It's always nice to get exclusive looks at things here at Hexus, even if it is just a UK exclusive. It's also nice to get a look at a product that at some point in its conception, might not have seen the light of day.
NVIDIA were in two minds about releasing the baby of the GeForce4 Ti family, the Ti4200. There is certainly the opportunity for some price overlap between NVIDIA products at this end of the scale. This made the decision to release the card a difficult one for NVIDIA. Do they release a budget Ti card to let more people get their hands on the tech and sell a few more NV25's or do they leave it at Ti4400 and Ti4600 to differentiate the Ti from the MX?
I think they made the right decision with releasing this iteration of the NV25 GPU since I'm not really a fan of the GeForce4 MX (NV17) and I'd rather a full hardware DX8 accelerator proliferated rather than an inferior, poorly marketed, product with not much price difference.
So what exactly differentiates this little card that NVIDIA sent us from the other GeForce4 Ti's we've seen so far? Well for starters it's just a baby PCB compared to the monster sized ones that hold a Ti4400 or Ti4600. According to this little article here @ The Tech Report the PCB is different due to the different memory types and electrical routing requirements of these slower clocked cards. Looking at a Ti4400 or Ti4600 and you'll see the extra electrical hardware, especially at the extremes of the board at the far edges.
No such hindrance here on the test 64MB Ti4200. That brings us on nicely to one of the other differences, memory size. The Ti4200 is the first NV25 based design to feature a 64MB memory size. While it's possible for Ti4400's and Ti4600's to address the smaller memory size since they are after all the same GPU, the low end Ti4200 is the first product to offer it.
You'll see 128MB Ti4200's hit the market but they'll feature slower clocked memory than the 4ns Hynix memory clocked to 500MHz (250MHz DDR) on the test board. We've spoken before about the effect a 128MB frame buffer has on performance and right now it's nothing significant, but in the future it might be a defining reason to avoid some graphics cards and buy others.
The rest of the card is just as you'll find with a regular GeForce4 Ti. We have a DVI port which is a good feature but I wonder how much use it'll see on a card in this price range. Shipped with a DVI to VGA convertor and it makes more sense but that's up to the manufacturer. TV-Out using the Conexant CX25871 makes an appearance too just like all the other GeForce4's we've seen so far (bar the Philips VIVO controller on the MSI Ti4400) and the quality is excellent. Again, nView using a TV as the 2nd head works like a charm and lets me keep using my Sony for windows duties and watch video on the TV.
So onto the formal spec of the card. This is taken from NVIDIA's website for the most part.
• 113 Million vertices per second
I covered nView, LMA II and the updated shader hardware in our Ti4600 review here so if you need a refresher course on the tech behind the NV25 and these new cards, take a look at that article.
• 4 Billion anti aliased samples per second
• 1.03 Trillion operations per second
• 8GB per second internal memory bandwidth
• 64MB DDR memory
• DVI, D-Sub, TV-Out outputs using nView logic
• nFiniteFX II
• Lightspeed Memory Architecture II
• nView Display Technology
• Accuview Antialiasing
Installation and Bundle
There wasn't any bundle at all which is normal with reference boards. Just a UPS-delivered white box with the card wrapped up in an antistatic bag and bubble wrap. I had to download some drivers for the card since my 27.42 Windows XP Detonator set didn't support the card. I grabbed the latest unofficial Det's for the card (28.80's) and used them for the testing. All the other card results you'll see were grabbed using the 27.42 set (including the GeForce3 Ti500 results).
Installation was incredibly simple. Remove the previous drivers using the Windows XP control panel, shut down the test rig, remove the old card, install the Ti4200 and install the 28.80 driver set. It couldn't be simpler.
Performance will be measured using our same graphics card test suite as previous card reviews. This test suite is undergoing some changes in the near future so future reviews will show a couple more tests than we do now to cover more bases.
Windows XP was the operating system and the installation was unchanged from the MSI GF4 Ti4400 review so all the test software remained installed. A reboot was done between tests to maintain consistency with other reviews although this may change to reflect real world usage. As I say, we're refining the test setup at Hexus as we speak!
The 28.80 drivers were used on the Ti4200 and the 27.42's on all other cards.
