When it comes to building solutions for computer platforms and providing SLI support, Nvidia is in a rough spot right now. Where it once had a very profitable chipset business that spanned to both Intel and AMD platforms, it now has only Ion and rebranded nForce chipsets for AMD. Business for the new LGA 1156 is far from sight and as for LGA 1366 chipsets, we will probably never see them.
There are reasons for this, other than chipset licences Intel might or not want to concede.
First, there's the LGA 1156 interconnect to the PCH, the new name that Intel gave to the chipset, which stands for Platform Controller Hub and is connected to the CPU via DMI, which is essentially how all Intel ICH southbridges had connected to the northbridge until the LGA 1366 came along and brought a single chipset and QPI. The new P55 chipset has two connections to the CPU: DMI and FDI:
FDI stands for Flexible Display Interface, which is needed for "Clarkdale" based processors that feature an integrated GPU, called iGFX.
The DMI bus is something rather narrow, which apparently delivers only 2GB/s, or 1GB/s in each way. Some diagrams have 4 lanes drawn with the DMI word on it, which translates to 2GB/s as DMI is reportedly PCIe v1.1 derived. These four lanes, at 500MB/s (upstream+downstream) that the PCIe v1.1 calls four, delivers the aggregated 2GB/s.
The old FSB architecture is better well known: the northbridge connected through a wide enough FSB to the processor and had the memory controller attached to it. Typical FSBs for Core 2 range from 1066MT/s to 1333MT/s, on a 64bit wide bus. This translates to a one way bandwidth of 8.5GB/s or 10.6GB/s in case of the 1333MT/s(or 1333MHz) of the FSB. But the memory controller has moved to the CPU and the FSB just doesn't cut it for multi processor platforms, hence QPI and HyperTransport.
The old architecture also had one plus going for it, when the chipset had an integrated graphics core, it could have access not to just 10.6GB/s but to 21.3GB/s, when paired with two DIMMs of DDR3-1333. That was plenty for Nvidia to strap a 9400 core on a chipset and call it the GeForce 9400M. It is slower than the 9400GT but not by much, which only has access to 12.8GB/s in reference designs. The chipset has to share bandwidth with the CPU, so performance ends up similar.
Now, if you're Nvidia, where do you put your integrated graphics core on the LGA 1156 platform? You can't.
If you put the graphics core on the PCH, you only have 1GB/s access to main memory, shared with six 3Gbps SATA ports, 14 USB 2.0 ports and a PCIe x4 interface. I don't remember the last time graphics processors had so little bandwidth to their local framebuffer. Don't believe me? Dial your AMD's HT bus down to 200MHz(which would deliver 800MB/s) and see the slideshow for yourself.
AMD also has to go over the HT bus, but since it's way wider right now - 8GB/s in current processors - there's still some bandwidth there.
So, If I were Nvidia, I would have to go discrete or put a framebuffer on the motherboard - something like AMD's sideport. The difference here is that the framebuffer would have to be a complete one, since there's no bandwidth to use TurboCache, which allows you to allocate more, slower memory than you have on the local framebuffer - everything would be a slideshow again. 512MB of dedicated memory on a motherboard? Like that will work when you're trying to compete in price/performance.
There's still the very slim change that Nvidia can put TurboCache to some use, since it works over the PCIe bus, and get the integrated core feeded while sacrificing all 16 PCIe lanes from the chipset or just 8. Eight lanes gives you 4GB/s in each way, which will hardly cut it, and the full 16 that LGA 1156 have would bring that to a more manageable 8GB/s in each way. That's still too little nowadays, but would allow for roughly GeForce 9300 class performance.
By switching to a DMI interconnect for the new LGA 1156 platform, Intel has locked Nvidia out of their platforms by engineering, not by court. Intel needed QPI on LGA 1156 if it hadn't moved the PCI Express 2.0 lanes inside the CPU, this way it will be fine with just DMI. Nvidia can't strap an efficient GPU on a PCH as Intel will have a lot more bandwidth available and a "good enough" GPU core already embedded in the CPU package with "Clarkdale". This is also the main reason why Intel has pushed "Clarkdale" cores to an earlier release, as they can't also strap a GPU to the PCH and sell mobile "Lynnfield" cores.
The FDI bus is of no use, it is just there to pass the final render of the GPU: around 500MiB/s suffices for 1920x1200@60Hz, which is around the same bandwith what a single link DVI also delivers (1920*1200*24*60/8/1024² = 395MiB/s).
The LGA 1366 is an even worse situation for Nvidia. A chipset could be built with a QPI license, it could have an integrated graphics core also, but to what end? There's an each way bandwidth of 12.8GB/s to tap on a 6.4GT/s capable Core i7 but the platform is so exclusive that it wouldn't pay for the development cost of a discrete chipset, let alone for an integrated one. The only markets would be the high-end desktop and the even smaller workstation, HPC and server markets.
When Nvidia was on the AMD front, things were considerably different. They could build a desktop chipset, not change it even a bit and sell it for servers and high-end workstations for a bit more of profit. With "Nehalem" that just isn't possible.
Nvidia is, reportedly, working on chipsets for the LGA 1156 and "Nehalem" mobile platforms. Unless they're targeting the discrete chipset market, which would be even stranger now that they're selling P55 chipset SLI licenses to anyone, a new chipset without the serious caveats described above is a very, very long shot.
No comments:
Post a Comment