'Graphics PCIe' myth

Main -> Hardware lies -> "Graphics PCIe" myth

"Graphics PCIe" myth

It seems that nearly everyone now believes that:
a) Graphics cards are so fast that they must be inserted only in PCIe x 16 v. 2.0 slot to work at their best.
b) The PCIe x 16 v. 2.0 slot nearest for the processor is for graphics cards use only and the graphics card inserted in other slot will loose in productivity.

This two myths are the aftermath of one of the most blatant lies in computer industry invented by nVIDIA - the myth about "Graphics PCIe v. 2.0". Well, should say it was quite eagerly supported and used by AMD (and even somewhat backed by Intel - by not disclosing the truth).

Let us see some technical data first:

PCIe v. 1.1 has the standard bandwidth of 2.5Gb/s per each lane. As the standard demands 2-bit parity check for each byte, this means 2Gb/s (or 250MB/s) of data per line. So, 16 lines in PCIe x16 slot can transfer 32Gb/s or 4GB/s.

AMD Hyper Transport v.1.0 bus specifies 3.2GB/s (which means 2.5GB/s of data - only about PCIe x10 v.1.1 link), later AMT HT specifications provide much faster throughput
Intel chipsets support at least no less than full PCIe x16 speed from chipset to processor or to memory

But the tests with ioDrives (and similar Micron devices) show that even the latest nVIDIA chipsets can not support more than 1.2GB/s PCIe-to-memory transfer (which is the theoretical bandwidth of only a PCIe x4 v.1.1 link!)

This means that NONE of the existing graphics cards needs more that true PCIe x4 v. 1.1 bandwidth! Where is the catch then?

Some history:
The predecessor of "Graphics PCIe" - the AGP bus was developed by Intel to allow the use of memory-less videocarts. The high cost of on-card memory prevented further graphics development, so the bus allowed fast dedicated card-to-memory connect, allowing the use of cheaper main memory instead of on-board card memory. This approach really needed wide bandwidth, so AGP 4x (8.533 Gb/s - 1.067 GB/s) became the-facto standard, with up to AGP 8x 64 -bit (34.133 Gb/s - 4.266 GB/s)) solutions in use.

For some time AGP bus co-existed in parallel with PCIe bus, used for other high-speed cards. But development of graphics processors made the main idea behind the AGP bus - common memory pool - an obstacle. AGP bus could not supply the fast GPU with data, so the cards first got on-board memory for textures and, with cheaper memory at hand, all the GPU memory pool moved back to the card.

This made the ultra-fast GPU-to-memory connection unnecessary and allowed the GPU cards vendors to adopt (less than 1% slower than AGP 8x at the time) PCIe x16 slot for graphics cards.

And here we come to the beginning of "Graphics PCIe" myth.

nVIDIA newer could produce fast chipsets (even now their chipsets do not fully support AMD HyperTransport bus speed or Intel FSB bus speed). So, to be on the safe side they implemented their early "PCIe" bus. Really the "Graphics PCIe" bus is a PCIe w/o parity checking - video may work nicely without it.. This allowed a bit more bandwidth for the much lower than the standard speed of nVIDIA "PCIe bus". Of cause only nVIDIA cards could work on the provided slot, any other true PCIe card would not work in it. But nVIDIA cards also could not work in standard PCIe slots - to hide this limitation nVIDIA invented nVIDIA-ready certification (later it transformed into SLI limitations, making SLI drivers working only on nVIDIA chipsets, despite the fact that SLI is pure software technology and will work nicely on any chipset).

This did not last for long - quite soon AMD began producing cards that will work nicely in this "Graphics PCIe" slot, Intel also supported this lie with their new chipsets supporting the feature.

nVIDIA soon understood the danger of this approach, and as PCIe 2.0 standard was already at hand, they built new chipset with support of both standard and (renamed "Graphics PCIe") "Graphics PCIe 2.0" - of cause it newer supported even PCIe v.1.1 full speed.

But nVIDIA "Graphics PCIe" forced most other hardware vendors to write in their cards manuals not to insert their cards into the first PCIe slot (though only nVIDIA chipsets had this non-standard "PCIe" connection)

If graphics cards do not need PCIe bus speed, why they use additional SLI / CrossFire bridges?

Well, this is very simple - as nVIDIA PCIe link is quite slow, it turned out to be more fast solution to organize direct GPU-to-GPU link. In its latest GPUs AMD also has taken this approach - this allowed AMD to allow its cards to work nicely in PCIe x8 slots with only PCIe x4 link between cards (well, this link uses only a fraction of even PCIe x4 v.1.1 bandwidth). By the way - this really mean that new AMD cards need only PCIe x8 v.1.1 connection to work, though nVIDIA cards still need x16 connection for full speed work.

This also explains why nVIDIA will have to quit the chipset market - latest AMD chipsets also support at least full HT 1.0 bandwidth (3.2GB/s), which nVIDIA chipsets can not support. nVIDIA is just not able to implement this speeds in silicon. As soon as Intel will make their Larrabee to work at least no worse than nVIDIA/AMD cards it will no longer need the "friend/enemy" nVIDIA hanging around and quite possible will start drowning it with the above-stated arguments.