Ultra-mode video card shopping guide

While this is a seemingly popular stance, it's frequently ineffective in terms of price / performance over time. The cost of many components is priced non-linearly; at the low end, a few dollars buy a lot of improvement, whereas at the high end, a lot of dollars buys only a little improvement. Additionally, Moore's Law and related changes has a strong effect on how much power you get from your money over time.

If you have an approximately fixed budget for computing over a span of several years, particularly 4-5 years as is common for students, blowing it all on a top of the line system at the start will on average give you far less performance by the end, than saving some for upgrades or even replacements.

E.g. if you have $3000 to spend over 4 years, buying a $3000 computer now is a bad deal. You're probably better off spending half to two-thirds of it now, and saving the rest for a major upgrade or partial replacement (moving some parts over) in about 2 years.

Also, in terms of Going Rogue, not only will we have a better idea of actual performance closer to launch, it's far enough away that video cards will hopefully have some major performance increases between now and then. The time leading up to Christmas is also somewhat problematic, as some of the better choices for graphics cards right now are in short supply, and cost more than they "should" because of that.

I have my doubts that Larrabee was ever going to apply downward pressure on the high end. I think Larrabee was designed at a time when Intel did not think AMD and NVIDIA would get to the stream/vector processing levels that they did. Although they are blaming general problems, I think Intel realized Larrabee wasn't even going to be competitive with the current gen of AMD and NVIDIA silicon, and decided to go back to the drawing board. If their architecture doesn't leapfrog AMD and NVIDIA completely, they are toast (because their novel architecture is almost certainly going to require significant time to software-optimize, so they need to be way faster in raw performance just to compete with the current GPUs in the short term).

If the cards are still hard to get by xmas, I'm probably going to just scoop up a 5850 (or possibly a 5870) at whatever price they are at, and call it a day.

Robert Cringely thinks Intel is going to buy nVidia.

I somewhat doubt this. Not will all of the bad blood between Intel and Nvidia.

In addition, right now Nvidia pretty much offers Intel nothing of value. The Nvidia memory controllers aren't as fast or as efficient as Intel's. Tegra's long-term mobile value is still iffy. Due to the recent settlement with AMD, Intel has access to all of AMD's patents and Intellectual property, so it's not like Intel's hurting for graphics technology. Yeah, Larrabee didn't make the performance cut Intel was aiming for, but since Intel is already sitting in the number volume seller of GPU's with it's integrated set-up, they shouldn't be in a hurry to enter the discreet GPU market.

The only value Nvidia could bring Intel is a mass contract deal. Let's say that Nvidia has actually captured a contract for the next Nintendo Handheld, or has captured the contract for the next Apple Ipod / Iphone system. Unlikely, but that's the only thing that would make Nvidia an attractive acquisition target for Intel. There is also the aspect that if Nvidia had captured such a contract, they wouldn't be in a hurry to sell themselves off to the highest bidder. They'd remain independent. So even if Nvidia did have a contract on the way, it'd still would not help an acquisition.

Cringley tends to think lots of things. Most of the time, his predictions tend to be right only when he predicts the continuation of a trend. When he goes out on a limb, he's worse than random chance.

I think anything is possible in the chip industry right now, but Intel acquiring NVIDIA seems highly unlikely to me. It all hinges on his theory that the ATI acquisition by AMD gives AMD a strategic advantage over Intel that Intel has to respond to. Larrabee suggests that Intel wasn't thinking that.

Intel doesn't care how many graphics cards AMD and NVIDIA sell, anymore than they care how many hard drives Seagate sells. What Intel wants control of is the motherboard, and everything on it. The motherboard is the defensive perimeter around the CPU. Intel *does* care about embedded graphics, but embedded graphics did not appear to be directly on Intel's mind when they designed Larrabee. A multicore superscaler x86 architecture is not the sort of thing you're going to then add to a motherboard or a SoC as built-in graphics most of the time.

Instead, Intel was most likely thinking two things. First, the real threat was NVIDIA (and to a lesser extent AMD) pushing the graphics cores into the computing space, ala HPC via GPU. If people start deciding that encoding video or other computationally high tasks can just run on their graphics card, they may decide they don't need the next big Intel CPU. That's bad. They probably felt they needed to have a foothold not in the GPU-space, but in the *ultra-highend* GPU space to make certain they couldn't be outflanked in the computational high ground (for those that don't follow graphics card technology, look up the specs on the top of the line NVIDIA and AMD GPUs sometime).

