Experts have one question about Nvidia Inc.'s public admission last week that it may offer its own PC processor: What took you so long?
As the only major system player that doesn't offer both a CPU and a graphics processing unit (GPU), Nvidia has no choice today but to get with the program, say experts.
Not having a CPU wasn't an issue for Nvidia when desktop PCs ruled the world and integrated GPUs were too anemic to support 3-D games, HD video, or even Windows Vista.
With powerful GPUs for add-on cards that were embraced by gamer and regular consumer alike, Nvidia's revenues grew from US$158 million in 1999 to $4.1 billion last year.
But the chip market has decisively shifted to notebook PCs and smartphones.
Not only is there no room for add-on cards, but with mobile devices, the GPU's raw speed is less important than size, wattage and playing well with other internal components, says In-Stat analyst Ian Lao.
"You could have the screaming-est device ever made on Earth, but if your Wi-fi is weak because your GPU is too bulky or draws too much juice, then your performance is going to stink," Lao said.
Intel dominates the integrated GPU market, in part because it can attractively price its bundled CPU and GPU. Nvidia is starting to make some inroads into integrated graphics with its GeForce 9400M chip.
But it has made little headway with its Ion bundle (GeForce 9400M plus Intel Atom CPU) because of Intel's twin strengths.
"Intel has bundled Ion out, so it's having trouble getting traction," said independent analyst Rob Enderle.
Intel is also challenging Nvidia on its home turf. Its integrated graphics chips can already support Blu-ray HD video. And Intel is readying Larrabee GPUs for the add-on graphics market.
"Nvidia keeps getting pushed upmarket into an ever-smaller niche, and they don't want to be boxed in," Enderle said.
So if Nvidia must get into CPUs, how should it proceed? Especially in order to avoid the mistakes of the three most recent failed challengers to the Intel/AMD duopoly, Cyrix, PowerPC and Transmeta Corp.?
Analysts offer these four suggestions:
Don't build a CPU from scratch
Nvidia may claim to have the smartest graphics engineers on the planet, and it may claim that GPUs are morphing into CPUs. But it would have to "commit hundreds of millions of dollars in R&D and lots of time" to build its own CPU from scratch, said Nathan Brookwood, an analyst with Insight64. "[And] while you're busy doing that, the market would have moved on," Brookwood said. Nvidia wouldn't just be starting far behind Intel and AMD -- it would be running straight into a "minefield of potential engineering problems," said Brookwood. For most of its 30-year existence, the x86 processor "evolved without a rigorous architectural definition. As a result, engineers creating a newer version had to ensure that it was compatible with all of the weirdnesses and bugs of the older one."
Both Intel and AMD have a strong handle on x86's quirks due to the "historical memory" of the thousands of engineers they employ, as well as the many systems they have for testing and development, said Brookwood.
These were key factors that hurt Cyrix when it unsuccessfully tried to challenge Intel in the mid-1990s, he said.
A reasonably successful maker of low-cost knockoffs to Intel's 386 and 486 chips, Cyrix's problems started when it touted its M1 processor as faster than Intel's Pentium.
Brookwood said that advantage, whether real or not, quickly vanished as early M1s proved buggy. Meanwhile, Intel churned out newer Pentiums that quickly topped Cyrix's chip.
Nvidia has hinted that if it built CPUs, it would focus on lower-end mobile processors such as those used in netbooks. Brookwood said such a move would let it sidestep some of Cyrix's problems -- but not enough of them.
License or acquire a CPU from an existing maker
Brookwood suggests that Nvidia buy Centaur, the semi-autonomous U.S. division of Via that designs its C7 mobile processor. The C7 was used in Hewlett-Packard's first-generation 2133 netbook. Enderle said that Via's C7 successor, the Nano, "is competitive with Intel's Atom though not as low-powered." Although the Nano hasn't attracted many customers, Brookwood attributes that to Via's lack of marketing muscle and its need to stay neutral as a major motherboard maker. Nvidia wouldn't be shackled by Via's constraints. "It would be the least bizarre of many alternatives," Brookwood said.
Enderle favors Nvidia approaching GlobalFoundries Inc., the now-independent chip fabrication division of AMD, to see if it is willing to manufacture for Nvidia CPUs based on AMD or Intel technology.
Negotiating the necessary cross-licenses, though tricky, isn't impossible, Enderle says, citing Microsoft's XBox processor, which is manufactured by IBM for Microsoft. It is based on Intel Pentium III technology.
What Nvidia has to watch out for as it considers deals with third-parties is being burned. Take Cyrix's deal with its manufacturer, IBM: It allowed IBM to build and sell the same processor under its own name. Cyrix hoped that would allow its processor to quickly gain volume, resulting in lower manufacturing costs and expanding the market.
IBM had other ideas. "Their sales guys just followed the Cyrix salesguys into the same customers and offered them a better deal," said Brookwood. "[Cyrix] opened itself up to cannibalization."
Don't over-promise and under-deliver
Although no surprise to industry insiders, Nvidia's statement of interest in making CPUs will inevitably raise speculation, and expectations. Nvidia's CEO, Jen-Hsun Huang, also has a history of making brash, and some would say rash, statements.
That was the trap into which Transmeta fell, says Brookwood. The company had a huge initial public offering (IPO) at the tail end of the dot-com boom. It then spent the several hundred million dollars raised on developing mobile CPUs, which failed to deliver on their promise of super-fast and low-power. Transmeta's chips are not widely used and it was acquired by Novafora late last year.
Don't think you know better than your customers
PowerPC was created in the early 1990s by an alliance between Apple, IBM and Motorola Inc. PowerPC chips were based on the then heavily-hyped Reduced Instruction Set Computers (RISC) architecture, which sought to raise performance through efficient chip design, rather than cramming more, smaller transistors into a chip, as exemplified by x86's Complex Instruction Set Computers (CISC) approach.
At the time, PowerPC was arguably faster than x86. Backers thought that would be enough to win software and hardware makers over to the platform, said Enderle. In the end, independent software vendors and original equipment manufacturers responded to customers, who were buying x86 PCs because of Windows and their larger pool of software.
The situation is reversed today. It's not the phone makers or even consumers calling the shots, says In-Stat's Lao. It's the wireless carriers.
Lao said that means Nvidia will need to make very low-powered chips if a carrier like AT&T believes that long battery life will sell more smartphones and win it more customers, or it may need to shrink chipsets down if Verizon Wireless decides that cute is in.
This story, "Ghosts of Cyrix, PowerPC, Transmeta haunt x86-bound Nvidia" was originally published by Computerworld.