NVIDIA NV18/NV28 and AGP 8x Investigation

by FastSite
12/06/2002 | 12:00 AM

On September 25, NVIDIA introduced its new graphics chips supporting the AGP 3.0 standard and 8x data-transfer speed.

When naming the chips, the company didn't exert imagination so we have got "NVIDIA GeForce4 Ti4200 with AGP8x" (unofficially known as NV28) and "NVIDIA GeForce4 MX440 with AGP 8x" (NV18). Actually, NV18 and NV28 only differ from NV17 and NV25 by their AGP 8x support. Besides, they also feature higher working frequencies:

• NVIDIA GeForce4 Ti4200 with AGP 8x: 250MHz chip frequency, 128MB 128bit DDR SDRAM working at 512MHz (256MHz DDR).
• NVIDIA GeForce4 MX440 with AGP 8x: 275MHz chip frequency, 64MB 128bit DDR SDRAM working at 512MHz (256MHz DDR).

The increase in the working frequencies is quite significant: "older" GeForce4 Ti4200 based graphics cards with 128MB of graphics memory worked at 250MHz/444MHz (222MHz DDR), while GeForce4 MX440 based cards - at 270MHz/400MHz (200MHz DDR).

In this review we will try to find out the advantages brought about by the increased frequencies and the AGP 3.0 support.

AGP 3.0: - What's New?

The AGP 3.0 standard is the evolutionary development of the AGP 2.0. It's the newest, but regrettably (or, fortunately) the last generation of the AGP standard, which appeared in 1996 and has been steadily developing since then. Intel, the inventor of the AGP, put the full stop in AGP development with this last modification.

The company sees further development of high-performance data transfer buses in serial interfaces and is actively promoting the new PCI-Express standard. Judging by the mess and noise around that new standard, AGP 3.0 is going to give way to faster and perspective data buses soon, despite of all the advantages it may have.

But back to the point. What's so new and impressive about AGP 3.0?

Not delving deep into details, let's list its main differences from AGP 2.0, which are interesting from the bandwidth point of view:

• 4x and 8x data transfer rates support, so the peak bandwidth of the bus in this modes is 1064 and 2128MB/sec, respectively;
• Out of two variants of AGP-transactions (SBA and PIPE), there is only one left - SBA.

  The transaction is a process that includes the creation of a read or write request and the reception of the response to the request. For example, during AGP-texturing, the graphics chip wants to get data from the textures stored in the system RAM and starts bombarding the bus with data-read requests. In response, it gets the requested texture samples.
  In the PIPE mode, the same channel is used for both creating and receiving requests, so the next request cannot be sent until there arrives a response to the pervious one.
  In the SBA (SideBand Addressing) mode, a separate request channel is used, so the requests may be sent before receiving responses to the previous ones. It's clear that the SBA mode uses up the AGP bandwidth more effectively.

• "Long transactions" support (when over 64Bytes are transferred) has been removed. Now, if the graphics chip needs to read over 64Bytes, it has to carry out several transactions;
• "High priority transactions" support has been removed. All the AGP transactions now have "low priority";
• Isochronous mode and isochronous transactions support has been introduced. The regular AGP transactions are now called asynchronous for distinguishing purposes.

  The introduction of transactions of the new type was caused by the fact that there are situations when the bandwidth is not enough and the transaction takes too long, although the AGP bus features high average bandwidth and low latency. This can result in no real trouble in common cases, for instance when working with a 3D graphics card. The worst thing that can happen: some delays of the frames output.
  But the delays accompanied by low buffering may lead to data loss in devices that process streaming data, such as video-capturing cards.
  The AGP 3.0 specification includes a rule that obliges the chipset to effect a set number of specifically marked transactions in a set period of time (1ms), so that it can guarantee no delays and to prevent the bandwidth from falling below a certain level. Transactions like that are called isochronous and can be interlaced with regular, asynchronous transactions. The device that uses isochronous transactions for streaming data transfer can determine the read/write requests ratio in the dedicated channel.
  Isochronous transactions can be used only in 8x data transfer mode. According to AGP 3.0 specs, the AGP bus bandwidth for isochronous transactions cannot be lower than 128MB/sec.

