The ability to cache the video and system BIOSes has been a feature in our BIOSes for a long time. But there's now a new BIOS option called Video RAM Cacheable. Normally found in the Advanced Chipset Features' menu, it's the option you can see boxed in red here :-

What's this BIOS option for? Well, according to the manuals, this feature enables or disables the caching of the video RAM at A0000h-AFFFFh region via the L2 cache. This is supposed to significantly speed up accesses to the video RAM. However, will the caching of this small 64KB region actually boost the performance of the graphics card and the system overall? Let's examine the theoretical aspect first :-

First, let's find out the RAM bandwidth of most graphics cards these days. The average graphics card should have a 128-bit wide memory bus with the onboard RAM running at about 150MHz (RIVA TNT2 as a model). That translates into a RAM bandwidth of 2.4GB/s. That's much lower than the RAM bandwidth of many higher-end cards (4.8GB/s with a 128-bit bus running at 150MHz DDR) which will soon become the norm but 2.4GB/s will be the number we will be using as an example.

Now, the Celeron 504MHz (that I'll be using for the tests) has a bandwidth of 4.0GB/s (64-bit x 504MHz). That's only 1.6GB/s higher than the average bandwidth of the average graphics card or 0.8GB/s slower than the memory bandwidth of higher-end cards. So, there's little use in caching the already fast video RAM. On the other hand, the motherboard RAM's bandwidth is only a pitiful 0.8GB/s (64-bit x 100MHz), which is way slower than the L2 cache (only a miserly 20% of the L2 cache's bandwidth). It would be much better to use the limited amount of cache to speed up the slow motherboard RAM instead of the much faster video RAM.

Finally, remember that data from the L2 cache has to pass through the AGP (or in some cases, the PCI) bus. Although the AGP bus will now be faster with the introduction of AGP4X, it's still way slower than the transfer rate between the onboard RAM and the graphics chip. What this means is that while the L2 cache is much faster than the onboard RAM, the graphics chip can only access the data there though the AGP (or PCI) bottleneck. Whatever performance boost from the speed of the L2 cache will surely be lost in the transfer through the AGP (or PCI) pipeline. Needless to say, such additional traffic on the AGP will also reduce the bandwidth available for texture transfers, etc...