Interview with ATI´s Eric Demers
November 6, 2003 / by Leonidas / page 1 of 1
Our interview partner is Eric Demers, Hardware Engineering Manager from ATI.
3DCenter: With the recently published Catalyst 3.8 driver set, ATI offers a new feature called OverDrive. It operates only with chips that incorporate a thermal diode. Does ATI plan to ship this functionality with all upcoming products, or will it be limited to the enthusiast market. Do you also plan to implement a fan throttling feature that allows silent 2D operation, or is this left to AIB manufacturers?
Eric Demers: OverDrive is a great new feature that we've introduced with the 9600 XT and the 9800 XT. It allows us to do "safe" overclocking, by monitoring the graphics' chip temperature. With this information, the driver can adjust the frequency of operation of the core graphics and memory clocks. While temperature information is required, it's not enough to know the "true" limits of a particular chip (we would also need to know exact process information on that chip). Given that, the OverDrive needs to be tempered a little bit (i.e. overclockers could still achieve higher results, but this would void their warranties). That being said, we now have the technology to implement OverDrive in our future products, as we see fit. And in fact, our current mobile products use PowerPlay in similar ways, to lower the clock frequency when the graphics load is light. I can't say specifically what future products will be OverDrive enabled or not, but when we introduce a new product, check to see if its OverDrive enabled!
The 9800 XT's already support a variable fan, and will throttle back that fan when the temperature is low, and vice-versa. We have to be careful to maintain a very silent operation when using this. We don't want our fan to become a leaf blower when you start running some 3D apps! And as far as perfectly silent operation, even in 2D there's some heat that you need to dissipate. There are various cooling solutions out there that would allow silent operation (I use a Zalman on my 9700p), at least in 2D. Our AIBs are certainly welcome to investigate those (and at least one has used this), as is ATI.
3DCenter: Over the last years desktop graphics cards showed a trend towards ever-increasing power consumption, which results in sophisticated cooling solutions and additional power connectors on the cards, severely stressing PSUs. Conversely, the mobile market is served with very efficient, power-saving devices. Is ATI considering to bring similar technology to desktop cards, or do you think this is too much of a niche market?
Eric Demers: On the mobile side, we are getting more and more elaborate to try to keep the power consumption down, while increasing the overall performance. It's probably the largest focus of the mobile market. Now, the desktop graphics' market has been trending upwards in power consumption, at least on the high-end systems. The midrange has been a little more conservative, and a 9600 non-pro ships with passive cooling. The increase in power consumption has come from running larger dies at higher clocks.
In general, with shrinking processes (150nm, 130nm and now 90nm), active power has gone down per transistor (smaller transistors and voltage lead to less power outputted when switching), but our transistor counts have increased at a larger rate, so that the active power has kept on going up. But with 130nm and even more with 90nm, static power is getting higher and higher. In a normal CMOS process, one can think that if a transistor is not switching, it's not using power. That's roughly right, but there is something called "leakage", which is a small current that leaks out of a transistor, even when not switching. As dimensions have shrunk, leakage currents have gone up by about an order of magnitude each time. This is leading to the problem that non-switching power is getting bigger and bigger. With 90nm, it's going to more than ½ the total power. This power is directly proportional to the number of active transistors, or plain die size. Consequently, as we shrink down, active power sort of goes down, but passive power goes up. With 90nm, power consumption is going to be a "Big Deal", even when the chip isn't doing anything.
Having said all that, even the high-end desktop parts cannot keep on increasing their power consumption indefinitely. Consequently, we are looking at power saving options to reduce the desktop cards' consumption. Both the mobile technologies as well as others are being considered. But we don't want to handicap desktop chips from a performance standpoint. It's a fine line that we are keenly aware of.
3DCenter: The current MacOS X drivers from ATI can optionally enable supersampling on R3x0 based cards. Will this feature be included into Windows drivers?
Eric Demers: It's a possibility, but is currently not in the plan. The driver support requirements on the Windows side are enormous for that feature, which does not make it very appealing. As well, the current AA quality we offer cannot be matched with a supersampling algorithm, since a SS algorithm would not be able to deal with gamma correction in the same way that our current AA algorithm does. The only advantage right now to SS would be for alpha on textures. That does not currently seem like a worthy trade-off.