So, there is no doubt that the AGP 3.0 8x mode provides the maximum data-transfer rate. The peak value is 2128MB/sec, which doubles the 1064MB/sec peak value of the 4x mode in AGP 2.0.

Considering the above mentioned things, it's no wonder that the today's leaders in the 3D gaming chips market, ATI and NVIDIA, and also less fortunate SiS and Matrox, accepted the standard at once and released (Matrox is just planning to release) new products with AGP 3.0 and 8x mode support.

Now let's dwell upon some peculiarities in relationships between graphics cards and AGP 3.0-supporting mainboards. You will understand later on, why it is so necessary :).

Universal AGP 3.0, Forced AGP 2.0

The connectors of the graphics cards and slots on the mainboards that support AGP 2.0 and AGP 3.0 may not differ physically, but there are certain "electrical" divergences, of course. Firstly, the connectors of AGP 3.0 supporting devices use contacts that haven't been used before. Secondly, some contacts changed their fucntions. Thirdly, while the high signal in AGP 2.0 equaled 1.5V, it's only 0.8V in AGP 3.0. Fourthly, ...

To cut it short, the graphics card that only supports AGP 2.0 and the mainboard that only supports AGP 3.0 wouldn't work together, in the best case. That's why there are Universal AGP 3.0 and Universal 1.5V AGP 3.0 specifications that add to compatibility of the AGP 2.0 hardware. The possible modes and speeds of graphics cards and mainboards that correspond to different AGP specs, are listed below:

So, I installed the graphics card into the mainboard slot. How does the mainboard recognize the AGP port mode? The system powers up and goes through POST. Then the graphics card and the mainboard that comply with Universal AGP 3.0 and Universal AGP 1.5V 3.0 specs determine the port mode by means of specific contacts on the AGP connector. It is "GC_DET#" contact (Pin #A3), which detects graphics card's ability to work in AGP 3.0 mode (if the card supports AGP 3.0 or Universal AGP 3.0, this contact is connected to the common wire, otherwise it's not connected to anything). The second one is "MB_DET#" contact (Pin #A11), which detects mainboard's ability to work in AGP 3.0 mode (if the mainboard supports AGP 3.0 or Universal AGP 3.0, this contact, like GC_DET#, should be connected to the common wire, otherwise it's not connected to anything).

So, if the both contacts are "grounded" in the graphics card and mainboard, they are checked and then AGP 3.0 is turned on during POST, otherwise - AGP 2.0. Later on, when the system is at work, you can't change the port mode, only the data-transfer rate can be set at either 4x or 8x for AGP 3.0 and 1x, 2x or 4x for AGP 2.0.

When we tested NV18 and NV28 based graphics cards with the mainboard supporting AGP 3.0, the AGP 3.0 mode is enabled without any options. This fact turned out a real nuisance during the tests.

By changing data-transfer speeds, we could see the real advantage of the transfer via AGP bus in test results. In other words, the advantage of 8x speed compared to 4x speed. But 4x in the AGP 3.0 mode and 4x in the AGP 2.0 mode are different things as AGP 3.0 forces all the new AGP 3.0 features, no matter what data-transfer rate has been set. And it means that it would be incorrect to compare 4x and 8x speeds in AGP 3.0, as the effective bandwidth in 4x AGP 3.0 may differ from 4x AGP 2.0, that is, from the mode supported by NVIDIA NV17/NV25 based graphics cards.

Nevertheless, in spite of this "incorrectness" we will compare the graphics cards performance in AGP 3.0 mode at 4x and 8x speeds.

Another way to estimate the advantages of AGP 3.0 and 8x data transfer speed is to compare NV17/18 and NV25/28 by setting equal core and graphics memory frequencies. But again, this method lacks accuracy, as we will compare different graphics cards, with different PCB designs, different BIOS, memory timings and different graphics chips…

Although we did carry out such a test, bear in mind the above mentioned stipulation.

