabbing down to 45nm)
Stacko! Within the 65nm die lies some 64 shader processors, which by default pushes data around at 1750MHz. The graphics core is clocked at 650MHz, which isn't quite a big number by today's standards. NVIDIA's partners do know that too; out of the four candidates with us today, no less than three are factory-clocked above mainstream specifications. Memory wise, we're still looking at 1ns gDDR3 BGAs clocked at 900MHz on a 256 bit memory bus. You can be sure NVIDIA didn't shortchange users on memory bandwidth this time around. Below is a comparison chart between the vanilla 9600GT specifications, and that of a few other recent options floating around the midend market.
By now, it should be apparent that the G94 chipset is poised to replace the G84, more commonly known as the GeForce 8600 series. G84 had just reached it's end-of-line status on Valentines Day about a week ago. As usual, the G94 9600GT will be priced according to the performance hierachy of the GeForce family. NVIDIA AIB partners meanwhile are in a rush to clear their (large) inventories of last-gen G84 graphics accelerators to faciliate the introduction of the 9600GT.
Losing balance... Previous generations of midend DX10 graphics accelerators didn't quite cut it for practical DX10 gaming. Gamers on a budget are hoping (or rather, expecting) either of the two graphics giants to come in with new offerings to fill the performance gap. In the past months, we've heard immense bitchings about the 8600GT/8600GTS's measly 32 Shader Processors (SP) on the G84. Simply said, the 8600GT/GTS ain't quite as fast as what consumers expected midend graphics accelerators to be. At that point in time, the G80 8800GTS/GTX already had 3-4 times that number of SP at hand. The G94 happens to be NVIDIA's answer to gamers' prayers; beefed up to 64SP, the 9600GT drops neatly onto the middle rung of the performance ladder, somewhat below the dearer 8800GT. From Titans to Mutants Unlike previous generations of graphics accelerators (think Radeon 9600XT, GeForce 4 MX) where creativity breeds competitive designs, today's graphics accelerators are all-too-often OEM offsprings. The dawn of a new chipset for NVIDIA will not attempt to stray from this safe trend; three of the four 9600GT graphics accelerators with us today are found with the P545 reference PCB design, complete with spoonfed component choices and layout.
Specifications wise, the 9600GT cards with us today share much similarities. Clockspeeds and features would decide the desirablity of these graphics accelerators. We've prepared a 7-way shootout for these cards. How they do on paper is shown in the above table. Like the 8800GT, the 9600GT comes with a 2-phase VRM, we will touch further on that later in the review. More on setup configuration in the following CPU-Z dump:
CPU cooling was provided by a Swiftech Apogee GT looped in with a 2.120 heatercore and a LAING DC pump. Seagate 7200.7 80GB harddisks provided storage. Operating System was Microsoft Windows Vista Ultimate x86 with ForceWare 174.13 drivers. References to P545 - Inno3D & XFX As mentioned, we will be testing three 9600GT cards based on the P545 PCB in this review. Two of them are from Inno3D, and one from XFX. All three sport different BIOSes, yet share the same PCB design and components onboard. Initial problems with the 9600GT (outlined here) appear to have been ironed out (no pun intended, soldering iron wielders), so the P545 boards are "good to go". In subsequent pages, we'll be grinding the cards under our millstone, one by one.
The lineup, ready for an onslaught of tests. Inno3D 9600GT 512MB & "Overclock"Inno3D was rather excited about sending us samples of their GeForce 9600GT graphics accelerators, excited enough to drop not one, but two samples of their latest with us. One of the two is overclocked, whilst the minimalist theme sticks.
The Inno3D boxes. The overclocked is distinguished only by the "OVERCLOCK" sticker.
Packaging is simple, but sturdy. I particularly like the way the card is held within the box by foam supports.
Provided are minimalist accessories.
The Inno3D cards with different overlays. XFX 9600GT 512MB "XXX" EditionDespite sharing the same PCB design, the XFX is marketed in a totally different manner from the two Inno3D graphics accelerators. The gamer-oriented theme sticks as you ruffle through the package.
Antistatic Teaser.
Provided accessories.
Instructions, drivers and a hound to keep you isolated while you frag. Company of Heroes is included with the XFX Card.
So wicked-sick, you can't tell the nine from the six. A Peek Under The Oddball XpertVision's 9600GT is the most unique of the lot, offering a myriad of display connectivity options and a different PCB design. 2 Dual Link DVI ports, a DisplayPort and a HDMI 1.3 port. The stock cooling solution is a striking heatsink-fan resembling a Thermaltake Golden Orb of yore.
Fully Functional Familiar Funky Fins.
Myriad of display options. Brands offering the same PCB will sport identical connectivity options.
VRM comparison between Palit's 9600GT and 8800GT. Both use the RT8802A PWM IC for GPU voltage regulation. We've got similar cards with 3 phase power from Gainward and Vvikoo as well.
Tomb Raider is included in the package, along with a couple of VGA converters and an S-Video Cable. Full Monty - Visually Molesting The P545
P545 VS P393 (8800GT).
9 inches long. Have something longer?
This 6 pin power entry soaks up much of the 9600GT's 95W.
