Two weeks ago NVIDIA formally launched the retail GeForce GT 640, catching up to their OEM and laptop offerings with their first GK107 based video card for the retail desktop market. GT 640 is designed to be NVIDIA’s entry-level Kepler video card, joining several Fermi rebadges as the members of the GT 6xx series. With the enthusiasm behind Intel’s Ivy Bridge in the laptop market and the boost in sales it has provided for NVIDIA’s mobile GPUs, NVIDIA is hoping to accomplish the same thing in the desktop market with GT 640.

Today we’ll finally be taking a look at the GT 640 in action. We’re expecting NVIDIA will launch a GDDR5 variant at some point, but, for the first round of cards, GT 640 is exclusively DDR3. This has important performance repercussions. Meanwhile, as is common with entry-level video cards, there is no reference design intended for retail sale and NVIDIA isn’t sampling any such card. However, NVIDIA’s partners are stepping up to sample cards to the press. Our sample comes from Zotac, who sent over their single slot based Zotac GeForce GT 640.

	GTS 450	GT 640 DDR3	GT 630 GDDR5	GT 630 DDR3
Previous Model Number	N/A	New	GT 440 GDDR5	GT 440 DDR3
Stream Processors	192	384	96	96
Texture Units	32	32	16	16
ROPs	16	16	4	4
Core Clock	783MHz	900Mhz	810MHz	810MHz
Shader Clock	1566MHz	N/A	1620MHz	1620MHz
Memory Clock	3.6GHz GDDR5	1.782GHz DDR3	3.2GHz GDDR5	1.8GHz DDR3
Memory Bus Width	128-bit	128-bit	128-bit	128-bit
Frame Buffer	1GB	2GB	1GB	1GB
GPU	GF106	GK107	GF108	GF108
TDP	106W	65W	65W	65W
Transistor Count	1.17B	1.3B	585M	585M
Manufacturing Process	TSMC 40nm	TSMC 28nm	TSMC 40nm	TSMC 40nm
Launch Price	$129	$99/$109	N/A	N/A

Diving right into the GT 640’s specifications, this is the same GK107 GPU as the 640M and other laptop/OEM GK107 products, so we’re looking at the same basic features and specifications. GT 640 ships with all of GK107’s functional units enabled, which means 384 CUDA cores organized into 2 SMXes sharing a single GPC. Further attached to that lone GPC is a pair of ROP blocks and memory controllers, giving the GT 640 16 ROPs, 256KB of L2 cache, and a 128-bit memory bus. All of this is augmented by the common features of the Kepler family, including the NVENC hardware H.264 encoder, VP5 video decoder, FastHDMI support, and PCIe 3.0 connectivity.

Thanks to the fact that this is a retail desktop product the GT 640 will ship at a fairly high clockspeed of 900MHz, which puts its clockspeed ahead of its OEM DDR3 counterpart but behind the GDDR5 version. Like the laptop and OEM versions boost clock is not present, so performance is rather straightforward here. Similarly, for those of you looking to make Fermi comparisons, GT 640 and other Kepler based video cards do away with the separate shader clock in favor of additional CUDA cores, so GT 640 has a much lower shader clock and far more CUDA cores than its predecessors.

Unfortunately things don’t look nearly as good on the memory front. NVIDIA is launching with DDR3 here, which means that even with the 128-bit memory bus this card is extremely memory bandwidth starved. At just shy of 1.8GHz it only has 28.5GB/sec of memory bandwidth. DDR3 versus GDDR5 has been a recurring issue in this market segment, and as both GPU performance and GDDR5 performance have increased over time the gap between DDR3 and GDDR5 card variants has continued to grow. By the time we’re up to this many ROPs and shaders the memory bandwidth requirements are simply enormous. In traditional fashion these DDR3 cards are outfitted with more memory overall – the DDR3 GT 640 ships with 2GB – so it has a memory pool every bit as large as the GTX 680’s but lacks the memory bandwidth to make effective use of it. So expect the obligatory 1GB GDDR5 version to be much faster here.

As for physical specifications, the official TDP on the GT 640 is 65W, putting this card distinctly into the PCIe slot powered category. Idle power consumption on the other hand is spec’d at 15W, which at least on paper is actually a bit worse than GT 440 and competing 28nm cards. Meanwhile the die size on GK107 is 118mm2, virtually identical to the 116mm2 die size of GF108. Product naming aside, due to the similar TDP and GPU die sizes the GT 640 is clearly going to be the direct successor to the GT 440 from a hardware perspective.

All things taken into account, GT 640 (or rather GK107) is a rather powerful video card given the entry-level market segment it will be occupying. GT 440 (GF108) only had 4 ROPs so NVIDIA is no less than quadrupling their theoretical pixel throughput here. At the same time the CUDA core count is greatly expanding thanks to the smaller manufacturing process and Kepler architectural changes, and after compensating for those architectural changes NVIDIA has effectively doubled their shading and texturing performance. A healthy boost in the number of functional units is of course typical with any new manufacturing process, but because NVIDIA pared down Fermi so much for GF108 the additional functional units on GK107 should significantly improve performance.

NVIDIA’s official guidance on performance is that GT 640 should come in just behind GTS 450, which makes sense given the similar theoretical shader and ROP performance of the two cards. Ultimately we wouldn’t be surprised if a GK107 card surpassed GTS 450 thanks to the former’s higher clockspeeds, but that will have to wait for a GDDR5 GT 640 or something similar. As it stands the DDR3 performance handicap will keep GT 640 from catching up to the GTS 450, even with the clockspeed advantage. Consequently while NVIDIA is pitching GT 640 as a solid performer at 1680x1050 it can really only achieve this at lower quality settings. So for most of our readers GT 640’s real strength is going to be HTPC usage thanks to its combination of video features (VP5/NVENC), ample shader performance for post-processing, and its sub-75W TDP.

Wrapping things up, let’s quickly talk about pricing and availability. NVIDIA’s goal for the GT 640 DDR3 was for it to be a $99 card, however with their partners free to design their own cards and set their own prices, they have for the most part not gone along with this. Currently only a single GT 640 is available at Newegg for $99, with everything else (including the Zotac card we’re reviewing today) going for $109. The better news here is that unlike the GTX 670/680 availability shouldn’t be an issue here as GK107 is far easier for NVIDIA to produce in volume even with TSMC’s capacity constraints. Cards have readily been available for over 2 weeks now and that’s not expected to change.

Finally, because of its de facto price of $109 the GT 640 DDR3 is in direct competition with AMD’s equally priced Radeon HD 7750, along with last-generation cards such as the GTS 450, GTS 550 Ti, and Radeon HD 5750. Unfortunately for the GT 640 it’s the only card packing DDR3 at this price, so it should come as no great surprise that its performance is significantly lagging its competition. As we alluded to earlier, any significant success for GT 640 is going to have to rely on its role as a sub-75W card, where it’s one of the more powerful cards in that segment.

