SLI on X58: That Simple?

Intel’s next-generation processor has already been previewed to the point where there probably won’t be any surprises when it launches in October of this year (OK, maybe one; o ). Of course, the processor has been shown to be phenomenal. It’s Intel we’re talking about, after all. Intel, the company that just three short years ago was being so thoroughly trounced by AMD in the processor market that it was hard for anyone not on the inside to think they would ever be able to snap out of it. Intel, the company that for three years has not been able to do wrong. Yes, it’s the very same Intel we’re talking about. Nehalem is going to be amazing. Let’s put that to rest right now.

SLI on X58: That Simple?

However, one can make an argument that Intel’s success over the past few years has been at least in small part thanks to NVIDIA of all people, since their generally excellent nForce 600 and 700 series chipsets merged industry-leading graphics with industry-leading processors. Consumers looking to get the best of both worlds ended up going with NVIDIA chipsets so they could do 8-series SLI with their brand new Intel Core 2. Intel’s own 975X and 965 chipsets were successful for sure, but anyone will tell you that the lack of multi-(NVIDIA)GPU support was a big detriment to sales. X38 and P35 were a pretty significant improvement, but again, those chipsets owe at least part of their success to somewhat competitive ATI graphics cards (HD3800) that were making waves at the same time. And even now – even when NVIDIA is no longer in the top spot of the graphics world – people prefer nForce chipsets over their ATI-compatible counterparts from Intel. The reputation that NVIDIA chipsets have steadily built since nForce 2 is incredibly strong.

So it would stand to reason that the next-gen NVIDIA chipset, the one that supports Nehalem, will be equally successful. The only problem: there won’t be a next-gen NVIDIA chipset that supports Nehalem. Why? Well, it sort of depends on who you ask. At first, everyone thought Intel shut the door on NVIDIA and disallowed them to create a chipset for Bloomfield (first generation Nehalem). Now, it seems that the official story is that NVIDIA has no interest creating a chipset for Bloomfield. Good idea? Given everything we know about Nehalem so far, probably not.

We definitely don’t know when NVIDIA made this decision, but it’s safe to say AMD had not yet released the Radeon HD4800 series. Now that they have, the threat of a Nehalem/CrossFireX setup might be a little scary. Given that Bloomfield will be the only Nehalem option for the better part of a year, and that AMD does not have anything competitive in the processor segment on the horizon, NVIDIA probably thought it might be a good idea to get in on some Nehalem action. So, they let their nForce 200 chip out into the market so motherboard manufacturers could slap it on to X58, Intel’s enthusiast chipset for Bloomfield, and magically have SLI support. A simple solution. But is it really that easy?

X58

The first step in understanding how Nehalem/X58 works are to take everything you know about Intel motherboards and throw it out the window. The next step is to remember that Intel has eliminated the need for a memory controller hub (MCH) by taking a page out of AMD’s book and integrating the memory controller onto the Nehalem die. Consequently you would think that Intel’s usual three-chip chipset (CPU/MCH/ICH) would be cut down to two. But it hasn’t. Instead of the MCH communicating with the PCI-Express graphics lanes, the dual channel memory, and the ICH chip, Intel now has something called the IOH (I/O Hub). The IOH is connected to the CPU, just as the MCH was. However, now there is no such thing as a front side bus. Previously, the front side bus ran between the memory, MCH, and CPU. Now, there is no need for anything between the MCH and CPU, since they are on the same chunk of silicon. On X58, Intel has implemented something called a QuickPath Interconnect. QPI is an extremely quick, high-throughput means of communication between the IOH and the CPU. QPI is the same kind of concept has HyperTransport on AMD platforms, but… kinda better.

In any case, the result is that the IOH has a simplified job. All it has to do is communicate with the PCI-E graphics lanes and the ICH10/R. So there are still three chips on the new Intel chipset, their roles are just revised and things are sped up, significantly.

nForce 200

On current nForce 7 series chipsets, there are 4 chips (including the processor). There is the nForce SPP, nForce MCP, and nForce 200. The SPP basically does memory, nForce 200, and the MCP. The MCP does PCI, SATA, USB, audio, and other I/O features like IEEE1394. The nForce 200 does the PCI-E graphics. It is pretty complicated, but it has worked incredibly well. Now, it is key to note that you don’t need the nForce 200 chip to do SLI on an NVIDIA chipset. 680i didn’t have one. The advantage of nForce 200 is you get more PCI-E lanes for greater bandwidth (PCI-E 2.0) and, consequently, performance. So to put it simply, nForce 200 is sort of a great enabler for SLI.

The other side of the story is that nForce 200 is a pretty simple chip. It basically adds PCI-E lanes (?) and does something or other to allow two graphics cards to communicate. It doesn’t seem to be able to do anything else. By using the nForce 200 on a motherboard, you do nothing to negate the need for an additional chip to control/provide other functionality like I/O.

Smackover

In case you haven’t heard, Intel chose the totally cool and not cheesy at all name of “Smackover” for their next generation enthusiast chipset. Smackover joins Bonetrail (and Bonetrail 2!) and Skulltrail as the brain children of the Intel Extreme Desktop Board team. Pictures of Smackover have been on the web for a while now, and instead of showing you the actual final design of the board, we will just regurgitate the existing photos (which show the final configuration anyway).