S/PDIF pass-thru for HDMI.
G94 in full naked glory.
1ns Samsung RAM.
Analog Devices' VRM chipset.
Mysterious memory VRM.
Vgpu in 3D. Vmem hovers around 2V. DirectX9 & Before - Benchmarks, Games
Synthetic BenchmarksBefore we begin, the system specs as below
Industry watchers tend to place much emphasis on synthetic benchmarks from Futuremark, and they do very well in pulling apart minor differences between graphics accelerators. 3DMark01 is running out of favour amongst enthusiasts due to the ever shifting dependence on system power.
3DMark01 may no longer be as graphically bottlenecked as it used to be, but it shows the overall system performance, especially when it comes to driver overheads and texture/memory management on, and off the graphics accelerator.
3DMark03 remains responsive to graphical power scaling even today, largely thanks to GT4: Mother Nature. The mix of DX7 and DX8.1 graphics goes on to display compatibility.
Prior to 3DMark Vantage's debut, 3DMark06 continues to be the de-facto synthetic benchmark for bottlenecking graphics accelerators. Here, we see the XFX and XpertVision/Palit cosy on the framerates. Coincidental? We'd see.
'lil Gaming
Team Fortress 2 runs on the wildly flexible Source engine. We ran the incredibly raucous TF2 at 1600x1200 on below settings:
Notice how the the "X" lines up? Incredibly similar in performance! DirectX10 - Game Benchmarks + OpenGL
More Games!
What use is a new graphics accelerator looping benchmarks? Here's a few of the more graphically intensive games on the market right now. Crysis was run at 1600x1200 on Medium settings throughout.
A while ago, we noted how NVIDIA was working on drivers to improve DX10 performance here. The work appears to have paid off, as the drivers worked with the new 9600GT to bring about playable framerates on Crysis. Such a feat used to be handled by the G94's faster (and more expensive) kin, the G92. 1600x1200 resolution was the order for the day, in Shangri-la at least. Texture and Character details were both set to 5.
Not bad at all! All the 9600GT manages the resolution with aplomb to create playable conditions in the fast paced game. Another tie between the "X" cards. Next up? OpenGL! 1600x1200 Medium settings in Enemy Territory: Quake Wars.
Similar settings for ETQW at 1920x1200, just a comparison between the 9600GT. ETQW is definitely playable.
UT3 with details at 5, smooth as silk.
Crysis's fps are running low, still playable even thought a tad low..
Device Thermals - TM62 Exposed The 9600GT earns itself a new thermal solution known as the TM62, engineered by none other than the guys at CoolerMaster. Rather well built, we dissected the TM62 for a shot or two. Victim? Inno3D's vanilla 9600GT cooler. As expected, the cooler is pretty much similar to the 8800GT's stock cooler.
Heatpipes are soldered to the fins, which are angled to dump air diagonally. We're glad to see that NVIDIA has included a large diameter axial blower for the 9600GT. After the noise problem in single slot coolers for early 8800GT releases, they'd be damned not to have learnt their lesson. With PWM control of the blower, NVIDIA has succeeded in proving the TM62 an effective, relatively quiet thermal solution.
31 fins line up on the TM62 cooler.
There exists room for improvement in thermal interface. A better contact would show little heatsink compound residue.
GPU die is heatpipe cooled, common in current generation midend and highend graphics accelerators.
I'm pretty sure Merlin has retired, because these days, Magic is actually Made-in-China. Cooling AssessedBelow is the temperatures logged over three 3DMark06 loops. Highest temperatures recorded are shown, representative of the peak load temperatures experienced by the GPU onboard thermal senseor. Please ignore clockspeed readings as they are misread in the monitoring window of RivaTuner at the moment. Dynamic Fan Control
Full Fan Speed
Not surprisingly, the winner was the Sonic with it's two-slot cooler. Notice that even amongst the reference designs, temperature differences exist. Silicon tolerance plays an important role deciding how hot things run. Tweakin' n Fiddlin' - Overclocking Our readers should be familiar with the art of overclocking, all the more so if you're the sort who have been following the VR-Zone kind of hardware guides and reviews. No one is bothering to advise bullcrap about voiding warranty these days, because it has already been done, right at the start even before these graphics accelerators find themselves packed. Gone are the days where raising a MHz seemed like a ticket to zero-support. The 9600GT is a new graphics chipset that has yet to be supported by the majority of overclocking tools. You could use nTune for it, but options would be limited. To fiddle more with these 65nm creatures, you've got to download RivaTuner. In it's 2.06 guise, you won't get to recognise the 9600GT until you drop this configuration file (right click, save as) into the installation folder. Unlinked clocking allows for more play when it comes to maximising performance. All clocks tested passed 3DMark06 with no visible artifacts