Spring 2012 GPU Pricing Comparison
AMD	Price	NVIDIA
Radeon HD 6870	$159	GeForce GTX 560
Radeon HD 6850	$139
Radeon HD 7770	$129
	$119	GeForce GTX 550 Ti
Radeon HD 7750	$109	GeForce GT 640/GTS 450
Radeon HD 6750	$99

 

Meet The Zotac GeForce GT 640 DDR3

As we mentioned in our introduction NVIDIA is not sampling a reference GT 640, leaving that up to their partners. As always NVIDIA’s partners have spread out with a variety of designs ranging from low-profile cards to relatively large double-slot cards, with a few even trying their hand at factory overclocking. Today we’ll be taking a look at Zotac’s GeForce GT 640, a design that’s right in the middle of those extremes and fairly close to NVIDIA’s internal reference design.

Taking things from the top, at this moment Zotac is in a rather unique position with their design. They currently offer the only single-slot GT 640 available at retail, with every other card being a double-slot card to fit a larger cooler. Consequently for users with ITX cases where space is at a premium, Zotac’s GT 640 is generally going to be the only option.

Because the Zotac GT 640 is a single-slot card, Zotac’s cooler of choice is a fairly wide but shallow aluminum heatsink that spans roughly half of the card. At the center of the heatsink is a fairly small 2pin 55mm fan, which provides the necessary airflow to keep the card cool. The card’s DDR3 RAM is sitting underneath the heatsink, but no contact is made nor is it necessary due to DDR3’s very low operating power and temperatures.

Removing the heatsink we see the bare PCB, with the GK107 GPU and DDR3 mounted on it. Physically GK107 is an unremarkable GPU and the entire package appears to be roughly the same size as GF108’s GPU package. On this Zotac card it’s paired with 8 2Gb Micron DDR3-1866 modules operating in 16bit mode, which means for casual overclocking there should be a bit of headroom, but GPU DDR3 is rarely capable of going much past its rated speeds.

The rest of the PCB is solid yet simple; users worried about coil whine will be glad to see that Zotac is using solid chokes here as opposed to ring chokes, and we didn’t notice any coil whine during operation. The card measures 5.75” long – the minimum length necessary for a PCIe x16 card – so it should fit in virtually any full-profile case.

Meanwhile looking at the ports on the card’s bracket, we find NVIDIA’s favored configuration of 1 DL-DVI-I port, 1 DL-DVI-D port, and a mini-HDMI port. As one of the major improvements in the Kepler family NVIDIA now has more than 2 display controllers on their GPUs, so the GT 640 can actually drive all 3 ports simultaneously. You likely wouldn’t want to use the GT 640 for gaming but it’s certainly powerful enough for desktop work, and this is one of the few situations where that extra 1GB of VRAM might come in handy.

Top: Zotac GeForce GT 640. Bottom: NVIDIA Reference GeForce GT 640

Unfortunately the display ports on Zotac’s GT 640 also expose its one flaw, and it’s a big one. On NVIDIA’s reference design the mini-HDMI port is centered at the middle of the card, similar to the DVI ports. However for reasons unknown to us, Zotac has moved the mini-HDMI port on their GT 640 down by about 2mm. This doesn’t sound like much, but by putting the mini-HDMI port so close to the edge of the card it introduces a crippling flaw: it doesn’t leave any room for a cable to attach to it.

Zotac GeForce GT 640 Installed. Note the lack of clearance around the mini-HDMI port

Specifically, because the port is so low it’s right on the edge of the usable area of the bracket, as everything below the port will be covered by the I/O shielding of the computer case. Consequently if you attempt to plug in a mini-HDMI cable or adapter, the boot of the cable will run into the case’s I/O shielding before the cable is fully inserted, preventing the cable from getting a good connection and/or locking into place. The HDMI specification is actually rather strict about the size of the boot on mini-HDMI cables/adapters, and after having tested a few different adapters everything we’ve encountered is within spec, so this is poor planning on Zotac’s part. NVIDIA’s reference design and cards similar to it do not have this problem since if the port is properly centered it leaves plenty of space for the boot, which is why this is the first time we’ve run into this issue.

We’ve already brought this up with Zotac and they’ve told us that they intend to fix it once they’ve exhausted their current supply of brackets and mini-HDMI connectors, but for the time being all of their GT 640 cards will have this flaw. In the meantime the problem is not unworkable – with enough tampering it should be possible to force a mini-HDMI cable/adapter in there – but Zotac really shot themselves in the foot here by making the mini-HDMI port so inaccessible. On that note, if you do intend to take advantage of this port you’ll need to bring your own gear (Zotac doesn’t provide a mini-HDMI adapter), and you’ll want to either use a cable or a more specialized mini-HDMI-to-HDMI adapter. The stubby adapter Monoprice and most other retailers carry won’t work because the port is so close to the top of the bracket, which has been a recurring quirk with NVIDIA cards since NVIDIA started using the mini-HDMI port.

Moving on, rounding out the package is the typically bare minimum collection of extras found on budget cards. Along with a driver CD and quick start guide Zotac includes a DVI-to-VGA adapter, but not a mini-HDMI adapter. Zotac typically bundles Ubisoft games with their cards but on a budget card like this that isn’t really possible, so the GT 640 comes with the next best thing, which is a 3 day trial of TrackMania 2.

Finally, as we stated earlier the Zotac GT 640 is currently priced at $109 at Newegg and most other retailers, which makes it on-par with most other GT 640 cards. Meanwhile for the warranty Zotac is offering a base 2 year warranty, which is extended to a rather generous full limited lifetime warranty upon registration of the card.

HTPC Aspects : What is New?

We covered discrete HTPC GPUs in detail last year, and noted that the GT 520 was the sign of interesting things to come from NVIDIA.

The GT 520 was the first GPU from NVIDIA to come with support for VDPAU Feature Set D (also called as VP5 by some). VP5 is faster than VP4 and also brings support for decode of 4K x 2K videos. Unfortunately, the GT 520 didn't have the necessary hardware to output 4K resolution over any of its video outputs. The number of shaders in the GT 520 was also too low to support advanced deinterlacing algorithms. On the whole, despite the updated video processing engine (VPU), we couldn't recommend the GT 520 as the ideal discrete HTPC GPU.

All our concerns were supposed to be laid to rest with the launch of the Kepler series. NVIDIA started off at the high end with the GTX 680, a card which couldn't be called HTPC-friendly by any stretch of imagination. The more HTPC-friendly GK107 did see a simultaneous launch, but in mobile-only form as the 640M.

