Saturday, July 23, 2005

An Eye on DLP, No. 1

Samsung HLN617W

Digital Light Processing (DLP) is one of the big "players" in HDTV technology. Texas Instruments invented it in 1987 and introduced it commercially in 1996. In 2002, TI introduced the first DLP implementation that boasted hi-def resolution: 1280 x 720 pixels. (720 pixels is the minimum vertical resolution considered high definition.)

This year, 2005, inaugurates DLP technology that (using a clever trick) simulates full-fledged 1920 x 1080 resolution, for 1080i and 1080p support. Full-fledged 1080p resolution (or, actually, at 2048 x 1080 pixels, the slightly better "2K") is now used in digital cinema projectors in commercial theaters.

A 61" Samsung HLN617W DLP-based rear projector, a 2003 model, sits in the living room of yours truly. It has 1280 x 720p resolution.

HD2 "chip"
Texas Instruments

TI makes the integrated circuits that make DLP images; other companies such as Samsung, LG, and RCA build these chips into actual TV products. TI's so-called HD2 or "Mustang" chip, an early incarnation thereof, is inside my Samsung. This chip has been improved more than once since 2003 and is now called HD2+.

In the middle of the chip, as you can see, is a silvery window the size of a small postage stamp. The window reproduces the 16:9 aspect ratio of the TV screen. Inside this window is the "guts" of DLP technology: the Digital Micromirror Device, or DMD. It's not one mirror; it's 921,600 tiny mirrors arranged in 720 rows of 1,280 micromirrors each.

How a DLP rear-projection TV works
The diagram to the right shows how the DMD chip makes a picture on the screen of a rear-projection TV. (Click on it to see a larger version.)

Each individual micromirror is either on or off at any given moment. If on, light from the light source strikes it and bounces through the projection lens onto the screen. If off, the micromirror swivels off-axis on its tiny hinge, and light reflected from it strikes only a light-absorbent material, ideally never reaching the screen.

At any given instant, every micromirror is assigned a primary color, red or green or blue, by virtue of the white light from the light source passing through a filter of that color built into a rapidly spinning wheel. As the wheel spins, the three colors alternate so rapidly that the eye (usually) cannot tell that the TV is producing a red, then a green, then a blue picture on the screen. Instead, the eye sees a full-color image.

A revised image is flashed on the screen numerous times during each video frame. A video frame lasts 1/30 second, but this time period is divided into (at least) three subframes of red, green, or blue. Actually, in most DLP implementations these days, the color wheel has two red segments, two green segments, and two blue segments — plus one or two extra segments I'll get to later — so each video frame's time period is subdivided even further than that.

What's more, each primary-color subframe is further split into multiple time slivers in order to produce a grayscale. More on that below.

Question: how do you render a broad enough range of grays (mentally factoring out hue entirely) when the micromirrors can be only on or off — nothing in between?

Answer: you subdivide each primary-color subframe into several time slices or intervals. Then (let's assume you want the individual micromirror to produce a 50% gray, ignoring hue) you arrange for the micromirror to be on during half of those time slices and off for the other half.

Basically, a technique called binary pulse width modulation (PWM) is used to decide during which intervals, and for how many of them, the mirrors will be on. For instance, for 5o% gray the interval whose duration is exactly half that of the subframe is "on," the other intervals "off." For 25% gray, the interval which is one-quarter that of the subframe is the only interval that is "on." For 37.5% gray, the 1/4-subframe (25%) and 1/8-subframe (12.5%) intervals are both activated. And so on.

The intervals that are activated for a particular level of gray are added to get their total duration. Then that duration is divided into many, many ultra-wee time slices which are parceled out evenly over the entre period of the primary-color subframe. This technique is called "bit-splitting."

Bit-splitting minimizes the effects of certain objectionable grayscale artifacts which some people are able to see when their eyes move to follow an object traveling across the DLP image. (I'm getting much of this information, by the way, from "DLP from A to Z," an excellent article by Alen Koebel in the August 2005 issue of Widescreen Review magazine. Magazine subscribers may download the article in PDF form by clicking on the appropriate hotlink on this page.)

One of the few drawbacks of DLP technology is its "rainbows." Just as swift eye movement can introduce grayscale oddities, it can cause the seemingly full-color picture to fractionate, just for an instant, into distinct red, green, and blue images. What is happening is that the primary-colored images flashed to the screen in rapid succession are getting fanned out, as the eye moves, to different areas of the retina. The neural pathways of human vision can't integrate the images' colors into, say, the intended teal or lemon or mauve.

Some people are more sensitive to "rainbows" than others. For some people, they completely ruin the experience of watching a DLP display. For others — I'm lucky enough to be in this group — they don't happen all that often, if at all.