The only correct way to figure out the actual advantages of AGP 3.0 8x seems to be like that. We should somehow make NV18/28 based cards work in a Universal AGP 3.0 mainboard in both modes: AGP 3.0 and 2.0.

Do you think that it is impossible? Not at all :).

AGP 3.0 specs say that if one party - the graphics card or the mainboard - doesn't support AGP 3.0, the AGP 2.0 mode will be enabled instead.

So it's simple: if we insulate the GC_DET# contact in the graphics card connector, the mainboard will think that the card doesn't support AGP 3.0. Similar: if we insulate the MB_DET# contact in the AGP slot, the graphics card will check it and find that the mainboard doesn't support AGP 3.0. As a result, if we insulate both contacts, then both: graphics card and mainboard, will enable AGP 2.0 mode (of course, if they both comply with the Universal AGP 3.0 or Universal AGP 1.5V 3.0 specs, i.e. are compatible with the "old" AGP 2.0 standard).

The insulation was made with tiny pieces of scotch tape.

The reference card based on NV28 (NVIDIA GeForce4 Ti4200 with AGP 8x) will serve as an example. That's the way its AGP connector looked like before taping the contacts (the MB_DET# and GC_DET# contacts are marked with arrows):

   

When taped up, the MB_DET# and GC_DET# contacts looked like that:

The card is ready for testing in the AGP 2.0 mode: the MB_DET# and GC_DET# contacts are insulated just like in AGP 2.0 graphics cards, where they are not laid out at all. You can see it in the following snapshot of the NVIDIA GeForce4 Ti4600 based graphics card:

   

So, after making clear what and how we are going to test, let's go over to the "heroes of the occasion": the new GPUs from NVIDIA with AGP 8x support and the graphics cards based on them.

Closer Look: NVIDIA NV18 and NV28 Graphics Cards

New NVIDIA chips support AGP 3.0 and 8x data transfer rate that is why their engineers had to re-design the layout a bit. The NV18 (GeForce4 MX440 with AGP 8x) based reference card differs from its predecessors most notably:

   

As you can see in the pictures, the graphics card features BGA memory chips, just like NVIDIA GeForce4 MX460. However, this time all the memory chips are placed on the front side of the PCB.

The GPU is cooled by a passive heatsink. Under the heatsink we find a chip marked as: "MX440-8x" meaning "NVIDIA GeForce4 MX440 with AGP 8x".

The GPU works at 275MHz frequency.

The graphics card is equipped with 64MB of 128bit DDR SDRAM graphics memory from Samsung with 3.6ns access time.

The graphics memory of NVIDIA GeForce4 MX440 reference card with AGP 8x works at 512MHz (256MHz DDR).

The card is equipped with a DVI and D-Sub connector, as well as with a composite TV-In/TV-Out connector. The output onto the TV-set is performed by the TV-Out unit, which is integrated into all GeForce4 MX chips. The external SAA-7114H chip from Philips is responsible for video capturing and decoding.

The NVIDIA GeForce4 Ti4200 with AGP 8x based graphics card doesn't differ much from the cards that follow NVIDIA GeForce4 Ti4200 reference-design:

   

The reference card is equipped with a rather out-dated cooler, similar to what they used in NVIDIA GeForce3 based cards. Under the unassuming cooler we saw a new NVIDIA chip: GeForce4 Ti4200 with AGP 8x:

The GPU frequency here is 250MHz.

The graphics card is equipped with 128MB of 128bit DDR SDRAM memory from Samsung with 4ns access time.

The graphics memory frequency is 512MHz (256MHz DDR). According to the manufacturer's specs, the maximum working frequency for 4ns chips is only 500MHz (250MHz DDR). So, the graphics memory here works at an irregular frequency. To put it short, it's overclocked. We may only wonder what prevented NVIDIA from using a bit faster graphics memory chips in their card :).

NVIDIA GeForce4 Ti4200 with AGP 8x reference card is equipped with a DVI, D-Sub and TV-Out connectors. The traditional for the Ti4200 family Sil164CT64 chip from Silicon Graphics controls DVI-output, while the SAA 7104E/V1 chip from Philips forms the video signal for TV-Out.