Given the configuration of GK107, it appeared likely that a desktop version would tick off all the boxes necessary for a HTPC. Does the Zotac GT 640 fulfill our expectations? The short answer is: Yes, it does! It improves upon the performance of the GT 430 with respect to madVR, thanks to the extra computational power and memory bandwidth. Meanwhile the updates to the video outputs (HDMI PHYs) and the retention of the VPU from the GT 520 enable decode and display of 4K videos in their native resolution.

However, these updates don't mean that NVIDIA's GPUs are perfect for HTPCs. Just like every other HTPC GPU vendor out there, NVIDIA has a list of things which need to be fixed from a HTPC perspective, which we'll dive into in a moment.

Another important update present in the Kepler series is the on-board H.264 encoder. The practice of integrating a H.264 encoder in the GPU was started by Intel in Sandy Bridge. While Intel has the second generation version of QuickSync in the Ivy Bridge processors, AMD and NVIDIA are just now starting to ship their first generation encoders (VCE and NVENC respectively).

Both VCE and NVENC are yet to gain widespread support amongst the software vendors, and NVIDIA themselves indicated that full support for NVENC in CyberLink's and ArcSoft's offerings would be coming sometime next month. Keeping this in mind, we have decided to postpone NVENC coverage to a later date.

In the next few sections, we will look at the HTPC aspects of the Zotac GT 640. Before delving further into that, the details of our testbed are provided below:

Note that we used three different HDMI sinks for our testing. While the fancy Sony VPL-VW 1000ES was used to test out 4K resolution output, the Sony KDL46EX720 + Pioneer VSX-32 was used to verify HD audio bitstreaming. The rest of the tests (including HQV benchmarking) were performed with the Acer H243H monitor.

HTPC Aspects : 4K Decode and Display

From a HTPC perspective, GPUs over the last two generations have done little to tempt users to upgrade since HD audio bitstreaming became a commodity feature. The appearance of 3D TVs called for some updates from the GPU vendors, but the technology didn't really pick up at the pace that the industry wanted it to.

With the 3D craze having been milked dry, it is now the time for a new buzzword: 4K. Retina displays have become the focus of much talk, thanks to Apple's promotion, and, at Computex, we saw the introduction of products with 11" and 13" screens having 1080p resolution. It is not tough to imagine 4K resolution panels becoming commonplace in 32" and larger sized TVs and even 24" and larger sized monitors.

The one aspect that 4K has going for it is the fact that the higher resolution (when it comes to videos, at least) is unlikely to have any ill effects on the viewers' health. Unlike 3D (which caused discomfort to a number of consumers), we expect 4K to have a much smoother sailing in gaining acceptance in the marketplace. In addition, 4K is the natural step towards a more immersive experience. As such, we are more positive about 4K from a consumer as well as industry perspective than we ever were about the 3D initiative.

In terms of being an early adopter, the current issue with the 4K ecosystem is the fact that HDMI officially only supports up to 4096 x 2160 @ 24 Hz and 3840 x 2160 @ 30 Hz. For a smooth desktop experience at 4K resolution, it is imperative that we get 60 Hz refreshes at 4096 x 2160. This is scheduled to come in the next update to the HDMI specifications. It is also unfortunate that we are restricted to 4096 x 2160 as the maximum resolution, when the official 4K cinema specifications are only slightly larger at 4096 x 2304.

In any case, the Zotac GT 640 that we are looking at today is compliant with the current HDMI 4K specifications. The 4K resolution is available over all the three ports (using an appropriate DVI to HDMI converter). Both the DVI ports are also capable of carrying audio.

AMD's GCN lineup is also compliant with the HDMI 4K specifications, but it is the GT 640 which has excited us enough to talk about this in detail. While AMD's 4K hardware decode remains an unusable feature for the general consumer right now, NVIDIA's Kepler implementation fares much better.

Due to the aforementioned issues with the mini-HDMI port on Zotac's card, we tested out the 4K output over the dual-link DVI port connected to the Sony VPL-VW 1000ES through a DL-DVI to HDMI adapter.

Currently, native DXVA mode implementations tend to crash the system. However, using LAV Filters 0.50.5 in the CUVID mode or DXVA2 Copy-Back mode, we are able to decode H.264 streams with resolutions greater than 1080p using the GPU.

 

The screenshot above (click on the picture for full 4K resolution) shows the playback of a 4096 x 2304 H.264 stream at 24 fps.

We see that the GPU's VPU has approximately 60% load. EVR-CP doesn't load up the GPU core too much (less than 50% core utilization). Note that the maximum refresh rate possible at 4096 x 2160 is only 24 Hz, as indicated by the EVR-CP statistics. Another point to note is that the LAV Video Decoder is operating in CUVID mode.

CUVID acceleration is also possible for videos with arbitrary resolutions greater than 1080p. The screenshot below (again, click for full Quad FHD resolution) shows flawless decode acceleration of a 3412 x 1920 video at 25 fps. At 3840 x 2160 (Quad FHD), the GPU is able to drive the desktop with a refresh rate of 29.97 Hz. In this case, the VPU load is a bit lower (around 45%) as per expectations.

 

How well does 4K decode and rendering work with other combinations of decoders / renderers? The usage graphs below present the CPU package power, GPU core load, memory controller load, video engine load and video bus load when playing our 4K test clip (original version of this YouTube video) on a 1080p display (which is probably going to be the way that most consumers are going to enjoy 4K content for some time to come). As usual, we accept no quality tradeoffs in madVR and go with the high quality settings that we have used in previous reviews.

 

It is immediately obvious that the GT 640 is not in any way up to task for madVR processing on 4K content, even when it is just downscaling to 1080p. As evident from the above graph, the core is maxed out whenever we choose madVR as the renderer irrespective of the decoder used. Our suggestion is to retain EVR-CP as the renderer for all 4K content.

HTPC Aspects : Custom Refresh Rates

AMD and Intel GPUs don't offer the end users an easy way to create custom refresh rates for their displays. While Intel does offer a control panel for custom timings, it is heavily tied to the EDID, rendering it unusable for the most part. On the other hand, AMD GPUs have had a history of being close to the desired refresh rates out of the box. NVIDIA's GPUs have always needed some tweaking, and the Zotac GT 640 is no different.

As we have recounted in earlier HTPC reviews, a GPU should ideally be capable of the following refresh rates at the minimum:

•     23.976 Hz
•     24 Hz
•     25 Hz
•     29.97 Hz
•     30 Hz
•     50 Hz
•     59.94 Hz
•     60 Hz

Some users demand integral multiples of 23.976 / 24 Hz because they result in a smoother desktop experience, while also making sure that the source and display refresh rates are still matched without repeated or dropped frames. The gallery below shows the refresh rate handling for 24, 25 (x2 = 50 Hz), 29.97 (x2 = 59.94 Hz), 30 (x2 = 60 Hz), 50, 59.94 and 60 Hz settings.