Before we hit the numbers, here's a quick rundown of the test system.
• EPoX 8K3A, KT333 Chipset, Socket A AMD DDR Motherboard
• Unlocked AMD Athlon XP1500+ Processor (1.33GHz, 10 x 133)
• 2 x 256Mb Samsung PC2700 DDR Memory Modules (CAS2)
• Visiontek Xtasy GeForce4 Ti4400 128MB
• Gainward Ti550 GeForce3 Ti500 64MB
• NVIDIA Reference GeForce4 Ti4200 64MB
• Adaptec 39160 PCI SCSI Dual Channel U160 controller
• 2 x 73Gb Seagate Cheetah U160 10,000rpm SCSI disks
• Plextor 12/10/32S SCSI CDRW
• Pioneer 6x Slot-load SCSI DVD
• Creative Soundblaster Audigy Player Retail
• Windows XP Professional Build 2600.xpclient.010817-1148
Nothing different except the use of the 28.80's for the Ti4200. Overclocking was done using RivaTuner RC10.1. I overclocked in 5Mhz steps on the core clock until I encountered instability and dropped down until the card was stable running all the tests and was able to loop 3DMark 2001 SE. Then I did the same with the memory clock to find its stable limit, dropping down 1MHz at a time when instability or memory artifacts were encountered.
• DetonatorXP 27.42 and 28.80 NVIDIA drivers
• Aquamark v2.3
• Quake3 v1.30
• 3DMark 2001 Professional Second Edition
• Serious Sam: The Second Encounter Demo
Benchmarks were ran 3 times and the upper and lower results were discarded. Any runs that were outside a 1% variance were discarded and a fresh set of 3 runs were done until 3 results within 1% of each other were recorded and the selection process was done again.
We'll be looking at performance from 4 points, 2 featuring the test card.
• GeForce4 Ti4400 at stock speeds (275/550) with 27.42 drivers
I left out GF4 MX numbers to keep graph clutter down and because the Ti range is aimed at replacing the GeForce3, not replacing a GF4 MX.
• GeForce4 Ti4200 at stock speeds (250/500) with 28.80 drivers
• GeForce4 Ti4200 overclocked to 277/571 with 28.80 drivers
• GeForce3 Ti500 at stock speeds (240/500) with 27.42 drivers
Lets take a look at 3DMark 2001 SE first. As you know, 3DMark 2001 SE is a full system benchmark with some tests relying on card performance, some on system memory bandwidth and some on CPU clock. But given the same base system it becomes a valid graphics card benchmark.
The Ti4200 actually wins out here when overclock and managed to overclock a little bit faster than a stock Ti4400. Obviously overclocking the Ti4400 would make the win for the Ti4200 disappear but it shows the card is a decent overclocker and that it responds well to the overclock.
Getting close to 9000 out of the box, untweaked, on such a slow processor (not that 1.33GHz Athlon XP's are slow!) is something to shout about. The new nFiniteFX logic on the NV25 really helps the card pull out a lead over a similarly clocked GeForce3. The extra vertex shader really helps in the Dragothic High Detail test where there is a ton of vertex data being moved about and the driver does well to load both vertex units. The other advantage over GF3 comes from the LMA II crossbar which improves effective memory bandwidth over an identical memory clock on the GeForce3 Ti500.
What about Aquamark? Aquamark is a shader hardware heavy test that really needs a competent DX8 accelerator to do well. Onboard solutions such as the nForce and GPU's like NV17 (GF4MX) and NV15 (GF2) really don't do well in Aquamark on the test CPU since the CPU has to do more offloaded work than with a full DX8 implementation.
Here are the results.
It's a pretty close run thing across all 4 test points, mainly do to the low clocked processor. Given a higher speed base, the Ti4200 would have put some distance between it and GeForce3 Ti500. Of course it would beat a regular GeForce3 and GeForce3 Ti200 by increasing margins.
Again we see the small win over stock Ti4400 when overclocked. The stock clocked result slots in right where expected between Ti550 and Ti4400. We hinder the card here with the CPU but it performs well and beats NV20 which is important. From the graph, it's not sure how hard the extra vertex unit is being made to work by the driver since the results are so close together. We do know however that Aquamark is shader heavy so we can maybe hypothesise in Aquamark that the 2nd unit isn't being used all that much?
