Integrated graphics will kill discrete

Posted: 06.20.05
FIGURE 1. Simple projection of Intel's IGP growth in market share.
(Source: Jon Peddie Research)

I've made a lot of predictions in my 20+ years here at JPR/A, and fortunately most of them have been correct, making profits for our clients and allowing me to keep my job. One prediction I made in 1998 as SGI was about to launch the Cobalt IGP (which they called Integrated Visual Computing, IVC, architecture) for their 320 and 540 workstations was that it was a harbinger of graphics processors for the future. Mistakenly, sitting at Dave Orton's desk (he was senior VP and in charge of SGI's workstations at the time), I said integrated graphics will kill discrete. He's never forgotten it and says it in public from time to time when he thinks I need to be put in my place.

I also have the privilege of having told Jen Hsun-Huang repeatedly while he was still at LSI Logic, Don't go into the graphics chip business.

Orton left SGI and took over the reins of ArtX, which at the time had a contract to build the next generation of graphics for Nintendo (with the team at SGI that built the first graphics for the Nintendo 64 ), which became the Flipper chip. That design too was an integrated part, and extending the concept further, in 1999 ArtX developed the 3D graphics processor (the first IGP for the PC) for Acer/ALi's new Aladdin 7, a 128-bit Socket 7 integrated graphics chipset for AMD K6-II and K6-III processors.

Did this not clearly show the way of the future? From workstations to game consoles, integrated graphics were being developed and entering the market. Integrated graphics were not a novel idea—SiS offered its SiS6205 IGC in 1995 and VIA had its Apollo MVP4 in 1998. ATI announced its integrated graphics and core-logic program in the spring of 1999, and Nvidia and ALi introduced the Aladdin TNT2 integrated graphics chipset with ALi's Slot 1/Socket 370 north bridge design in the fall of 1999. And in 1999, Intel introduced the 752 core in the 810 and 815 SDRAM IGC chipsets for the low-end Celeron PCs; although they did have the Intel 430VX PCIset in 1996 (and, in those innocent days, it was referred to as unified memory architecture—UMA; it is now called Shared Memory Architecture—SMA). And just to be pedantic for a moment, I have to mention that Dell was actually the first company to try it with an experimental ASIC it developed in 1991 called JAWS and it used UMA for graphics memory. All these parts are IGCs, integrated graphics controllers, and not Integrated Graphics Processors (IGP), which, as the difference in name implies, offers programmability.

It took ArtX to prove that an IGP wasn't going to work in the high end. The ArtX IGP had T&L, a notion that was just beginning to get traction in discrete PC parts with the introduction of Nvidia's Riva in early 1999 (the workstation folks like 3Dlabs had it for years).

What ArtX found out, the hard way, was there just wasn't enough bandwidth in the systems of 1999 to do a UMA, and so the part was dropped, and no one came out with a true IGP again until Nvidia announced its Nforce for the AMD platform in 2001.

In the meantime Intel steadily took market share with their IGC from the discrete segment, growing from zero in 1999 to 55% by the second half of 2001—that got a few people's attention. And in Q1'05 the company surprised everyone with their dramatic growth in integrated graphics in mobile platform due their remarkably successful marketing campaign with Centrino—possibly the most successful campaign since Intel Inside. If you project the curve they will own the entire market in a short matter of time (see Figure 1).

But can they sustain it, and is it the death knell for discrete?

I don't think so, and here are the reasons why.

Party's over. Intel was super successful with the Centrino marketing program, and surprised everybody in Q1, including themselves. They've gotten that pop, and it's winding down—every-one who wanted a Centrino has one, so all that's left is the late adopters.

New competition. Two powerful companies have entered the IGP segment, ATI and Intel. That has had a couple of impacts.

1. There are eight market segments in the PC: Desktop and Mobile—which are split into AMD and Intel, which are split into Discrete and Integrated. Intel participates in two of the eight, and ATI and Nvidia participate in all eight.

2. AMD is growing market share. AMD-based IGPs will grow with it, decreasing Intel's overall market share in Integrated.

3. An incumbent in any market always loses market share when new companies enter.

4. ATI and Nvidia, as well as VIA/S3, will introduce much more powerful IGPs than Intel and thwart any growth by Intel into a higher-performance category.

Cached GPU. Nvidia introduced the concept of using system memory with an external GPU (UMA again), and called it Turbo Cache. ATI matched them and introduced Hyper Memory. Both concepts rely on the high bandwidth of PCIe (and remember, bandwidth was what stopped higher performance IGPs in the past) and allow for the abandonment of local private frame buffer memory. This will further stop the growth of IGP and offer a much greater price/performance ratio for a mid- to low-end mainstream solution.

Longhorn. To get all the graphics richness Longhorn is promising the systems of 2006 will have to have powerful GPUs and lots of overlay memory. IGPs can't do it, and so there is a special limited class of Longhorn for them and lower power or older systems.

Need for high end. There will always be a need for high-end graphics. The challenge the GPU suppliers have been facing is how to expand the market. When the lame-ass game developers catch up with GPU development and start to offer more challenging and visually stimulating games, GPU sales will rise. There're 25 to 30 million users out there just waiting for a good reason to buy up.

Jon Predicts

So here I go again out on a limb with a saw—just can't help myself it seems. First, let me be very clear about what I'm saying. Intel recently suggested they could hit 80% market share on desktop graphics (based no doubt on their position of 72% in integrated desktop parts Q1'05 and steady growth of share—see Figure 2). I'm saying they can't do it, for the above reasons and a few more. However, I'm not saying they are going to crash and lose all the IGP market—they'll still do damn well, and have significant share. Intel has substantial marketing and sales power, branding, co-op marketing dollars, bundling deals, one-size-fits-all motherboards, and good old-fashioned arm-twisting. I am saying I think they've hit the peak and they will start to see market share loss, albeit not significant loss, but a break in the slope of the curve.

FIGURE 2. Jon's crystal ball forecast. (Jon Peddie Research)

Integrated will continue to steal market share from Discrete until demanding applications show up that make better use of the GPUs. ATI and Nvidia will participate in the erosion of the GPU space with their own IGPs and so the migration won't kill the two companies that are single handedly providing all the development in 3D and graphics today. Intel will ride that pony for free basically, and enjoy the power of its brand.

Like I said, don't go selling your Intel stock just yet, but also don't go believing there's no limit to the company's growth in the integrated market.