3DCenter: With the growing use of complex pixel shader effects, the likeliness that some of them introduce aliasing artifacts increases. Will it be the full responsibility of the developers to work around this issue, e.g. with gradient instructions, or will ATI support them with new features?
Eric Demers: Our MSAA algorithm currently operates perfectly with shading, since we offer both center sampling and centroid sampling, on a per texture basis, to deal with some of the artifacts of shading. As far as procedural texturing aliasing goes, yes, it can be an issue in certain circumstances, where the procedural texture has too much high frequency information that leads to aliasing. There are things that the developers can do to reduce the artifacts, such as rendering to a high resolution texture, then low pass filtering that before using it. Developers could also just generate textures that don't have as much high frequency information. In the current generation of games, this hasn't been a problem. For the future, I can imagine that it might become a bigger deal. And, yes, we will support our developers with tons of features (such as centroid sampling), generally well in advance of when they will need them. At the end, however, it is the ISV's responsibility to decide what their game looks like.
3DCenter: The Radeon 9800 series introduced a new feature, the F-Buffer. Is this rather a stop-gap solution, or will it be ATI's efficient way to handle the challenge of long, dynamic shaders in the future?
Eric Demers: The F-Buffer concept is method to generalize multi-pass and actually make it useful. And you need multi-pass solutions to deal with very long shaders and other challenges such as larger temp storage or larger number of iterators, etc. However, it's not the panacea that just allows for limitless shading either. The way we are planning on exposing it (through an OpenGL extension) allows the savvy developer to use them. We are working on improving its ease of use for future products. As well, there are other solutions we are looking at that make it easy for developers to write, for example, longer shaders. I'm sure that future products will have those features as well.
3DCenter: With the 6x sparse multisampling offered by Radeon 9500+ cards, ATI set a new standard in antialiasing quality on consumer hardware. Do you feel there is a need for another increase in quality in the near future, or is the cost for even better modes too high to justify the result?
Eric Demers: There's still not a single PC solution out there that gives AA quality like we do. Not only do we have a programmable pattern (which is currently set to sparse), we are also the only company offering gamma-corrected AA, which makes for a huge increase in quality. Honestly, we've looked at the best 8x and even "16x" sampling solutions, in the PC commercial market, and nobody comes close to our quality. But we are always looking at ways to improve things. One thing people do have to realize is that if you use a lossless algorithm such as ours (lossy ones can only degrade quality), the memory used by the back buffer can be quite large. At 1600x1200 with 6xAA, our buffer consumption is over 100 MBs of local memory space. Going to 8xAA would have blown passed 128MB. Consequently, with a lossless algorithm, the increase in subsamples must be matched with algorithm changes or with larger local storage. We've looked at things such as randomly altering the programmable pattern, but the low frequency noise introduced was worst than the improvement in the sampling position. Unless you have 32~64 subsamples, introducing random variations is not good. So we are looking at other solutions and algorithms. Current and future users will be happy with our solutions. Stay tuned.
3DCenter: The Parhelia chip from Matrox offers the so-called "Surround Gaming" feature. Do you think this is just a gimmick, or is it bound to become the common form of gaming over the next years? How do you rate alternatives that provide the gamer with increased immersiveness, like e.g. 3D glasses?
Eric Demers: I can't really speak for ATI on this. Our products might support these features, if the company deems them worthy and cost effective. In my opinion, the "surround gaming" feature sounds cool, but it generally impractical, so it will end up being either niche or even "gimmicky". I don't actually think that it's a selling feature. As well, it adds to the packaging costs and board costs, so that every user has to pay for that feature, which isn't great. As far as other "immersiveness" items, such as 3D glasses or 3D LCD displays, they have their place. I've used 3D glasses on high-end systems, way back at SGI, and they were very cool and worked well. The affordable glasses that I've tried more recently have left me lukewarm on them, since the 3D effect is not always there, and the low fps (1/2 peak) removes much of the "immersiveness". I hope that the technology keeps on improving, since that kind of experience is great, when it works.
We like to thank Eric Demers for the interesting conversation and his detailed answers.