AMD R680 - Dual GPU graphics card
R680 and its glimpses of fame
AMD's R680 has been about as low key as a Britney Spears shopping spree at Palace of Wigs and AMD seems to be just fine with that. After all, they were the ones running the displays and demos. The R680 card was seen at the original AMD Spider demo day in Tahoe, the first AMD RV670 / HD 3870 editor's day and as recently as this past CES show in Vegas, even inside Intel's Skulltrail platform demo.
AMD also didn't seem to mind a little bit of information leaking out; we knew from the get go that R680, a codename for a card design not really a new graphics chip, was essentially made up of two RV670 GPUs tied together in a CrossFire configuration on a single board. We also knew that AMD intended the R680 to compete for the performance crown against the likes of NVIDIA 8800 GTX and Ultra models.
Now we have had time to analyze the technology and actually test it; we have the answers to the questions that have been nagging you.
AMD's R680 Design
The design of the new AMD Radeon HD 3870 X2 graphics card, previously called the R680, is really quite simple at its core: take two RV670 GPUs, put them on the same PCB with their own 512MB frame buffers, connect them with a PCI Express bridge chip and enable CrossFire for a better performing single card solution. In reality of course we all have much bigger questions than that and AMD has provided most of the information we were looking for. R680 is indeed a good display of engineering that ties together nearly all aspects of its design into a neat package.
The target segment for the Radeon HD 3870 X2 is the highest AMD has targeted in some time: performance leadership. That's right, AMD is planning on taking the 8800 Ultra and GTX for a ride all while pricing the R680 at $449 - less than current 8800 Ultra pricing and about the same as most 8800 GTX cards.
Reference R680 Radeon HD 3870 X2
The first you'll notice about the card is that it is big: about as long as an 8800 Ultra but much heavier.
Rarely do we see board design diagrams with this kind of detail from either AMD or NVIDIA but AMD wanted to show off what they consider to be one of the most refined board designs they've ever done. You can see three chips in this diagram above: the two RV670 GPUs are the higher pin-count chips while the similarly sized but less complicated chip between them is the PCIe bridge chip. Each core has four memory chips attached to it, via the identical 256-bit memory bus that RV670 used before, totaling 1GB of on-board frame buffer for the Radeon HD 3870 X2. This diagram is a bit dated though, as you'll notice the power connectors shown in the upper right hand corner are moved around slightly on the actual product.
The bridge chip is PCI Express 1.0 technology that takes a single x16 connection in and splits it between the two RV670 cores on the card. Each card gets x16 PCIe 1.0 bandwidth to the bridge allowing for fast GPU-to-GPU data transmissions. This is in fact the exact same PCIe bridge that was used in AMD's dual-GPU HD 2600 X2 card. Because this is a PCIe 1.0 bridge, the Radeon HD 3870 X2 is actually NOT a PCIe 2.0 graphics card as the HD 3850 and 3870 cards are. This is will certainly be one of NVIDIA's contention points going forward as they release faster high end cards but for now even NVIDIA's flagship 8800 GTX and Ultra cards are PCIe 1.0 - and in reality the bandwidth difference isn't a big performance factor in these GPUs quite yet.
As I've mentioned a few times, the R680 is really taking two RV670 GPUs are combining them on a single card and this diagram shows that to be exactly the case. Obviously still built on the 55nm process technology that allowed the HD 3870 and 3850 to perform at such a high level while reducing power consumption over the HD 2000 series, the HD 3870 X2 should be just as power efficient when compared to a pair of 3870s. The specs on the table above are clearly just taking the 3870s specs and doubling them: 640 stream processors, 32 texture units, 32 ROPs. These are nice to see on paper but in reality there is a big difference between having these on one chip and two as we'll see throughout our testing.
One interesting spec that has changed is the clock speeds. The core frequency has actually gone UP from the stand alone HD 3870 card: from 775 MHz up to 825 MHz while the memory clock has decreased from 2.25 GHz to 1.8 GHz. The increase in the core clock is impressive since we are actually combining two RV670s on a single board and under a single heatsink though I am guessing the increase heat of the core clock is offset slightly by the slower memory clock.
The pure processing numbers that AMD provides here, basically doubling the numbers from the RV670, are astounding: getting 1 Teraflop of power on a single card! But as we noted above, getting the theoretical peak out of two GPUs is even harder than getting it out of a single one.
Taking a look at the rest of the features on the new HD 3870 X2 we see that in fact the original 3870 uses GDDR4 while the new card runs a slower clocked GDDR3 module. According to AMD at least the board power is less than a pair of standard HD 3870s though it DOES require two 6-pin PCIe power connectors despite previous reports to the contrary.
Obviously one of AMD's key selling points for this GPU is what it holds for the future: later in Q1 AMD will be releasing drivers that allow for pairing two HD 3870 X2 cards for quad-GPU rendering known as CrossFireX. By doing so you'll essentially have 1280 stream processors and 2GB of memory at the games' disposal with 2 Teraflops of computing power. Obviously the hard part here is getting it all to play nicely with the games that take advantage of it - we'll have more on that next month I am told.