To minimize the "rainbow" effect, TI learned two tricks early on. First, put two wedges of each primary, red, green, and blue, on the color wheel, not just one. For each rotation of the color wheel, there are accordingly two subframes of each primary color. Second, make the color wheel spin real fast, so that each primary-color subframe is real brief, and there are lots of primary-color subframes within each video frame's timespan. (Those micromirrors will have to pivot a lot faster, of course, but what the hey?)

Using 3 DMDs
There's even a way to eliminate the "rainbow" artifact entirely: use three DMDs, not one.

You use beam-splitting prisms to route white light from a light source to three DMD chips operating in parallel. One beam contains red light for the red image, one green-for-green, and one blue-for-blue.

Thus, three images are formed independently by their respective DMDs; each is given its own primary hue. The three primary-color images are optically combined and shepherded together to the screen.

There's no color wheel in 3-chip DLP. Nor does the video-frame time interval have to be as finely chopped to allow for multiple primary-color subframe intervals — which makes things a lot simpler and allows a more smoothly gradated grayscale.

The problem with 3-DMD DLP projectors is that they're much more expensive than 1-DMD projectors. As a guess, I'd say that the optics alone cost more — imagine having to make sure that three beams carrying three images line up just right.

The real cost boost is in adding two more chips, though. The cost of making DMD chips is fairly high because the manufacturing yield — the percentage of chips that aren't rejected as imperfect — isn't all that high, yet. (As expected, the "yield curve" rises as TI's experience with making the chips accumulates over time.) So when you triple the number of chips, you add quite a lot to the cost of the projector.

As far as I know, there are no 3-chip rear projectors. Only front projectors (including those for digital cinema) are expected by customers to be pricey enough to warrant going to three DMDs.

The Achilles' heel of DLP is its inability to render deep blacks, in the way that CRTs, on the other hand, give deep, satisfying blacks.

No fixed-pixel technology — not DLP, not plasma, not LCD or its variants, LCoS, D-ILA, and SXRD — does black well. As the Widescreen Review article says, "The evolution of DLP technology, since its commercialization in 1996, could be described, as much as anything, as a search for better blacks."

Each fixed-pixel technology has a different reason for weak blacks. In the case of DLP, poor blacks are due to light from the light source that reaches the screen when it shouldn't.

Imagine an all-black image. The uniform-black signal causes every micromirror to turn off, meaning that it swivels away from the direction that carries reflected light to the screen. Instead, all the light is supposed to be beamed into a light absorber.

Well, what happens if the light absorber doesn't absorb all the light? The remaining light from the light source ends up bouncing around ("scattering") inside the RPTV or front projector ... and eventually arrives at the screen, polluting the image by washing it out just slightly.

A second problem has to do with the spaces between the micromirrors. Light reaching these gaps can easily end up bouncing off the DMD's underlying structure and eventually reaching the screen.

A third problem concerns the tiny "dimple" or "via" in the middle of each micromirror where it attaches to its hinge. The larger the via, the greater the amount of light that bounces crazily off it and reaches the screen when it shouldn't.

Problem One, the scatter problem, was addressed early on, at the time of the first commercial DLP displays, by providing the best light absorber possible. Those early DLP displays had an "on-off contrast ratio" of about 500:1. That is, an all-white field produced 500 times more luminance at the screen than an all-black field.

Problem Two, the gap/infrastructure problem, was addressed in two stages. The first, says WR, "was to coat the hinges and surrounding structure under the mirrors with a low-reflectance (light absorbing) material." That "dark metal" technology was dubbed DarkChip1, and it boosted on-off contrast ratio to 1,000:1. Later on, DarkChip3 reduced the size of the gaps between the mirrors, and contrast ratio went up even further.

In between DarkChip1 and DarkChip3 came DarkChip2, which addressed Problem Three, the dimple/via problem. The dimple or via was simply made smaller. That boosted the contrast ratio by giving less area of irregularity on each micromirror off of which light could bounce askew. That change alone deepened black renditions, improving contrast. And it further boosted contrast ratio by offering more area on each micromirror that could properly reflect the source light, making the overall picture brighter.

DarkChip3, when it came along, made the dimple or via yet smaller than DarkChip2 did, in addition to shrinking the inter-mirror gaps. Now, with DarkChip3, DLP black levels are much more respectable than in earlier versions. In fact, they're way better than those on any other non-CRT display.

One of these DarkChip revisions — I'm not quite sure which — made the micromirrors swivel further off-axis in their off position than they had before, reducing image-polluting light scatter within the display or projector. In the jargon, the "micromirror tilting angle" was increased from 10° to 12°. My guess is that this happened at the time of DarkChip1, since the tilting-angle increase surely exposed more of the structure under the mirrors.

More on DLP technology in An Eye on DLP, No. 2.

No comments: