The i740: Intel's Lead Architect Talks

This may be the year when we see some big-time thinning in the 3D accelerator herd. Intel's entry into the 3D accelerator chip world, the Intel740 (or i740), has upped the ante for all companies competing in the market. Not even current leaders in the market like 3Dfx and nVidia are underestimating Intel, and companies like Cirrus Logic have already decided that success in 3D accelerator market might be beyond their reach. Certainly the introduction of the Intel740 has made this already crowded market much less appealing to second and third tier 3D chip makers.

The first card on the market to use the Intel740, the Real3D Starfighter, has shown impressive performance. Wide OEM support for the chip means many different flavors of i740-based graphics cards will appear on the market this year. Some analysts have said that by the end of the year, Intel could control 20% or more of the 3D accelerator market.

I recently interviewed Jay Sturges, the platform architecture manager at Intel who's responsible for the planning and development of all graphics platform technical programs related to graphics chips. Jay gave a glimpse into some of the company's i740 design decisions, and explained what the future holds for Intel's thrust into the consumer graphics hardware business.

What prompted Intel to jump into the 3D chip market in the first place?

We continued to see that the graphics accelerator market (especially 3D) was not balancing the Pentium II processor and the overall platform. There's a Pentium II processor which can deliver bucket loads of 3D geometry, yet graphics accelerators could only consume spoonfuls. This lead us to develop the Intel740. It's a high quality, high performance 3D accelerator (not to mention 2D and video as well) which was designed to balance the Pentium II processor platform.

So Intel felt the current 3D market leaders like 3Dfx, Rendition, PowerVR, and nVidia weren't reaching their potential?

Our intent was to deliver balance between performance, high quality graphics, and price to the mainstream segment of the market. In addition to 3D, this segment requires a balance of 2D and video performance. It requires all of this, and yet the graphics accelerator can't become expensive. Dedicated 3D-only chips or solutions not balancing all requirements do not solve this problem.

Which graphics board OEMs are you working with, and when can consumers expect to see the first boards on the market?

To name a few, AUStek, Diamond, Elsa, Leadtek, Number Nine, STB, and Real3D. Our list of graphics boards OEMs has grown fairly quickly. There are many more world wide offering the Intel740 on a board. I expect to see the first boards on the market in late March. You should see systems equipped with Intel740 based boards shortly after that.

What APIs does the chip support? Will Intel develop a chip-specific API like 3Dfx's Glide for optimal performance?

We support Direct3D and OpenGL. In fact, our natural data type for the Intel740 is identical to Direct3D's. So Direct3D spits out a TLvertice and the Intel740 takes the structure in directly without modification. We looked into adapting Glide (in fact we have a license to it) for AGP support, and determined that it started to look like OpenGL. By going with the standard APIs we feel we can deliver the highest quality and performance while not further fracturing the market with proprietary APIs

Will the i740 fully compliant with all Direct3D features? If not, what features won't it support and under what conditions is this true?

Because there are so many different combinations and revisions of DirectX it's better to say what the Intel740 supports. We fully document all the capabilities in Direct3D (and OpenGL) we support within our Software Developer's Manual. In addition we have a complete discussion on the performance considerations to a software developer. The manual can be downloaded online at http://developer.intel.com/design/graphics/manuals/.

What is your opinion of tiled rendering architectures like that of the PCX2?

I'm not closely familiar of the PCX2 implementation. Tile rendering approaches ultimately try to minimize the bandwidth required to generate a sequence of scenes. The key limiter to wide spread adoption is its methodology on content creation and the inherent unknown quality in real-time playback.

Having a number of your own fab plants must be a tremendous cost and scheduling advantage for the i740. How much do these kinds of advantages help you compared to companies that contract time at external fab plants?

Intel's fabs are world renowned in performance, quality and reliability. Having this base technology for our graphics chips gives us the ability to deliver competitive pricing, high performance, quality and reliability to our customers. For example, we integrated a technology called "scan" into the Intel740. With this, we can achieve about a 300-500 DPM (defects per million) right out the chute whereas current state-of-the-art in graphics accelerators achieves about 5000 DPM, then over 12-18 months they refine software test vectors to lower this value to say 500-1000 DPM. So in what typically takes 12-18 months for others we do instantaneously.

How is Intel approaching the game developer market for support of the i740?

We have been engaged with software developers since 1996 about the Intel740 (formerly code-named Auburn). In mid-1997 we started to deploy sample Intel740 systems along with our Pentium II processor-based platform samples. Our samples are based upon our customer reference design for an Intel740 card. Thousands of these systems have been sent to developers. This is a continual process, as having exciting applications that show the true performance of the Intel740 is what everyone is looking for. Developers all over the world have been working with an Intel740 for now about nine months. We continually elicit their feedback and incorporate this into design plans.

What is Intel's take on the Talisman architecture?

While Talisman is interesting in concept, I see it as a bucket of parts: a number of technologies that might be integrated in different architectures. We did not choose to implement any of these into the Intel740 because we didn't feel they would provide additional benefits for the upcoming generation of 3D applications.

Can you talk about the i740's precise-pixel interpolation feature, and give some examples of how it improves visual quality of games?

