Share This Article

4
inShare

When Nvidia launched the GeForce GTX 690 nearly a year ago, conventional wisdom predicted that AMD would respond with a dual-GPU version of the Radeon 7970. That didn’t happen. Even when Nvidia followed up with an unprecedented second $1000 GPU in February, the GTX Titan, there was no official word on whether or not AMD would answer with a $1000 card of its own. Today, that changes. AMD has decided that it wants to play in the $1000 segment this cycle after all, and its put its own official HD 7990 together to make that happen.

The HD 7990 consists of two 7970s sitting side-by-side on a single PCB. The difference between this card and the “unofficial” HD 7990s from companies like Asus and PowerColor? At a high level, not very much. Both the unofficial and official variants use a 48-lane PCIe 3.0 connection chip from PLX to handle cross-GPU communication, the PowerColor and AMD 7990 run at similar clock/memory speeds, and all three cards are a full Tahiti implementation (32 CUs, 2048 shader cores, 384-bit memory bus).

What makes the official HD 7990 interesting is the fact that AMD’s new card draws significantly less power than some of the alternatives. PowerColor’s solution, for example, is a triple-wide card that requires three 8-pin PCI-E connectors. The 7990 from AMD gets by with just two, and fits the chips into a dual-slot cooler. Core clocks and memory frequencies are right up there with what you’d expect from two cards running in Crossfire; there’s no penalty for dropping the chips side-by-side.

The other difference between the unofficial and official cards is the port allocation. The PowerColor card offered two DVI connectors, two mini-DisplayPort connectors, and an HDMI port. AMD’s version, on the other hand, has just a single DVI port and 4x mini-DP ports. That means if you want to use three monitors for gaming, you’ll have to shuck out an extra $25-$30 per adapter for a DVI-to-mini-DisplayPort dongle. You’ll need an activeadapter, not a passive one, and reports indicate it’s a good idea to source your choice off AMD’s list.

Never Settle goes for broke

Since last fall, AMD has been bundling some solid games in with its HD 7700-7900 video cards. For the HD 7990, Team Red has gone a little nuts. Buy one of these $1000 GPUs, and you get copies of:

• BioShock Infinite
• Crysis 3
• Deus Ex: Human Revolution
• Far Cry 3
• Far Cry 3: Blood Dragon
• Hitman: Absolution
• Sleeping Dogs
• Tomb Raider

At current prices, that’s just over $300 in video games. It’s the best gaming bundle we’ve ever seen a manufacturer field. Obviously the final value depends on whether or not you own or want to play these games, but these are all good titles. BioShock Infinite is a strong contender for 2013′s Game of the Year. Tomb Raider is well regarded, Deus Ex: Human Revolution is awesome (if a bit old), and the other titles are at least decent.

AMD is clearly doubling down on the Never Settle bundle and, by extension, on its commitment to gaming. That’s particularly interesting given some of the results we’re going to talk about today.

Over the past few months Nvidia has been steadily updating its desktop product lineup. The GTX Titan kicked off the year with a bang and new single-GPU performance records, theGTX 780 offered most of Titan’s performance for $350 less than Nvidia’s halo product, and the GTX 770 is a slightly faster GTX 680 for about the same price.

If the sudden deceleration from “record-breaking graphics card” to “Slightly faster than last year’s model” left you feeling whiplashed, we apologize, but that’s the general shape of things this time around. The GK110 (found in the GTX 780) boosted high-end performance dramatically, but Nvidia’s cards below the $650 price point aren’t moving much.

The GTX 760 is Nvidia’s new midrange card for the $249 price point, and it’s somewhat different from the GTX 660 Ti that it replaces. Specifics on the two are shown below:

There’s a little information that these charts don’t show you. The GTX 660 Ti has 112 TMUs (texture mapping units) but just 24 raster operations pipelines (ROPs). The new GTX 760 has fewer TMUs (96) but more ROPs (32) and thus a significantly higher pixel fill rate (33GP/s as compared to 23.5GP/s for the GTX 660 Ti). The GTX 660 Ti has 17% more cores than the 760, but the 760 has 33% more memory bandwidth and makes up for the core differential with a 7% faster clock speed.

Both the GTX 760 and GTX 660 Ti are based on the GK104 GPU. That chip is built with 1536 cores by default — Nvidia locks out damaged cores when it bins the product for board mounting. This makes it relatively easy for the company to fine-tune the GPU core/TMU/ROP ratio, and I suspect that’s what happened here. The original GTX 660 Ti is still a strong card, but the 760 Ti has had its features tuned to better balance performance and cost.

The price comes down a bit, to $249. Average performance should be quite close to the GTX 660 Ti, with the GTX 760 possibly taking an edge in higher resolutions due to increased memory bandwidth. Benchmarks I did at 1920×1080 with 8x MSAA — all detail levels set to maximum — show the Radeon 7950 consistently outperforming the GTX 760, though the gap is in line with the difference in price. Currently, the Radeon 7950 is selling for $299 — $50 more than the $249 Nvidia card.

The GPU market cools down — temporarily

Some GPU launches are more exciting than others. This, admittedly, is one of the less exciting ones. The GTX 760 doesn’t really change Nvidia’s overall lineup — it just solidifies the company’s $249 price point with a better-positioned product. There’s still a significant gap between the $249 GTX 760 and the $399 GTX 770 — AMD’s Radeon 7950 and 7970 are currently lodged in this area, and Nvidia has opted to preserve its ASPs rather than slash prices on the GTX 770 and attempt to drive Team Red out of those spots.

AMD has publicly stated that its next major GPU release, Volcanic Islands, will occur before the end of the year. Given that the company has also prominently advertised the HD 7000 family (and the associated HD 8000 rebrand) as “Stable throughout 2013,” we can assume that the introduction will happen in the November/December time frame. Nvidia, meanwhile, has promised that the Maxwell family arrives “in 2014″ but hasn’t been specific as to when.

The timing of both products will depend on which process nodes they utilize — and where they’re built. While there’s little doubt that Nvidia will continue to use TSMC, AMD is contractually obligated to switch to GlobalFoundries. The terms of AMD’s amended Wafer Supply Agreement (WSA) with GF are quite specific and states: “The second amendment grants the Company certain rights to contract with another wafer foundry supplier with respect to specified 28nm products for a specified period.”

But therein lies the rub — while GF is definitely in risk production and has shown 20nm wafers (and a few 14nm wafers that include its hybrid technology), it hasn’t said the magic words: “volume production.” Will it be ready by the end of the year? Maybe. Or maybe Volcanic Islands is a second-generation 28nm refresh. It wouldn’t be the first time AMD has delivered significant GPU improvements without shifting to a new node, and a second-generation 28nm chip produced at TSMC might give them an early window compared to a 20nm part from Nvidia later in 2014.