The native 23 Hz setting, unfortunately, resulted in a 23.9724 Hz refresh rate.

However, with some custom timing setup, we were able to achieve 23.97622 Hz, which is off by just 0.000196 Hz. In my experience, this is the closest to the optimum refresh rate that I have ever achieved with a NVIDIA card.

The custom timing feature is usable, but not without its quirks. Adding a custom resolution is straightforward. Setting the vertical parameters to values similar to the ones in the screenshot above achieves desired results, but the 23 Hz resolution gets saved as 24 Hz. We already pointed out the details in our review of the GT 540M in the ASRock Vision 3D 252B. We hope NVIDIA fixes this annoying issue in one of the upcoming driver releases.

HTPC Aspects : HQV 2.0 Benchmarking and Video Post Processing in Action

HTPC enthusiasts are often concerned about the quality of pictures output by the system.  While this is a very subjective metric, we have decided to take as much of an objective approach as possible. Starting with the Core 100 review in 2010, we have been using the HQV 2.0 benchmark for this purpose.

The HQV 2.0 test suite consists of 39 different streams divided into 4 different classes. The playback device is assigned scores for each, depending on how well it plays the stream.  Each test was repeated multiple times to ensure that the correct score was assigned. The scoring details are available in the testing guide on the HQV website.

In the table below, we indicate the maximum score possible for each test, and how much the Zotac GT 640 was able to get. As mentioned in the previous section, we used NVIDIA Graphics Driver v301.42 for the benchmarking.

HQV 2.0 Benchmark - Zotac GT 640
Test Class	Chapter	Tests	Max. Score	Zotac GT 640
Video Conversion	Video Resolution	Dial	5	5
		Dial with Static Pattern	5	5
		Gray Bars	5	5
		Violin	5	5
	Film Resolution	Stadium 2:2	5	5
		Stadium 3:2	5	5
	Overlay On Film	Horizontal Text Scroll	5	5
		Vertical Text Scroll	5	3
	Cadence Response Time	Transition to 3:2 Lock	5	5
		Transition to 2:2 Lock	5	5
	Multi-Cadence	2:2:2:4 24 FPS DVCam Video	5	5
		2:3:3:2 24 FPS DVCam Video	5	5
		3:2:3:2:2 24 FPS Vari-Speed	5	5
		5:5 12 FPS Animation	5	5
		6:4 12 FPS Animation	5	5
		8:7 8 FPS Animation	5	5
	Color Upsampling Errors	Interlace Chroma Problem (ICP)	5	5
		Chroma Upsampling Error (CUE)	5	5
Noise and Artifact Reduction	Random Noise	SailBoat	5	5
		Flower	5	5
		Sunrise	5	5
		Harbour Night	5	5
	Compression Artifacts	Scrolling Text	5	5
		Roller Coaster	5	5
		Ferris Wheel	5	5
		Bridge Traffic	5	5
	Upscaled Compression Artifacts	Text Pattern	5	3
		Roller Coaster	5	3
		Ferris Wheel	5	3
		Bridge Traffic	5	3
Image Scaling and Enhancements	Scaling and Filtering	Luminance Frequency Bands	5	5
		Chrominance Frequency Bands	5	5
		Vanishing Text	5	5
	Resolution Enhancement	Brook, Mountain, Flower, Hair, Wood	15	15
Video Conversion	Contrast Enhancement	Theme Park	5	2
		Driftwood	5	2
		Beach at Dusk	5	2
		White and Black Cats	5	2
	Skin Tone Correction	Skin Tones	10	0
		Total Score	210	178

We find that score closely tracks what we had for the GT 540M in the ASRock Vision 3D 252B review. In fact, the only difference is the fact that the horizontal scroll response time has been improved a bit, enabling it to score two more points in that test. Given that the GT 540M had no trouble deinterlacing 1080i60 content, it was not a surprise to find that the GT 640 sailed through those tests. In the next section, we will look at some rendering benchmarks to see how deinterlacing operations load up the GPU. Chroma upsampling algorithms are passable, and there is no difference in quality between what was obtained through the 540M and what we got with the GT 640.

In our review of the video post processing features of the GT 540M, we had indicated that the contrast enhancement and skin tone correction features didn't work. We found no change in the v301.42 drivers. However, we did find contrast enhancement working with the black level testing clip in the AVS HD 709 calibration suite. This just proves that the dynamic contrast enhancement feature in the NVIDIA drivers doesn't work as effectively as Intel's or AMD's.

Should the low HQV score or lack of proper dynamic contrast enhancement prevent you from choosing the GT 640 for your HTPC? Definitely not! The nice aspect about NVIDIA GPUs is the fact that there are lots of HTPC software packages available to take advantage of the GPU resources. As long as the hardware deinterlacer works (it does, as the HQV scores for those tests indicate), and there are enough shaders and other computing resources available to let madVR work its magic, the HTPC end-user has no reason to worry. Advanced HTPC users tend to distrust any post processing done by the drivers, and would rather not let the driver mess with the video output by applying its custom post processing algorithms (which tend to break with every new driver release).

However, video post processing algorithms are not the only issue-prone HTPC aspects in the driver. Proper black levels are necessary irrespective of the color space being output. The gallery below shows that the behavior of the driver doesn't correlate in any way to the settings in the control panel. NVIDIA drivers seem to adopt two modes for the limited (16-235) and full (0-255) settings, one for global (desktop, photos etc.) and one for videos. Global mode is chosen to be limited (16-235) for all in-built resolutions and full (0-255) for all custom resolutions when in YCbCr mode with no way to change this (the gallery below shows correct dynamic range being chosen in RGB mode for still photos / desktop). The dynamic range for video and desktops are also different. Toggling the dynamic contrast enhancement box also seems to affect this setting. In addition, there is no way to specifically choose RGB Full or RGB Limited in the current drivers.

This dynamic range issue was apparently present in the Vista days, and fixed earlier. There appears to be a regression in the state of this bug recently, and we have been observing problems since May 2011 at least. A method to fix the issue has been outlined on Microsoft's official Windows community forums. It is disappointing to note that NVIDIA has still not fixed the issue despite the bug being a major annoyance for many HTPC users.

HTPC Aspects : Decoding and Rendering Benchmarks

We introduced our HTPC decoding and rendering benchmarks in the ASRock Vision 3D 252B review, and also used it in the 4K decode and display section of this review. In this section, we will look at how the system responds to the various test streams under various renderers and decoders. You can roll the mouse over the various entries in the first / last rows of the table below to compare the resource graphs.