They were probably also, in a related line of thought, wondering if super-core x86 could eventually scale up to compete with the superscaler GPUs. Since Intel's current strategy is to keep increasing cores until people can't make use of them any more, *one* way to make sure people are buying those Core i-eleventybillions is if they push the software industry to start writing code designed for multiple x86 cores. One way to do that is to release a GPU designed around multicore x86. If graphics driver writers could be convinced to start writing code for Larrabee, those same skills would be useful for eventually writing for Core iX. Eventually, Core iX would implicitly *have* a built-in GPU without even needing to integrate a special graphics core (they'd basically just need to extend the cores with Larrabee vector extensions). You'd just have an Intel CPU with twenty cores, and four would run your graphics. This is Intel's best-case wet-dream scenario.


There are a number of reasons why Intel might pull Larrabee. If you're Cringley, Intel is lying. But if they are telling the truth, then Intel has said two contradictory things about Larrabee. Prior to the announcement, Intel was all roses about Larrabee's performance. In fact it achieved a teraflop on a real benchmark not that long ago. But they are now saying there are performance issues with it. Two big possibilities loom for me:

1. In spite of prior information to the contrary, on real workloads the system was unable to achieve its benchmarked performance levels, or unable to achieve *stable* performance levels. My guess: internal memory bandwidth or architecture issues. This has bit the big two in the past as well.

2. When actual silicon was tested, it turns out the design has fabrication issues that would make it difficult to ship in volume. It may even be analogous to the problems AMD and NVIDIA are having now with their 40 nm process.

Its also possible that the skills required to make an actual video driver for the thing that will work under all conditions at high performance simply don't exist yet. In that case, Larrabee might be a "software development platform" (what Intel is now saying) specifically to build the skills internally at Intel just to program the thing. A GPU without video or HPC drivers is just a hunk of sand.

Unless dwindling supplies of a now out-of-production product and the possibility of ongoing SLI demand serve to keep their prices right where they are.

Well yea, even if the price stays the same I wouldn't mind too much. The 9800GTX+ I got about a year and half ago was $200, now it's like $130, can hope the price goes down heh. I just bought a new 28" monitor so I'm waiting till I have a little more money before I drop another $300 on a video card.

Anyways does this have DX11 support? I don't see anywhere saying if does or not.

Wait, what? He's blogging again?? That rat! he said he was quitting two years ago.

Will have to keep my eye out and wait and see the price vs. performance.

Do keep in mind, you won't be comparing Fermi to the RadeonHD 5870 set of launch cards.

If Nvidia is able to make the February time table, they'll be launching shortly before ATi's spring product refresh.

If Nvidia makes a March time table, they'll be launching into the tail end of the current RadeonHD 5x00 lineup.

If Nvidia makes an April or May launch, they'll be alongside the second generation of Evergreen series cards.

Given past ATi product refreshes, we can probably accurately predict pricing and performance of the refreshes, just looking at the changes from the Radeon 9700 series to the 9800 series, the x800 series to the x850, the x1800 to the x1900, and so on.

This, of course, depends on AMD deciding to stay on it's product refresh plans. If things go wrong for Nvidia, or the gaming performance of Fermi just isn't that good, AMD may be able to hold off on a product refresh, save itself some manufacturing cash, and keep pushing the current HD 5x00 series.

No. The only DirectX 11-compliant graphics cards right now are the ATI 5xxx series.

Most computers you buy "off the shelf" (i.e., that aren't custom-made to your specs) don't have power supplies like that. If you're custom-ordering one, you should be able to specify what parts you want in it, including the power supply. If you're buying one already put together you'd just need to purchase a better power supply and install it before installing a high-end video card.

Yeah, am ordering from the HP but they don't allow the power supply to be specified...so will probably have to buy the power supply like you said afterwards. Would any power supply fit?

HP has a history of making proprietary case/PSU combos, but it's been a while since I've used an HP tower. I'd call up their customer service line to ask about the PSU before ordering it, just in case.

Originally Posted by Nethergoat it's NEVER too late to pad your /ignore list!

We both know and don't know.

Since the HeroCon Going Rogue Demonstration was run on Crossfired RadeonHD 4870x2's, we can presume that yes, Multi-GPU setups will accelerate the game.

However, nobody from the developers side has actually come out and said that Multi-GPU setups will give a performance boost.

No, they said that Windows 7 isn't required for Ultra Mode. Specifically, they meant the version of DirectX that Win 7 uses.

City of Heroes runs off of a graphics package called OpenGL instead, so no particular DirectX version is required.

Originally Posted by Nethergoat it's NEVER too late to pad your /ignore list!

Just got off the phone with them again, its a bit frustrating; can't seem to get a straight answer beyond the fact that they want me to buy the PC first and then they will send me the specs for new graphic cards and power supplies. From the cards specified above, it seems that we still have no data on the 5770 am not sure what to do. I know their sales are going away at the 1st of the year... their customer service is polite but totally unknowledgeable