This technique is the used to ensure the life-like quality in rendered triangles. We calculate (and re-calculate) the key rasterization coefficients of a triangle (e.g., x, y, z, cr, cg, cb, sr, sg, sb, fog, alpha, s, t, u, v, lod) for each and every pixel. This yields much higher image quality than, say, calculating the fog or alpha value once for the entire triangle. For example, in SHADOWS OF THE EMPIRE from LucasArts, if your ship is damaged, smoke begins to appear. If fog is calculated on a per-triangle basis, the result is large black hole in the middle of your screen. Whereas when fog is calculated on a pixel basis, the result is realistic smoke. Along with this we carry up to 18 bits of sub-pixel and sub-color precision. This results in much higher color gradient quality render on a triangle and finer resolution of perspective correct texture mapping. For example, if I had a game which was designed to use small textures (say 32x32 pixels each) and these textures were being mapped on a large (visible on screen) object, a blocking pattern would result from using lower bits of sub-pixel precision vs. higher. Another example is a small color gradient change to a large screen pixel change. Consider a right triangle the width of the screen, with color 200 on the upper left vertex (in the corner of screen) to color 202 on the lower right vertex (in the opposite corner of the screen). You would expect the color 201 to show up right in the middle of the screen. With lower bits of sub-color precision this isn't the case. In fact the lower right corner may even end up with a different value than 202 (based on interpolation algorithm).

Explain the significance of the i740's parallel data processing.

The setup engine in the Intel740 takes a Direct3D TLvertice data structure directly. Every element within the structure is interpolated in parallel. This allows the highest level of sustained performance over a TLvertice's data. By interpolating the elements in parallel in a single pass, the performance of a game will not degrade significantly while additional rasterization features are enabled. For example, turning on specular highlighting does not force the rendering engine to perform multiple passes.

How does the i740's direct memory execution (DME) work?

DME is the ability to execute textures directly from system AGP memory. There is no need to copy them to local memory, then execute. In essence this is treating AGP as yet another memory channel, increasing the overall effective bandwidth within the Intel740 to 1.3GB/second of bandwidth (525MB/second AGP + 800MB/second local video memory). By keeping and executing textures directly out of system memory, you eliminate the step of copying to local video memory - hence the improved overall performance. Not to mention the fact that there's no need for least-recently-used algorithms to fetch textures from local video memory.

How does Intel feel about 3D benchmarks, like ZD's 3D WinBench? How important are these benchmarks to you, and do you think they're relevant to the real world?

Benchmarks are a good means, when they are put into context, to help consumers measure value of one thing over another without being an expert on the subject. Ziff-Davis has made a great attempt to measure quality and performance in an environment. We also support application-based benchmarking with a focus on application quality and performance. Both forms offer different views. 3D WinBench covers the entire domain of what could be 3D accelerated (like blending, fog, alpha, etc.) and do this in a way which is representative of a wide spectrum of applications but in a controlled environment. Application testing focuses on the matter-of-fact "real-world" experience, yet may not cover a wide spectrum of what is being accelerated by the hardware (e.g., the application may not use fog or MIP-mapping, etc., thereby limiting the value of the test).

Why no geometry acceleration on the i740? Is that something you'd consider in the future?

The Pentium II processor's floating-point engine delivers the best price/performance for game-intense geometry today and into the future. For the mainstream segment of the market, integrated geometry processing is far too expensive for the usage model. Intel's processors will continue to improve geometry performance, so I don't see the future holding dedicated geometry processing in a graphics accelerator.

Are there any plans to integrate i740 functionality into Intel processors, as AMD did with their K6+3D? Why or why not?

We actively look at the best price/performance balance of the overall PC platform. For certain segments of the market, integration has value. We will continue to offer products that deliver the best in value, performance and price to meet our customers' requirements.

Some OEMs are creating PCI versions of boards that use the i740. What's your reaction to this? Did you try to discourage these manufacturers from doing so, or is it of no consequence to you?

The value of AGP is that it allows the graphics accelerator not to be limited to the amount of memory available in the local frame buffer for textures. A PCI card with 12MB of texture memory cannot replace that. Of course, we don't discourage our customers to implement this if they feel there is a value to their end users, but we feel that only AGP can truly unleash the full potential of the Intel740 graphics chip.

NEC recently announced that their second generation of PowerVR accelerators will target a broad range of platforms, from set top boxes and consoles through PCs and into the arcade. As Intel is a founding member of the OAA, the latter at least seems like a reasonable platform for the i740. Does Intel plan to target other platforms?

I think someone could build an very cool arcade box based upon the Pentium II processor and an Intel740 graphics accelerator. It could potentially rival some of the "dedicated" design systems in price/performance and time to market.

Will there be some announcements later this year regarding the i740 and the console or arcade markets?

While it's entirely feasible for a hardware vendor to develop a product based on the Intel740 graphics accelerator that might address these market segments, I don't know of any announcements about such products. Currently our focus with the Intel740 is on the mainstream PC segment of the market.

It seems that graphics chips are evolving an increasingly rapid pace. What are Intel's goals for updating the silicon? Will we see a second generation i740 in '99?

You can be assured that the Intel740 is not a point product. Graphics is a business for us and we will continue to bring graphics technology to market.

Alex Dunne ([email protected]) is Editor in Chief of Game Developer Magazine.