Onto Quake3, a benchmark that seems to have performed well on cards since the dawn of time! It's been around for a while and is OpenGL based and doesn't need specific shader hardware to work well. NV15 does well in Quake3 so it'll be interesting to see what NV20 and NV25 do. Here's the graph.
This graph is quite interesting and shows you the effect that new drivers can have on proceedings. In the highest resolution, 1600x1200, things are in order. Overclocked Ti4200, Ti4400, Ti4200 and GF3 Ti500 all order themselves according to card clock speed. The cards all are made to work harder at the high resolution and fillrate becomes important. But at 1280x1024 and 1024x768 things are a bit different.
The stock Ti4200 proceeds to beat the stock Ti4400 at those resolutions buy a few, but perfectly reproducable, frames per second. Since nothing else has changed, we can only put it down to the driver used. 28.80 seems to have a nice WindowsXP OpenGL driver! Overclock the Ti4200 and the gap widens. An interesting result and one that highlights the importance of keeping an eye on driver releases and giving the odd driver a try on your hardware.
All cards perform excellently and it's nothing to worry about at all. A nice, if a little biased, victory for our little challenger!
Finally, Serious Sam: The Second Encounter. Like, 3DMark 2001 SE, it's a full system test for the most part when you are benchmarking it. The demo used in our test, Valley Of The Jaguar, works the graphics card hard so it's a nice test. It also treats a GeForce4 Ti exactly like a GeForce3 would at those clocks. Therefore a 275/550 clocked GeForce4 like the Ti4400 will perform broadly similar to a 275/550 clocked GeForce3. Here are the results.
At high resolution, you can see that effect since all the cards are bunched up tight. The fact that the benchmark seemingly performs identically on NV20 and NV25 at the same clocks doesn't help us show off the NV25 very well. As things drop to lower resolution we see a more pronounced difference in the scores but again, nothing exceptional.
The Ti4200 overclocked wins by a cats whisker but essentially it performs identically to stock Ti4400 and that's to be expected since the clocks are nearly identical. You can also see that card memory clock doesn't do much here and it's all about the core.
Anti-Aliasing and Anisotropic Filtering Performance
Here's a treat for you! It's not often that I do anti-aliasing and anistropic filtering analysis since usually the hardware I run doesn't quite do AA quick enough, I can't see the effect it has and in the case of anisotropic filtering, the performance hit is too great.
NVIDIA themselves make big things about Accuview, the logic on the NV25 that performs full scene anti aliasing on your rendered scenes to eliminate 'jaggies'. Jaggies are the jagged edges you see on aliased lines, the steps in the gradient of the line as it moves one pixel at a time in whatever direction.
In the games I play, AA makes no difference to me but that's not to say it wont for you hence a performance analysis is worthwhile. Anisotropic filtering (I'm not sure what algorithm NVIDIA use for their hardware aniso) filters the texture data applied to a surface and generally helps improve visual quality, especially on mip map levels.
Mip maps are intermediate texture levels generated from a base texture but lacking detail. They are used in place of the base texture on sufaces far (or not so far!) from the viewer so that the card has to do less work texturing surfaces you aren't really focussing on. Hardware or software can generate mip map levels (usually hardware, to save pulling the mip map data across the AGP bus from memory, sometimes disk memory which is slow).
Since mip maps are by their very definition ugly (less detail than the base texture) then any filtering effects that help improve their visual quality is welcome. Of coure, aniso filtering is applied not just to mip map levels but they are the textures that seem to benefit most.
Also, aniso filtering helps visual quality most when surfaces are rotated at non right angles to the viewer.
The test consists of a simple 3DMark default benchmark run with varying levels of anti-aliasing and anisotropic filtering enabled.
As we've said at Hexus a few times, a 6000+ score in 3DMark 2001 is indicative of good all round game performance on a system. So what we are looking for is some possible combination of aniso and AA that will give us that score and let us enable these visual enhancing options and still retain performance that's usable.
Here's the graph.