AVCODEC + EVR-CP	AVCODEC + madVR	LAV CUVID + EVR-CP	LAV CUVID + madVR	DXVA + EVR-CP	DXVA Copy-Back + madVR
AVCODEC + EVR-CP	AVCODEC + madVR	LAV CUVID + EVR-CP	LAV CUVID + madVR	DXVA + EVR-CP	DXVA Copy-Back + madVR

Resource Usage Comparison - Software Decode vs. DXVA2 vs. LAV CUVID / EVR-CP vs. madVR

The GT 640 can be made to effectively work with madVR without issues. Unlike GT 540M, it is not necessary to carefully configure madVR to avoid dropping frames. With the queue sizes at the maximum, we were able to go through our rendering test suite in both full screen exclusive and full screen windowed modes without dropping frames.

Starting with this review, we also want to look at the power consumption profile of the system when subject to the rendering benchmarks.

Zotac GT 640 HTPC Testbed Power Consumption (W)
Idle	51.4 W
Benchmark Stream	CUVID		avcodec
	EVR-CP	madVR	EVR-CP	madVR
480i60 MPEG-2	67.5	70.9	58.9	69.3
576i50 H.264	67.5	67.1	58.8	57.7
720p60 H.264	69.9	74	67.9	77.8
1080i60 H.264	74.5	77.2	80.4	83.1
1080i60 MPEG-2	73.9	76.2	71.6	77.7
1080i60 VC-1	73.8	77.1	80.5	84.6
1080p60 H.264	72.7	76.3	74.3	85.7

madVR does carry a bit of a power penalty. As expected, software decode is more power efficient for lower resolution streams (up to 720p60) / MPEG-2 encodes. CUVID based hardware decode turns out to be more efficient with the 1080i and 1080p streams. Note that the benchmark streams were played off the local primary hard drive. The power consumption (measured at the wall outlet) also includes the hard drive activity.

As our coverage of the Zotac GT 640's HTPC aspects comes to a close, we would like to underline the fact that it is one of the best HTPC cards available in the market right now if madVR capability is a must.

For the general consumer, Intel's HD 4000 based system should be more than enough. However, in terms of looking into the future as well as current software infrastructure available, it is hard to go wrong with the GT 640. If it were not for the shortcomings of the NVIDIA drivers, we would have had no hesitation in crowning the GT 640 as the next undisputed HTPC king.

We are aware of the fact that AMD 7750 is a competitor to the GT 640 in more ways than one. We already covered AMD 7750's HTPC performance here. However, we will shortly be carrying out a review of the Sapphire Ultimate 7750 passively cooled edition using the same metrics considered in this review and the latest drivers from AMD.

Musing About Memory Bandwidth & The Test

As we discussed in our introduction, NVIDIA is launching the GeForce GT 640 exclusively as a DDR3 part. Because of the lack of memory bandwidth this is going to hold back the performance of the card, and while we don’t have a GDDR5 card at this time it will still be pretty easy to see what the performance impact is once we jump into our gaming performance section and compare it to other cards like the GTS 450.

In the meantime however we wanted to try to better visualize just how little memory bandwidth the DDR3 GT 640 had. It’s one thing to say that the card has 28GB/sec of memory bandwidth, but how does that compare to other cards? For the answer to that we drew up a couple of graphs based on the ratio of theoretical memory bandwidth to theoretical performance.

The first graph is the ratio of memory bandwidth to color pixel ROPs (bytes per color operation), which is an entirely synthetic metric but a great illustration of the importance of memory bandwidth. Render operations are the biggest consumer of memory bandwidth so this is where DDR3 cards typically choke.

Because of the nature of this graph cards are all over the place, with cards with particularly unusual configurations (such as the 4 ROP GT 440) appearing near the top. Still, high-end cards such as the GTX 680 and Radeon HD 7970 have among the highest ratio of memory bandwidth to color render operations, while lower-end cards generally have a lower ratio. At 1.97 B/cOP, the DDR3 GT 640 has the lowest ratio by far, and is only 66% of the ratio found on the next-lowest card, the GeForce GT 440 OEM. The fact of the matter is that because of its high clockspeed and 16 ROPs, the DDR3 GT 640 has by far the lowest ratio than any other card before it.

Our second graph is the ratio of memory bandwidth to shader operations (bytes per FLOP). Shaders aren’t nearly as bandwidth constrained as ROPs thanks to liberal use of register files and caches, but I wanted to take a look at more than just one ratio. Compared to our ROP chart the GT 640 isn’t nearly as much of an outlier, but it still has the lowest ratio of memory bandwidth per FLOP out of all of the cards in our charts.

Now to be clear not all of this is NVIDIA’s fault. Memory speeds have not kept pace with Moore’s Law, so GPU performance has been growing faster than memory speeds for quite some time leading to a general downwards trend. But the GT 640 is a special card in this respect in that there has never been a card this starved for memory bandwidth. This is further impacted by the fact that while GDDR5 speeds have at least been increasing as a modest rate over the years as GPU memory controllers and memory chip production have improved, DDR3 memory speeds have been locked in the 1.6GHz-1.8GHz range for years. Simply put, the gap between GDDR5 and DDR3 has never been greater. Even a conservative memory clock of 4.5GHz would give a GDDR5 card 2.5 times the memory bandwidth of a typical DDR3 card.

Of course there’s still a place in the world for DDR3 cards, particularly in very low power situations, but that place is shrinking in size every day. If and when it arrives, we expect that the GDDR5 GT 640 will quickly trounce the DDR3 version in virtually all gaming scenarios. DDR3 for a card this power hungry (relatively speaking) and with this many ROPs just doesn’t look like it makes a lot of sense.

The Test

For our test we’re using the latest NVIDIA drivers at the time our benchmarks were taken (301.42), and for AMD’s cards we’re using the new Catalyst 12.6 betas. For analysis purposes we’ve thrown in a couple of additional cards that we don’t normally test, such as the GDDR5 GeForce GT 240. The GDDR5 GT 240 has 43.5GB/sec of memory bandwidth but with a far older and less powerful GPU, which makes for an interesting comparison on progress.

