Fermi Goes Mobile: AVADirect's Clevo W880CU with GTX 480M

GTX 480M: Fast but Mixed Feelings

Back when we took the ATI Mobility Radeon HD 5870 and NVIDIA GeForce GTX 285M and pit them against each other, the 5870's victory was met with some disappointment because it just wasn't the Hail Mary we had hoped for. Notebook graphics performance had been stagnating for so long with no competition at the top of the heap, allowing NVIDIA to refresh the G92 an absurd number of times, and yet when ATI finally decided to come out and play, the best they could do was beat the GTX 285M by about 10% on average. ATI didn't deliver a knockout blow; they just flicked NVIDIA behind the ear over and over again. Now with a cut down Fermi chip powering the GeForce GTX 480M, NVIDIA's response is to say "quit it!" and slap at ATI's hands.

The most impressive thing about the 480M isn't its performance; it's the fact that NVIDIA was able to get the sucker into a notebook to begin with. Sure, it's a nine pound notebook cooling a TDP of 100 watts, but credit where credit is due: Fermi isn't exactly well known for being economical with power on the desktop. Really, the GTX 480M raises more questions than it answers.

I suspect most of us agreed when the GeForce GTX 260 and GTX 280 came out that there was no way NVIDIA would ever fit those chips in a notebook, and in some sense we were proven right with refresh after refresh of G92 at the top of the mobile graphics food chain. With mobile Fermi, it looks like NVIDIA more likely chose to remain with a tweaked G92 in order to focus resources elsewhere—i.e. dropping to 55nm to save power and boost clocks over the 65nm original. Obviously, we wouldn't have wanted a trimmed down GT200 chip this late in the game, but cutting down the GF100 to fit into a notebook had to have been far more onerous a task than trying to get a 55nm GT200b die into the same power envelope (or trying to respin GT200b at 40nm). Unfortunately, GT200b doesn't have DX11, so really NVIDIA had no choice. The result is a GF100 die that sips power at idle (relatively speaking) but still guzzles the juice under load. (Not that you'd run a gaming laptop on battery power.)

As for ATI/AMD, they seemed unable to deliver Mobility Radeon HD 4800s in any kind of reasonable quantities, and in general there was a lack of interest. Contrast that against being able to buy an HD 5800 series laptop from a variety of vendors today. They're not the fastest mobile parts any longer, but they are far more affordable. $1500 for the ASUS G73Jh makes the Clevo W880CU look like highway robbery! Go one step further and start asking ATI the same questions. Cypress is a monster to be sure, but it's no more a beast in terms of power and heat than its predecessors, the RV770 and RV790, were. RV770 made it into notebooks, but the best ATI says they can do is trim the clocks on Juniper and call it a day. We're left with a Mobility Radeon HD 5870 that offers a minimal improvement on its predecessor and wondering why a mobile chip based on the superbly economical Radeon HD 5850 isn't making the rounds. If NVIDIA can do a 100W TDP mobile part, AMD should be able to do the same. Certainly trimming Cypress too much has proven in some ways as troublesome as cutting down Fermi has been; the 5830 sports higher thermals and power draw than the 5850, and the GTX 465 landed on the market with a resounding thud, but desktop parts aren't the same battleground as notebooks and 5830 or GTX 465 levels of performance in a notebook would be substantially faster than what we currently have.

Really, NVIDIA got to sit on the top of the mobile GPU heap for far too long. It's good to see competition, and we can only hope that there's more to come from both companies. We're still a generation behind in terms of desktop performance; even if both companies are now using up-to-date parts, the final clock speeds are a far cry from desktop GPUs. What we really want is more of a Conroe style revolution for mobile GPUs where we get up to 25%-50% more performance without increasing power requirements—or even reducing them!—over last generation hardware. Then again, the P4 architecture was so poor that it made Conroe possible.

It's hard to believe there aren't better options for either manufacturer. Was NVIDIA so upset about losing the mobile crown to ATI—even though the margin wasn't that great to begin with—that it was worth curtailing Fermi's performance so brutally? Wouldn't the prudent thing to do have been to let ATI have their cake for the time being and try and push GF104 into laptops? Or would that just be suggesting NVIDIA do the same thing we're accusing ATI of? Like we said, the GeForce GTX 480M raises more questions than it answers, but all of us armchair engineers have to be wondering why mobile graphics aren't improving faster.

Looking at the big picture, the limiting factor on mobile GPUs is power. Desktop cards keep getting faster, sure, but power requirements are generally increasing as well. ATI's 4870 has higher load power than 3870, and 5870 leapfrogs 4870. Likewise, NVIDIA's GTX 285 needed more than the 9800 GTX, and the GTX 480 ups the ante. Move over to notebooks, and we hit the power wall hard. The biggest power bricks are still 240W (give or take), so there's no going over that limit, even if you can dissipate the heat. We've had the same 220-240W power adapters at the high-end for years, and it doesn't look to be changing. 480M may have bumped the TDP up to 100W, but our battery life tests show that it's about the same as the 50W 5870 when it's not under load, and we've had dual-GPU notebooks that use a lot more power than a single 480M. It's not like you're going to load the GPU without plugging in, and at that point it's more a question of whether cooling is sufficient than how much power you need.

Perhaps a simpler way of stating things is that mobile graphics performance isn't increasing very quickly. AMD likely took the existing GTX 285M and did enough testing and research to make sure 5870 was faster by 10%. Now NVIDIA has gone and done the same thing to regain the lead. They pushed the power envelope harder, but that's more a factor of the Fermi design constraints right now. Give them time for revisions and we'll likely see that drop. Ultimately, process technology refinements and tweaked architectures are the primary means of performance improvements, and 25% faster per year looks to be the goal.