By the time we worked on the review, the NV18 based graphics cards had hit the stores already. So, we took a mass piece instead of the reference one. The first NV18 based card we got into our lab was a solution from Albatron.

A few words about the Albatron GeForce4 MX480 graphics card.

The card is shipped in a stylish looking retail package:

It includes the graphics card, the user's manual, two CDs with drivers and WinDVD, an S-Video-to-RCA adapter and an RCA cable:

The card follows the design of the NVIDIA GeForce4 MX440 reference card supporting AGP 8x:

   

The card features the NVIDIA GeForce4 MX 440-8x GPU and 64MB of 128bit DDR SDRAM from Samsung with 3.6ns access time:

   

The chip and memory working frequencies comply with those recommended by NVIDIA: 275MHz/512MHz (256MHz DDR) respectively.

The cooler on the GPU deserves a closer look: it covers not only the graphics core, but also the graphics memory chips. It's not that the memory chips in such a graphics card required any special cooling, or the GPU needed such a big cooler, but this cooling system does look impressive anyway.

As for the name of the card, it is not the best thing Albatron could do, actually. The number "480" in it may confuse the customer into thinking that this graphics card would work faster than NVIDIA GeForce4 MX460, which is not true. It's sad, but another graphics card makers may follow Albatron's example.

However, Albatron GeForce4 fully complies with NVIDIA's reference card in terms of frequencies and design and arouses no complaints about its functioning, which is just what we need now :).

Testbed and Methods

The most popular AGP 3.0-supporting chipsets today are SiS648 for Intel Pentium 4 and VIA KT400 for AMD Athlon XP.

So, the benchmarks were run on two systems:

The SiS648 based testbed was configured as follows:

• Intel Pentium 4 2800MHz CPU;
• ASUS P4S8X (SiS648) mainboard;
• 512MB PC2100 CL2.5 DDR SDRAM by Samsung;
• IBM DTLA 305030 HDD.

The VIA KT400 based testbed was configured as follows:

• AMD Athlon XP 2700+ CPU;
• ASUS A7V8X (VIA KT400) mainboard;
• 512MB PC2100 CL2.5 DDR SDRAM by Samsung;
• IBM DTLA 305030 HDD.

We used the following software:

• Detonator 40.52 for Windows XP;
• Windows XP:
• DirectX8.1;
• Quake3 Arena v. 1.30;
• Serious Sam: The Second Encounter;
• Unreal Tournament 2003 v.2107.

The higher data-transfer speed of AGP may lead to performance growth in "extreme" modes: in high resolutions, with turned-on full-screen anti-aliasing and anisotropic filtering.

In this case, some textures may not fit into the local graphics memory and the GPU will have to use AGP-texturing, that is, take samples from textures that are stored in system RAM. Quite naturally we expect that doubled AGP bandwidth in 8x mode will improve the results significantly.

On the other hand, we shouldn't expect these situations to happen often. Game developers do their best to avoid AGP-texturing, as the 4x AGP bandwidth (1064MB/sec) is much lower than the graphics memory bus bandwidth. For example, with 128bit DDR SDRAM working at 400MHz (200MHz DDR), the peak graphics memory bus bandwidth makes 6400MB/sec.

Besides the data, transferred via the AGP bus during the textures loading or AGP-texturing, the GPU also receives the descriptions of polygons vertexes. Transferring the polygon data via the AGP bus in low resolutions may take not so little time compared with what the drawing of these polygons on the screen takes. In this case, the higher AGP bus bandwidth would also positively impact the results.

Taking all this into consideration, we decided to test the graphics cards in both low and high resolutions as well as with enabled full-screen anti-aliasing and anisotropic filtering.

We used the following settings in the games.

Quake3 Arena:

32bit screen and textures color depth. Maximum graphics quality settings. Tri-linear filtering and texture compression enabled.

Serious Sam: The Second Encounter:

32bit screen color depth. "Quality" graphics quality settings.

Unreal Tournament 2003 v.2107:

Default graphics quality settings.
 