As you can see, the first result in the graph is our ideal situation. 1024x768x32 with 2x AA and 2-tap aniso and more than 6000 marks. This gives us a degree of anti-aliasing to remove some jaggies from our scene and also a 2-tap version of anisotropic filtering to improve the visual quality of textures in our visible scenes.
One of the things to notice is the essentialy free Quincunx anti aliasing you get when compared to 2x AA. There is obviously a speed hit but Quincunx is essentially a 'softer' version of 4x AA. You get all the benefits of 4x AA (essentially 4 samples per pixel) bit with the speed and memory requirements of 2x AA. Quincunx uses exactly the same amount of framebuffer memory, around 15-16MB in our 1024x768x32 test mode, as 2x AA. So you get soft (softer edges on your aliasing) 4x AA for the same speed hit as 2x AA.
4xS is like 4x but with a further 2x multi sample for the final pixel colour. The multi sample is taken from sub pixel colour data. Depending on application and rendered scene, 4xS can be slower or faster than 4x and in our case it's slower. The performance hit on the test setup is unacceptable and not recommended
Moving down the graph you can see that 4x and 4xS anti aliasing are close on 50% speed hits compared to the base level of 88xx marks. 2x and Quincunx are acceptable in their speed hits.
In terms of aniso filtering (generally regarded as a killer on NVIDIA hardware), we don't do too bad in 3DMark with this card although 4 and 8-tap wouldn't go to well with any of the aliasing modes. You'd need better base hardware for that.
So it seems that 2x AA and 2-tap anisotropic filtering is the best combination on NV25, even on the slowest version and this comparatively slow test processor. It's a testament to the cards fillrate and high speed rendering logic that we can enable both features and still break our 6000 point barrier on this hardware.
That should show you that anti-aliasing, especially at 1024x768 where it is appreciated more than at higher resolutions is doable on NV25 where it might not have been before and anisotropic filtering is a useful addition too. As always, make sure you can see the benefits in the games you play before enabling them.
There's not really much to say here. As you can see the card sits, in raw performance terms, inbetween NV20 in the form of our Ti500 GeForce3 and the Ti4400 GeForce4. That's to be expected and in terms of surprising us with stock performance, it hasn't really since we knew what to expect. The surprise in performance came with the overclocking. Ti4400 speeds were simple with this card. While it might not be possible to say that about the 128MB version that uses lower rated memory modules, this little 64MB version did the 550MHz memory clock with ease. Another 25MHz on the core proved easy too and the GeForce3 style cooler manages to keep NV25 nice and cool at Ti4200 clocks.
Whether card manufacturers choose the same cooler (and I doubt they will), remains to be seen but you can be sure their solutions will cool just as effectively.
Performance with AA and aniso filtering is also above the levels provided by NV20. Accuview really does a good job and 'free' Quincunx is a good thing provided you were enabling anti aliasing in the first place.
Excellent performance, especially given the target cost of Ti4200 cards. NV25 does the job yet again. The 64MB framebuffer is the only possible hindrace for this card. As game requirements get bigger, games with more than 64MB of texture and geometry data per scene will become more common. For the time being though, it poses no problem. Running at high resolution (1600x1200) may cause you some performance issues in the future but nothing shocking and nothing to put you off. Just means the card and driver will work a bit harder!
While it's only a reference card supplied from NVIDIA with no software bundle, no manual and no driver CD, we can still give it a useful workout against recent Ti4400's we've seen and also the card it replaces, the Ti500.
And replace it does, for an even lower price than Ti500 ever was. Performance is better than Ti500 in all areas and when overclocked, the gap widens further. You also get all the benefits of the bigger cards in the range and the benefits are numerous.
Since the GPU's are identical you get nView (a combo of technologies), Accuview anti-aliasing that can be used without a significant speed hit and the memory bandwidth technology enhancements. Performance is excellent and the price should be too.
A good addition to the NV25 family and my favourite NV25 card to-date due to the expected low price. I'm a fan of the bigger cards but at a projected good sub £200, you can't help but like this little one.
As always, it's nice to test new hardware so many thanks to NVIDIA for letting Hexus take an exclusive first look in the UK at Ti4200.