CPU:	Intel Core i7-3960X @ 4.3GHz
Motherboard:	EVGA X79 SLI
Chipset Drivers:	Intel 9.​2.​3.​1022
Power Supply:	Antec True Power Quattro 1200
Hard Disk:	Samsung 470 (256GB)
Memory:	G.Skill Ripjaws DDR3-1867 4 x 4GB (8-10-9-26)
Case:	Thermaltake Spedo Advance
Video Cards:	AMD Radeon HD 7770 AMD Radeon HD 7750-800 AMD Radeon HD 6670 AMD Radeon HD 5750 NVIDIA GeForce GT 640 DDR3 NVIDIA GeForce GTX 550 Ti NVIDIA GeForce GTS 450 NVIDIA GeForce GT 440 NVIDIA GeForce GT 240
Video Drivers:	NVIDIA ForceWare 301.42 AMD Catalyst 12.6 Beta
OS:	Windows 7 Ultimate 64-bit

 

Crysis, Metro, DiRT 3, Shogun 2, & Batman: Arkham City

Our first graphics test is Crysis: Warhead, which in spite of its relatively high system requirements is the oldest game in our test suite. Crysis was the first game to really make use of DX10, and set a very high bar for modern games that still hasn't been completely cleared. And while its age means it's not heavily played these days, it's a great reference for how far GPU performance has come since 2008

NVIDIA’s internal guidance on GT 640 DDR3 performance is that it should beat the Radeon HD 6670 by around 20%, however Kepler’s poor performance under Crysis means that isn’t going to happen here. At 31fps at 1680 with Mainstream quality the GT 640 is just barely playable here, with the Radeon HD 7750-800 (a similar sub-75W card) more than doubling its performance. Worse, the GT 640 actually loses to the GT 240 here, with NVIDIA’s two-generation old card beating it by 14%. To NVIDIA’s credit this happened to be the only test where that occurs, but this does a great job driving home the point that the GT 640 is heavily handicapped with DDR3.

On that note, given what we’ve seen with Kepler so far with GK104 cards and now with the GK107 GT 640, this further reinforces the idea that Crysis above all else a memory bandwidth hungry test. GTX 680 failed to greatly improve upon the GTX 580 when the two had similar amounts of memory bandwidth, and while the GT 640 does improve upon the GT 440 by quite a bit at times the fact that it loses to the GDDR5 GT 240 lends further proof to our theories. It will be interesting to see what happens here once we do see a GDDR5 card.

Finally, while we’ve focused thus far on the GT 640’s poor performance relative to its current competition, it’s not all bad news for NVIDIA. With Performance and Gamer quality in particular the GT 640 improves upon the GT 440 by a rather impressive 48% despite the fact that the two cards have similar memory bandwidth, reflecting just how much of an impact doubling the shader performance and quadrupling the ROPs can have.

On that note, as this happened to be one of only a couple of games where our test settings overlapped our iGPU test settings, we’ve also thrown in our Intel HD graphics numbers. The CPUs aren’t identical (all of our dGPU testing is on SNB-E), but we’re GPU limited to such a large degree that it doesn’t make a practical difference. NVIDIA wants to sell the GT 640 as an upgrade to i3/i5 systems with Intel’s HD graphics, and while its performance may be lacking compared to its competition, at the very least GT 640 handily surpasses any iGPU. Intel’s decision to ship most desktop IVB CPUs with HD 2500 means that GT 640 can nearly quadruple the IVB GPU’s performance under Crysis.

Looking at the minimum framerates the story is much the same. The GT 640 is well behind the 7750 and similar cards. At best it manages to beat the GT 240, most likely due to the former’s lack of total VRAM (it only has 512MB).

Metro

Paired with Crysis as our second behemoth FPS is Metro: 2033. Metro gives up Crysis’ lush tropics and frozen wastelands for an underground experience, but even underground it can be quite brutal on GPUs, which is why it’s also our new benchmark of choice for looking at power/temperature/noise during a game. If its sequel due this year is anywhere near as GPU intensive then a single GPU may not be enough to run the game with every quality feature turned up.

Relative to its competition the GT 640 improves slightly over what we saw in Crysis, but it’s still trailing the Radeon HD 6670, never mind the 7750. Performance has improved by over 50% over the GT 440, which is enough to push us past 30fps at 1680 with medium quality settings, but that’s as much as NVIDIA’s going to get out of the GT 640 here.

DiRT 3

DiRT 3 is our next DX11 game. Developer Codemasters Southam added DX11 functionality to their EGO 2.0 engine back in 2009 with DiRT 2, and while it doesn't make extensive use of DX11 it does use it to good effect in order to apply tessellation to certain environmental models along with utilizing a better ambient occlusion lighting model. As a result DX11 functionality is very cheap from a performance standpoint.

DiRT 3 is traditionally a game that favors NVIDIA here, and while the GT 640 finally surpasses the 6670, it’s by no means a great showing for the GT 640. Again it’s handily beaten by the 7750 and GTS 450, and for as light as DiRT 3 is, we still can’t even break 40fps at 1680 Ultra quality without AA. To achieve 60fps here it’s necessary to turn it down to Medium quality. Elsewhere performance relative to the GT 440 has increased by nearly 60%, which is a big jump for NVIDIA but not enough to surpass their competition.

Total War: Shogun 2

Total War: Shogun 2 is the latest installment of the long-running Total War series of turn based strategy games, and alongside Civilization V is notable for just how many units it can put on a screen at once.

Under Shogun the story is much the same. The GT 640’s performance relative to the anemic GT 440 has greatly improved, jumping up by upwards of 50%, but it trails everything faster than a Radeon HD 6670. At the very least Shogun 2 is a relatively non-intensive game at 1680, so even at high quality the GT 640 is still achieving better than 40fps.

Batman: Arkham City

Batman: Arkham City is loosely based on Unreal Engine 3, while the DirectX 11 functionality was apparently developed in-house. With the addition of these features Batman is far more a GPU demanding game than its predecessor was, particularly with tessellation cranked up to high.

Batman: Arkham City is another game that traditionally favors NVIDIA’s GPUs, but again this isn’t much of a help here. At 1680 with Very High quality the GT 640 can just crack 30fps, and if we drop down to High quality that becomes a far more playable 51fps. But at the same time this leads to it greatly trailing the usual suspects at all configurations, and even the 6670 pulls ahead at High quality. Even the performance gains relative to the GT 440 have tapered off a bit, with the GT 640 only picking up 34% at High quality.

Portal 2, Battlefield 3, Starcraft II, Skyrim, & Civ V

Portal 2 continues to be the latest and greatest Source engine game to come out of Valve's offices. While Source continues to be a DX9 engine, and hence is designed to allow games to be playable on a wide range of hardware, Valve has continued to upgrade it over the years to improve its quality, and combined with their choice of style you’d have a hard time telling it’s over 7 years old at this point. From a rendering standpoint Portal 2 isn't particularly geometry heavy, but it does make plenty of use of shaders.

Portal 2 ends up being one of the better games for the GT 640, both on an overall basis and on a competitive basis. Overall we’re able to run Portal 2 at 1680 with maximum quality and with MSAA, and still hit 60fps, marking the first time we’ve been able to use such high settings on this card. At the same time the GT 640 is about as close to the 7750 as it’s ever going to get; it still trails by over 15% but this is a far better showing than earlier with games like DiRT 3. Performance relative to the GT 440 also once again looks good, with the GT 640 improving on the GT 440 by 60%.