It may seem that we're awarding our overclocking trophy to the vanilla Inno3D, but in truth, too much variables exist for there to be a clear winner. Much thought has been put into the design of the Sonic, which could possibly improve overclocking performance. However, the extra phase of power would only prove useful to overclockers who intend to modify Vgpu voltages beyond two phase limits. The third phase unfortunately compromises efficiency. 30' of Pleasure For many PC users, work is paramount, and gaming remains secondary. Those who run a large display for intensive multitasking, and only play the occassional game, may not justify the purchase of a highend grahics accelerator for the occasional 3D. Here is the type of settings that can be run on a NVIDIA GeForce 9600GT (a generic one like the Inno3D) on Dell's 30 inch display. Unreal Tournament 3Left click captured frames (settings/image quality) above to open them in a new window at 2560x1600. With the above settings, timedemo ran at 50FPS, which was rather comfortable for such a fast paced game. The test was run in Shangri-La where lighting and foliage posed rendering challenges for the 9600GT. Even so, it was possible to run at pretty high detail settings. Not too shabby for native resolution. Crysis
Left click captured frames (settings/image quality) above to open them in a new window at 2560x1600. Crysis is understandably stressful on current generation graphics accelerators. With help of new generation drivers and the updated patch, performance was reasonable as long as users do not expect too much out of the Crytek Engine. Running anything less detailed than the above settings made Crysis look terrible. DX10's supposedly lower overheads need some real-world tweaking, at least on Crysis. Enemy Territory: Quake Wars
Left click captured frames (settings/image quality) above to open them in a new window at 2560x1600. ET: QW ran pretty on its Doom 3 Engine, although it wasn't exactly fast on the 9600GT. NVIDIA's stronghold on OpenGL didn't shine here, but then again, there was 12xAF applied. The tradeoff was easily justifiable, for images gained a sharper character. Happy Ending The G94 OutlookAs a graphics acceleration chipset, the G94 has done much to take over the lower end of the market segment that has for the past month, been dominated by G92 graphics accelerators like the 8800GTS and 8800GT. Despite the lower number of Shader Processors, the G94 does a fair job with rendering, easily matching up to the G92 GTS and trouncing the midend of ATi's lineup. Performance deserves a thumbs up as NVIDIA puts out another value winner. The solemn sounding die-shrink is the rosy answer to higher (potential) clockspeeds at lower power consumptions. A job well done? You bet! ImplementationThe design of the P545 remains puzzling to us, and unlike the performance, isn't something we'd like to praise. The graphics industry, it appears, has been covering more grounds than it can cope. More often than not, this issue affects overclockers the most, as they ponder over the gradual decrease of overclocking headroom on the graphics buffer.
Recommended memory voltage is 1.9V, as read from Samsung's datasheet for the K4J52324QE-BJ1A. As with the past few generations of gDDR3 graphics accelerators, NVIDIA's P545 reference design has been subjecting memory ICs to voltages higher than that recommended by the memory manufacturers themselves. Running a memory IC at higher than specified voltages improves stability for a given clockspeed, saving costs of implementing a higher specification part. This offers more performance per dollar in most instances, but preys upon the unsuspecting consumer. A moral dilemma for NVIDIA engineers juggling the two 'B's; Budget, and Bandwidth.
Allowing for meter-error, default 9600GT memory voltage still falls beyond Samsung recommendations. Inno3D 9600GTThe Inno3D 9600GT, as the epitome of the plain-Jane, vanilla edition 9600GT, forms the most value-for-money option for a unique bunch of users: those who either don't believe in factory-overclocks, or those who know what to do to get the most out of the G94, all by themselves. The package may be simple, but remains startingly effective and reasonably attractive. The masked motif may elicit terrorism alerts in some parts of the world, but you can't go wrong with this choice. The fact that it clocked the highest, and scored the best in our overclocking roudup only goes to show how marketing precedes actual speedbinning in the real world. Inno3D 9600GT "Overclock"Another offering from the same enthusiastic manufacturer, the "Overclock" is a higher clocked edition of the reference design. Unfortunately, after their half-hearted tweaks, performance still falls a little below the two "X" brand offerings. If you do not intend to separately overclock the graphics accelerator, you'd get better performance from the remaining two overclocked candidates. In the meantime, hold your horses and watch for much more promising iChill editions. For the price, the Inno3D "Overclock" remains a winner.
XFX 9600GT XXX EditionBranding has always been a strength of XFX, with the 9600GT XXX Edition being a case in point. Packaging is unique, with an almost-demonic "9" forming the theme for it's box and overlay. Amongst the bunch, it has one of the best out-of-the-box performance. The bundle is pretty cool too, with relevant accessories for the g33ks. PETA need not worry, for no Alpha Dogs were hurt in this review. 9600GT SonicThe only non-reference design in the bunch, the Sonic is one of the best performers, on par with the XFX. The dual-slot heatsink design proves effective in keeping heat down, but at the expense of an expansion slot. The extra phase of GPU power may prove useful in voltage modifications, but the less advanced design sacrifices some efficiency. A simple but effectively speedy package. Choosing between the XFX and the XpertVision can become the most difficult decision of your life. Private MomentsAs usual with pre-release material, a lot of going-ons behind the scenes are less-than-perfect. Here's some bloopers for you readers, as we end off the review.
Extra speed for the colour blind, taken during one of our overclocking attempts.
The first ever prototype NVIDIA chipset? Maybe we've dug up undocumented history. =P
|