Initial findings: It seems HL2 likes to set around 13MB to system memory regardless of how much VRAM you have. This doesn’t mean anything negative, its just the way this particular game manages it's memory allocation.
Hold mouse over the links below to see the difference:
The first test up is with maximum quality settings within HL2 but with AA, AF or HDR off. Under these circumstances we never see more than 220MB being consumed. Results are almost identical across the board but the 512Mb card does show a little performance boost in resolutions above 1600x1200, though according to the data it’s not because of memory allocation.
The above tests are not indicative of how end users will set HL2 to run anyway (no AA, are you crazy?), so though you can manage to hold all data in VRAM it’s a useless test in real world terms.
Hold mouse over the links below to see the difference :
Now we turn up the heat a little by setting 4x anti-aliasing and 8x anisotropic filtering. With absolutely no increase in texture usage we see some dramatic memory usage gains by turning on these common settings. Let’s take 2048x1536 as an example. With no AA and AF the memory usage was 217.78MB. With 4xAA and 8xAF the memory requirement has jumped to 375.52. The 256Mb card has needed to tap into 99.05MB of system memory while the 512MB card has no problems fitting the textures and geometry within VRAM.
The performance numbers are quite interesting. We see a healthy boost in every single test here from the 512MB card, but most note worthy are the results at 2038x1536. There is a 16.7% performance boost with the 512MB card here.
Hold mouse over the links below to see the difference :
Now we just max out the settings by adding HDR into the mix. Most people will want to run this game with these settings for maximum image quality. Why else would you spend £250 or more on a GPU? You can cleary see that the only resolution that can run within a 256MB framebuffer is 1024x768, anything higher and your going to be storing data in system RAM, but does this matter? Looking at the results above we get the first indication that a little texture swapping does very little, but a lot can cripple performance.
At 2048x1536 it’s using 434.2MB, and the 256MB card needs to put 143.54MB of texture data into the system RAM. Not only this, but the totals do not add up. VMW reports the 256Mb card is using a total framebuffer size of 388.23MB whereas the 512Mb card shows 434.2MB. Where has 45.97MB gone? Could it be that we’re seeing signs of cache thrashing? The performance numbers would suggest this, and so would my subjective look at the timedemo used. At specific times in the timedemo the 256MB card would crawl along, exhibiting typical stuttering caused from cache thrashing. I cannot confirm this is the case, but it looks this way to me.
Summarising all the tests, it looks as if 512MB clearly is necessary in this game but only really from a technical viewpoint, but we like those! 2048x1535 with everything on is unplayable on both cards, so I’d have to say that though 512MB is definitely faster it adds no real world gains to this game. Initial findings show that a little texture swapping has no real impact on performance and is nothing worth worrying about. Let’s see if the others games react differently, starting with Quake 4.
PowerColor PCS+ 5830 1GB
Best Price:
