jasoncross.org

Today was the start of Nvidia’s GPU Technology Conference. It’s really still just the NVISION conference, because it’s not much of a “industry-wide” conference if ATI and Intel aren’t there. The biggest announcement of the show is undoubtedly the unveiling of Nvidia’s next-generation GPU, code-named Fermi. I’m not sure why they named the chip after Enrico Fermi, who is best known for his work with radioactive substances and controlled nuclear reactions and stuff. But as code-names go, physicists are cool, so I’ll let it slide.

I won’t bother to summarize all the individual features that were revealed today. Tech Report has a excellent article on it, so does AnandTech. I’m just going to editorialize a bit with some of my thoughts based on what we know (and don’t know) so far.

First, boards based on Fermi are going to cost a considerable bit more than the Radeon HD 5870 and 5850, which are ATI’s competing DX11 cards that just launched. The RV870 GPU powering ATI’s cards is 334 mm². It has a 256-bit memory interface. Nvidia didn’t talk about GPU size, but it did say that Fermi is 3.0 billion transistors – 40% bigger than RV870’s 2.15 billion. So, figure a chip somewhere around the 460-480 mm² mark. That’s huge.

The chip being 40% bigger doesn’t mean 40% more expensive to produce, though. Imagine chips A and B. Both are 40nm chips made at TSMC. Chip A can fit 100 chips on a wafer, and Chip B can fit 60 chips on a wafer, because it’s 40% bigger. But as chip size grows, it’s harder for the whole chip to come out without flaws, so the yields are worse. Chip A has a yield of 75% – three-fourths of all the chips on the wafer function properly within the intended specs. Chip B has a yield of 60%, because it’s so much larger. That means you’ll get 75 good chips on a wafer for Chip A, but 36 good chips for Chip B. That’s less than half.

In other words, depending on how the yield situation works out, Fermi could be twice as expensive to produce as RV870. Hell, it could be worse. We really have no way of knowing, except to say that a 40% larger chip is usually well more than 40% more expensive to make.

It’s not just the chip, either. A 384-bit memory interface means Fermi-based cards will likely have either 768 MB (not likely) or 1.5 GB of RAM, so that’s higher RAM costs. It also means more PCB layers on the board itself. So aside from higher chip costs, the board costs of Fermi-based products will be higher than Radeon 5800 products.

So if Fermi-based products are going to be considerably more expensive than Radeon 5800 products, what about performance? Well, all Nvidia has talked about so far are the chip design elements that impact GPU compute, rather than traditional graphics. There’s quite a lot there. Nvidia has clearly spent a fair chunk of the transistor budget doing things like dramatically improving double-precision floating point performance, increasing cache sizes, ECC memory support, and so on. These things typically do nothing at all for typical graphics performance (games and stuff). So the chip is 40% more transistors, but that won’t necessarily translate into 40% higher frame rates.

Nvidia seems to be gearing the world up for this. The mantra they keep chanting is that “graphics performance isn’t enough anymore.” Compute really matters a whole heckuva lot, they tell us. This sounds like PR code for “the card is going to be 50% more expensive than the competition and not 50% faster in games, so please place as much importance on GPU compute apps as possible so we look like a better value.”

You know how drill sergeants tell recruits to begin and end everything they say with “sir?” Sir, yes sir! Sir, I didn’t mean to shoot the sergeants toe off, sir! That’s what it’s like listening to Nvidia these days, only with “CUDA” instead of “Sir.” For over a year, Nvidia has told everyone who will listen that GPU compute is super duper important, and has very aggressively flogged PhysX and CUDA. And you know what? Consumers just don’t care all that much. Maybe one day, when there are robust standards and quite a few GPU-accelerated applications that normal people use all the time, the average consumer will want a graphics card to make its non-gaming apps go faster just as much as it wants it to make its games go faster and look better. But we’re not there yet, and we’re not going to be there in the next six months, as much as Nvidia would like us to be.

So Nvidia’s facing a tough sell in Q1 2010 (or maybe late 2009) when the first Fermi-based cards go on sale. They’ll almost certainly cost $399 or more, judging by what we know so far. ATI has a chip and board design that will let them push Radeon HD 5850 cards below $200 and 5870 cards below $250 within the next six months, if they want to. Is a modest increase in frame rate and much higher performance in GPU compute apps going to be worth such a broad difference in price? Will it be a moot point, because the cards will be out of the cost and power budget for most consumers (and OEMs)?