Battlefield 3

Its popularity aside, Battlefield 3 may be the most interesting game in our benchmark suite for a single reason: it was the first AAA DX10+ game. Consequently it makes no attempt to shy away from pushing the graphics envelope, and pushing GPUs to their limits at the same time. Even at low settings Battlefield 3 is a handful, making it difficult to run on low-end GPUs.

We ultimately had to iterate through several different settings to find something the GT 640 could handle with BF3. 1680 is right out – even at minimum quality it could only hit 40fps in our test, which in a severe firefight means that framerates will bottom out at around 20fps. Only by finally backing off on the resolution to 1366x768 were we able to hit 60fps. So it’s definitely playable, but 1680 just isn’t going to happen.

Of course it goes without saying that the GT 640 once again gets left in the dust here by its competition. At High quality (our common benchmark for all single-GPU cards) it can’t even surpass the GT 440 due to the lack of memory bandwidth, and at lower settings the 7750 is still 20% ahead. At best you can point to our iGPU results and see that the GT 640 triples the HD 2500’s performance, reflecting the fact that even a slow dGPU is still faster than a slow iGPU.

Starcraft II

Our next game is Starcraft II, Blizzard’s 2010 RTS megahit. Starcraft II is a DX9 game that is designed to run on a wide range of hardware, and given the growth in GPU performance over the years it's often CPU limited before it's GPU limited on higher-end cards.

Starcraft II is another game that NVIDIA tends to do well in, and while it’s still not enough to make the GT 640 competitive with the likes of the 7750 or GTS 450, it’s enough to reduce the gap. The 7750 leads by only 15% here, the smallest lead of the day. Meanwhile the GT 640 improves on the GT 440 by 53% here, once again showcasing the impact of such a large increase in ROPs, shaders, and texture units.

Skyrim

Bethesda's epic sword and magic game The Elder Scrolls V: Skyrim is our RPG of choice for benchmarking. It's altogether a good CPU benchmark thanks to its complex scripting and AI, but it also can end up pushing a large number of fairly complex models and effects at once. This is a DX9 game so it isn't utilizing any advanced DX11 functionality, but it can still be a demanding game.

If we iterated through enough settings I’m sure we’d find something that the GT 640 could hit 60fps on, but thankfully it’s generally playable above 30fps, so 1680 at High quality should be fine for most users. Here it narrowly beats the 6670, but the 7750 and GTS 450 are well ahead, once again reflecting the relatively poor gaming performance of the GT 640.

Civilization V

Our final game, Civilization 5, gives us an interesting look at things that other RTSes cannot match, with a much weaker focus on shading in the game world, and a much greater focus on creating the geometry needed to bring such a world to life. In doing so it uses a slew of DirectX 11 technologies, including tessellation for said geometry, driver command lists for reducing CPU overhead, and compute shaders for on-the-fly texture decompression. And the release of a new expansion pack today should keep Civ V relevant to gamers for some time to come.

With Civilization V the GT 640 starts roughly the same as it started: poorly. The 7750 leads by over 50% at our highest settings, a gap wide enough to make the 7750 more than playable at these settings while the GT 640 struggles. The 56% improvement from the GT 440 means that NVIDIA is still making solid gains compared to the GT 440, but it’s just not enough.

Compute and Synthetics

Moving on from our look at gaming performance, we have our customary look at compute performance. Kepler’s compute performance has been hit and miss as we’ve seen on GK104 cards, so it will be interesting to see how GK107 fares.

Our first compute benchmark comes from Civilization V, which uses DirectCompute to decompress textures on the fly. Civ V includes a sub-benchmark that exclusively tests the speed of their texture decompression algorithm by repeatedly decompressing the textures required for one of the game’s leader scenes. Note that this is a DX11 DirectCompute benchmark.

Because this is a compute benchmark the massive increase in ROPs coming from GT 440 to GT 640 doesn’t help the GT 640, which means the GT 640 is relying on the smaller increase in shader performance. The end result is that the GT 640 neither greatly improves on the GT 440 nor is it competitive with the 7750. Compared to the GT 440 compute shader performance only improved by 28%, and the 7750 is some 50% faster here. I suspect memory bandwidth is still a factor here, so we’ll have to see what GDDR5 cards are like.

Our next benchmark is SmallLuxGPU, the GPU ray tracing branch of the open source LuxRender renderer. We’re now using a development build from the version 2.0 branch, and we’ve moved on to a more complex scene that hopefully will provide a greater challenge to our GPUs.

NVIDIA’s poor OpenCL performance under Kepler doesn’t do them any favors here. Even the GT 240 – a DX10.1 card that doesn’t have the compute enhancements of Fermi – manages to beat the GT 640 here. And the GT 440 is only a few percent behind the GT 640.

For our next benchmark we’re looking at AESEncryptDecrypt, an OpenCL AES encryption routine that AES encrypts/decrypts an 8K x 8K pixel square image file. The results of this benchmark are the average time to encrypt the image over a number of iterations of the AES cypher.

The GT 640 is at the very bottom of the chart. NVIDIA’s downplaying of OpenCL performance is a deliberate decision, but it’s also a decision with consequences.

Our fourth benchmark is once again looking at compute shader performance, this time through the Fluid simulation sample in the DirectX SDK. This program simulates the motion and interactions of a 16k particle fluid using a compute shader, with a choice of several different algorithms. In this case we’re using an (O)n^2 nearest neighbor method that is optimized by using shared memory to cache data.

All indications are that our fluid simulation benchmark is light on memory bandwidth usage and heavy on cache usage, which makes this a particularly exciting benchmark. Our results back this theory, as for the first and only time the GT 640 shoots past the GTS 450 and coms close to tying the GTX 550Ti. The 7750 still handily wins here, but based on the specs of GK107 I believe this is the benchmark most representative of what GK107 is capable of when it’s not facing such a massive memory bandwidth bottleneck. It will be interesting to see what GDDR5 GK107 cards do here, if only to further validate our assumptions about this benchmark’s memory bandwidth needs.

Our final benchmark is a look at CUDA performance, based on a special benchmarkable version of the CUDA Folding@Home client that NVIDIA  and the Folding@Home group have sent over. Folding@Home and similar initiatives are still one of the most popular consumer compute workloads, so it’s something NVIDIA wants their GPUs to do well at.

Folding@Home has historically pushed both shader performance and memory bandwidth, so it’s not particularly surprising that the GT 640 splits the difference. It’s faster than the GT 440 by 32%, but the GTS 450 still has a 25% lead in spite of the fact that the GT 640 has the greater theoretical compute performance. This is another test that will be interesting to revisit once GDDR5 cards hit the market.