3D Performance: SiS648 Based System

NVIDIA GeForce4 MX440 with AGP 8x

We will start in ordinary modes, without anti-aliasing and anisotropic filtering:

Unreal Tournament 2003 pays no attention to the change of the AGP mode and data-transfer rate. There is no difference at all in performance of NVIDIA GeForce4 MX440 with AGP 8x in different AGP modes.

But the NVIDIA GeForce4 MX440 based card overclocked to the frequencies of the new 8x one cannot catch up with it, though. The reasons can be numerous: different memory timings, BIOS's, different GPUs…

When running at its nominal frequencies, GeForce4 MX440 loses notably to GeForce4 MX440 with AGP 8x. It's really not surprising, as the latter has higher nominal frequencies.

It's all the same: NVIDIA GeForce4 MX440 with AGP 8x can boast no advantage in AGP 8x mode. The tiny difference in results can be written off as a measurement error.

GeForce4 MX440, overclocked to the level of the competitor, still falls behind it a little bit.

At nominal frequencies, GeForce4 MX440 appears considerably slower than GeForce4 MX440 with AGP 8x because of the difference in their frequencies: the graphics card based on GeForce4 MX440 with AGP 8x has 1.8% higher GPU and 28% higher graphics memory clock-rates.

No changes here: AGP 8x provides no evident results improvement.

It looks as if testing in "normal" modes hardly loads the AGP bus. All the textures stacked up nicely into the local graphics memory and the graphics cards didn't have to resort to texture uploading and AGP-texturing. The transfer of polygons didn't load the AGP much, too. One way or another, the graphics card based on NVIDIA GeForce4 MX440-8x got no real advantage through its higher AGP bandwidth.

Now we will increase the workload by turning on full-screen anti-aliasing and anisotropic filtering. The local memory amount reserved for textures will be smaller and the higher AGP bandwidth may come in handy:

Well, AGP 8x improved the performance in Quake3 Arena and Serious Sam: The Second Encounter, but the actual growth is just laughable.

So, the results indicate that during the tests of our NVIDIA GeForce4 MX440 with AGP 8x graphics card in a SiS648based system we couldn't see any benefit from the new AGP 8x.

However, the increased clock-rates of the card allowed it to beat its predecessor and turned to be the real advantage.

NVIDIA GeForce4 Ti4200 with AGP 8x

The graphics cards based on NVIDIA GeForce4 Ti4200 and GeForce4 Ti4200 with AGP 8x have twice as much graphics memory compared to GeForce4 MX440/MX440 with AGP 8x based cards, so there is even smaller chance that the game would require texture uploading or AGP-texturing.

The benchmarks proved our concerns: AGP 8x mode adds no benefits to the GeForce4 Ti4200 with AGP 8x.

Nevertheless, its increased graphics memory clock-rate provided 10% performance growth over the GeForce4 Ti4200 based card.

Now we are turning on anti-aliasing and anisotropic filtering:

And here again we can't see any effect of the praised AGP 8x.

So, the tests of the new graphics cards on SiS648 make it clear that for now they have only one advantage over their predecessors: higher working frequencies. Enabling AGP 8x mode led to no performance growth, at least in case of a SiS648 based platform.

3D Performance: VIA KT400 Based System

NVIDIA GeForce4 MX440 with AGP 8x

Unreal Tournament is still indifferent to all our manipulations with the AGP modes.

There is definitely something wrong with Quake3 Arena. Why does AGP 3.0 provide poorer results than AGP 2.0?

Note also that the higher is the resolution, the smaller is the gap between the AGP 3.0 card and its rivals. As we already know, the cards don't use AGP-texturing here, so the problem with speed may only be connected with the polygons transfer. The number of polygons in the scene is the same in all resolutions, as well as the time required to transfer the polygons via the AGP bus. That's why when the resolution is higher and it takes longer to build all frames, these constant delays become less prominent and affect the overall result not that greatly.

The story goes on: whatever the data transfer speed is, the graphics card does worse in AGP 3.0 mode, than in AGP 2.0 mode. The lower is the resolution, the bigger is the gap in results.