Of course, the first thing you will notice about this board is the location of the memory slots. Below that obviously is the LGA 1366 (Socket B1), and to the right of that is the IOH. If you look a few inches below the IOH you will see the ICH10/R chip. The rest of the board is pretty self-explanatory. Of course on a motherboard, there is always more than meets the eye. Motherboards are not just a flat thing of plastic with a bunch of little doo-hickeys on them. If you look closely, you will see hundreds upon thousands of little tiny lines that seem to be connecting everything. These are called traces. Traces are tiny conductive paths on the board that really do connect everything. What’s more, is that they are engineered to precise lengths to insure information gets to certain spots at certain times. Oh, and one more thing: traces cannot cross over each other. Since components need to be connected and traces cannot cross over each other, layers have to be added to the motherboard to accommodate more complex circuitry. Most motherboards are 4 or 6 layers thick. The fewer the traces on the motherboard, the simpler it is to create and the less expensive it is to produce.  Intel’s X58SO is 8 layers thick. Of course there have been 8 layer PCBs before, but the fact that Intel’s primary CONSUMER desktop board is 8 layers is fairly alarming. When you actually look at what X58 represents and how everything is connected, you begin to realize that this chipset is one of the most complex ever created.

On Smackover, Intel has mounted the memory slots right next to and right above the processor socket. This allows them to rout the memory traces (remember, three memory channels) directly to the processor in as straight and short a line as possible. Similarly, the IOH is mounted to the right of the CPU and connected by the QPI, which again is routed as simply as possible, in a straight line. Intel has chosen a very efficient layout for Smackover which should provide for very low overall latency. It is easy to see that a “traditional” motherboard layout would result in some over-complicated connections between the IOH and ICH, which would consequently add latency to the board.

The thing to take away from all this is that there are a set number of components on the board and they all have to be connected with a large amount of traces and it is extremely difficult to make everything fit together on an ATX form factor motherboard.

X58 + nForce 200

If NVIDIA could make their own chip to replace the IOH, they could fairly easily create their own chipset mimicking the setup of X58. However, in order to do this, they would need to have a chip with a QuickPath link on it so it could communicate with the Nehalem processor. When news hit that X58 would support SLI, many believed that NVIDIA had received a QPI license from Intel in exchange for their nForce 200 chip. Intel would put the nForce 200 chip on their board to run SLI, and NVIDIA would have a slice of the Nehalem pie. Well, that wasn’t exactly the case. There was no QPI license given by Intel. NVIDIA simply released their nForce 200 chip into the marketplace. Now, motherboard manufacturers had the option to put nForce 200 on X58 boards to enable SLI support.

But why would anybody do this? Intel had no desire to plop a new chip onto their fine-tuned X58 setup when they already had a perfectly good multi-(AMD)GPU capable configuration. If NVIDIA had a QPI license, they could manufacture their own “IOH” and run the PCI-E Graphics lanes off that and integrate or circumvent nForce 200 altogether. But they don’t. Instead, motherboard manufacturers will have to add another chip to an already extremely complex X58 chipset. nForce 200 does not replace any existing chip on X58. As mentioned before, the only thing it does is control video cards. So how hard could it be?

Well, as it turns out, pretty damn hard. Despite the way this diagram makes it look easy, motherboard manufacturers will now have an extremely complex mess of circuitry to deal with by putting the nForce 200 chip somewhere in the already overcrowded southeast corner of the motherboard. There is only so much real estate on an ATX motherboard, and adding extra things necessitates the use of even more traces. Existing paths will probably need to be completely re-routed, which – and we hate to sound like a broken record – is not going to be easy since there is pretty much no extra space available in the first place. Remember, X58 is already built on an 8 layer PCB. Oh, and one more thing: using nForce 200 to control the PCI-E lanes means you have one more link between the graphics cards and the QPI. This will increase latency/overhead and more than likely result in crippled performance over what is expected from SLI on nForce 200.

So, not only will X58 motherboards with nForce 200 be available (and they will be available) for a $30 premium (approximate raw cost of nForce 200), they will likely be even more expensive for two reasons: We will probably see at least some manufacturers going with a 10 layer design (which increases production costs), and more feature-rich boards like this (read: Skulltrail) are always more expensive.

For the sake of article length we won’t get into the rumor about X58 being able to run SLI without the nForce 200 chip, and that the only thing holding it back is NVIDIA being stubborn and not putting one tiny line of code into their driver. That topic is purely rumor at this point as far as we know, so there is no point in discussing it further.

Conclusion

Yes you will be able to run SLI on the X58 chipset. But at what cost? Not only will you end up paying more for the ability to do this, but inevitable performance deficits will also be present. On the surface, this looks like a huge score for NVIDIA. But when you think about it, NVIDIA didn’t really score anything. What they did do is put one of their previously exclusive chips on the market. In exchange, they get the possibility that motherboard manufacturers will put the chip onto Intel’s (their competitor’s) chipset, the $30 that they charge for it, and any potential revenue that will be generated by graphics cards sales as a result of Nehalem.

It seems like the real winner here is Intel. Not only do they seem to have the most logical and well-thought out X58 design, but now their main competitor, AMD, does not have full reign on mutli-GPU configurations (and the money/notoriety gained from sales of multi-GPU configurations) for their new platform, and their other competitor, NVIDIA, can only do multi-GPU on their new platform by bumping up the price considerably. We simply cannot wait to see what happens come October when Nehalem launches.

Leave a Reply

Your email address will not be published. Required fields are marked *