Synthetics

Jumping over to synthetic benchmarks quickly, it doesn’t look like we’ll be able to tease much more out of GK107 at this time. GT 640 looks relatively good under 3DMark in both Pixel Fill and Texel fill, but as we’ve seen real-world performance doesn’t match that. Given that the GT 640 does this well with DDR3 however, it’s another sign that a GDDR5 card may be able to significantly improve on the DDR3 GT 640.

Tessellation performance is also really poor here, however there’s no evidence that this is a memory bandwidth issue. The culprit appears to be the scalability of NVIDIA’s tessellation design – it scales down just as well as it scales up, leaving cards with low numbers of SMXes with relatively low tessellation performance. NVIDIA’s improvements to their Polymorph Engines do shine through here as evidences by the GT 640’s performance improvement relative to the GT 440, but it’s not a complete substitute to just having more Polymorph Engines.

Power, Temperature, & Noise

As always, we wrap up our look at a new video card with a look at the physical performance attributes: power consumption, temperatures, and noise. NVIDIA is breaking new ground for desktop Kepler with their first sub-75W card, so it will be interesting to see just what the tradeoff is for such low power consumption.

Zotac GeForce GT 640 DDR3 Voltages
GT 640 Idle	GT 640 Load
0.95v	1.00v

NVIDIA doesn’t do a lot of voltage scaling with the GT 640. At idle it runs at 0.95v, and makes a short jump to 1.00v under full load. For a 28nm GPU 0.95v under idle is a bit higher than what we’ve seen in the past, which may explain the official 15W idle TDP.

Earlier we theorized that the GT 640 would have worse idle power characteristics than the GT 440, and this appears to be the case. The difference at the wall is all of 3W but it’s a solid indication that NVIDIA has at best not improved on their idle power consumption, if not made it a bit worse. The good news for them is that in spite of this slight rise in idle power consumption it’s still enough to tie the 7750 and beat older cards like the 6670 and GTS 450.

Long idle on the other hand sees the 7750 jump back into the lead, as NVIDIA doesn’t have anything resembling AMD’s ZeroCore power technology.

Of all of the sub-75W cards in our benchmark suite, the GT 640 ends up having the lowest load power consumption. Under both Metro and OCCT it’s equal to or better than the GT 440, GT 240, 6670, and 7750. AMD’s official PowerTune limit on the latter is 75W versus the GT 640’s 65W TDP, so this is not unexpected, but it’s always nice to have it confirmed in numbers.  That said, for desktop usage I’m not sure 4-8W at the wall is all that big of a deal. So while NVIDIA’s power consumption is marginally lower than the 7750 they aren’t necessarily gaining anything tangible from it, particularly when you consider the loss in performance.

GPU temperatures look absolutely stunning here, and in fact it’s better than we would have expected. Zotac’s heatsink is not particularly large, and while these type of cards typically stay under 70C we would not have expected a single-slot heatsink to perform this well. If these kinds of temperatures can carry over into other designs there’s a very good chance we’re going to see some nice passively cooled cards in the near future. In the meantime buyers sheepish about high temperatures are going to find that Zotac’s GT 640 is an exceptional card.

Last but not least we have noise. Zotac’s card may be exceptionally cool, but that single slot cooler and small fan comes back to bite them when it comes to noise. That little fan simply doesn’t idle well, leading to the Zotac GT 640 being one of the loudest idling sub-75W card we’ve seen in quite some time. Zotac fares much better under load – where the 7750’s equally tiny and tinny fan leads to the 7750 ending up as the louder card by about 1dB – but really neither of these cards is particularly quiet for as little power as they consume. HTPC users who aren’t already looking at passively cooled cards are probably going to want to look elsewhere unless they absolutely need a single-slot card, as there’s a very clear tradeoff on size versus noise here.

Final Words

Bringing things to a conclusion, when Ganesh and I began working on reviewing the Zotac GeForce GT 640 DDR3 we set out with two cards and two different goals. I would take a look at gaming performance and physical characteristics while Ganesh would be free to focus on the HTPC side of things. Unsurprisingly we have come to two very different conclusions.

From a gaming standpoint the GeForce GT 640 is unremarkable if not flat-out bad. NVIDIA’s GK107 GPU may have a lot of performance potential, but its first desktop iteration as the GeForce GT 640 DDR3 does not. The decision to equip it with DDR3 clearly bottlenecks the card just as it has done to previous generation entry-level cards. So this is by no means a new problem, but it’s a recurring problem that always has the same solution: buy GDDR5. With that said, NVIDIA and their partners will no doubt sell GT 640 DDR3 cards by the truckload – make no mistake, having lots of VRAM moves lots of cards – but if you’re fortunate enough to be reading this article then you’re probably well aware of the performance difference between DDR3 and GDDR5.

As a result of the decision to equip the GT 640 with DDR3, compared to every other major card around $109 including the 7750, 5750, GTS 450, and GTX 550 Ti, the GT 640 is the slower card and typically by quite a bit.  Even if we look at the much narrower range of sub-75W cards, AMD’s 7750 has the clear upper hand in gaming performance. As it stands there’s just no good reason to get a GeForce GT 640 if you intend to game on it. A Radeon HD 7750 performs far better for the same price and effectively the same power consumption, and that’s really all there is to it.

This brings us to the second conclusion: HTPC usage. To some extent HTPC tasks are still reliant on memory bandwidth – post-processing in particular – but overall the measure of a good HTPC card is rooted in its features and power consumption rather than its raw performance. To that extent the GeForce GT 640 is nearly everything we expected from the moment we saw GK107. 4K video decoding via VP5 and 4K over HDMI are working, custom resolution/timings are working, there’s enough processing power to handle everything short of intensive madVR use-cases, and all of this is on a sub-75W card. While NVIDIA still has some room to grow and bugs to fix, at this point it’s certainly the best HTPC card we’ve tested yet.

Really the only thing we don’t have a good handle on for HTPC usage right now is video encoding through NVENC. We’ve already seen NVENC in action with beta-quality programs on the GTX 680, but at this point we’re waiting on retail programs to ship with support for both NVENC and VCE so that we can better evaluate how well these programs integrate into the HTPC experience along with evaluating the two encoders side-by-side. But for that it looks like we won’t have our answer until next month.

Wrapping things up I’d like to spend a few words on the Zotac GeForce GT 640 design in particular. Zotac has worked themselves into an interesting position as the only partner currently offering a single-slot card, and while the fan noise will probably drive some customers away it certainly fulfills its role well. Unfortunately for Zotac the mini-HDMI port problem we outlined earlier is a direct obstacle for the GT 640’s greatest strength: HTPC usage. It’s by no means an insurmountable problem, but it makes the Zotac card a poor out of the box HTPC product. Once Zotac resolves the issue they’re going to have among the finer HTPC cards available, but until then prospective HTPC customers will need to take extra care to mitigate the issue.