Well, there is a certain solace, though. The GeForce4 MX440 with AGP 8x has higher working frequencies, so it always outperforms the regularly clocked GeForce4 MX440.

If we compare NVIDIA GeForce4 MX440 with AGP 8x with GeForce4 MX440 in AGP 2.0 mode and with equally set frequencies, the new one will turn a little faster.

With the enabled full-screen anti-aliasing and anisotropic filtering the new graphics card still does worse in AGP 3.0 mode than in AGP 2.0. But here the gap is smaller than in the previous tests, due to higher workload on the graphics cards. Certain differences in BIOS, memory timings and GPU design allow the new card to perform better even in AGP 3.0 mode compared with its predecessor "overclocked" to the level of GeForce4 MX440 with AGP 8x.

NVIDIA GeForce4 Ti4200 with AGP 8x

NVIDIA GeForce4 Ti4200 with AGP 8x on the VIA KT400 based testbed behaves the same as GeForce4 MX440 with AGP 8x.

Full-screen anti-aliasing and anisotropic filtering increase graphics card workload, so the performance gap in AGP 3.0 mode gets smaller than in case the tests are run without the enabled FSAA and anisotropic filtering.

Conclusion

So, the test results suggest that you'll hardly get any performance boost in present-day games with AGP 3.0 8x and double AGP bandwidth. Even if the performance gets any better, this will be a really insignificant improvement.

When rolling out renewed GeForce4 Ti and GeForce4 MX with AGP 8x support, NVIDIA made a very wise move having considerably increased their clock-rates. The new graphics cards turned out to owe their performance growth to the higher working frequencies, and not to the AGP 8x support.

Anyway, the company once again managed to present its developments as something brand-new and earn a lot on its "old-new" GPUs with the minimal effort applied.

And what a nice opportunity for graphics cards makers! A colorful package, magic "AGP 8x" letters and a name like "GeForce4 MX480", as we see by the Albatron card: all this helps to entice the customer. And there will be customers for these "brand-new" and "up-to-date" graphics cards…

The same makes sense for chipset and mainboard makers. AGP 8x is a nice opportunity to roll out "ultra-modern" products. And if we can't say anything bad about SiS648 based mainboard, since the graphics card worked in AGP 3.0 mode at least not worse than in AGP 2.0, the functioning of the VIA KT400 based board couldn't keep us silent.

When AGP 3.0 mode was enabled, at any data-transfer rate, all the changes in AGP 3.0 specs are forced into action. And it's not important whether the OS "knows" about AGP 3.0, or not. In other words, the problems with AGP 3.0 in KT400 may lie on the hardware level, i.e. in the chipset itself, as well as in BIOS, drivers or traditional "4-in-1" from VIA.

But really, I don't care about what and where they did wrong with the chipset. Imagine that I got so much impressed with AGP 3.0 features that I made up my mind to buy a VIA KT400 based mainboard, an AGP 8x graphics card and where am I? AGP 8x does work, but the speed would be lower than in AGP 2.0 4x. Hm, doesn't seem to be what I have longed for?

There is still some hope left that the problems of KT400 have nothing to do with hardware, but lie in software, and can be cured in the new BIOS, "4-in-1" versions and so on. Otherwise, we may witness another proof of VIA's chipsets development law: rapid introduction of some "KT400A".

But what about the AGP 8x standard? The higher bandwidth will only be useful when there are really big amounts of data pumped through AGP and AGP bandwidth is a crucial parameter for games. It may come, as there appear DirectX9 games that will be able to load the newest ATI and NVIDIA products with data amounts proportional to their powerfulness.

The AGP 8x story doesn't end here. There are two more AGP 3.0 supporting chipsets: NVIDIA nForce2 and Granite Bay from Intel. Moreover, we have only tested graphics cards based on NVIDIA's chips, while there are new AGP 8x cards based on ATI RADEON 9700 PRO/ RADEON 9700/ RADEON 9500 PRO/ RADEON 9500 and on SiS Xabre 600/400. Stay with us for more! :)