Level 2 Cache 3MB vs 6MB

You probably won't find a 'conclusive' answer anywhere. It will make a slight difference. Slight.

I'm not concerned about the extra VRAM on the GPU.

Not sure why, but that's likely to make a more significant difference than the L2 cache.

Only you can determine if those differences are worth the cost.

Here's a Xbench, memory test comparison between my, returned, mb 2.4 (3MB), 4GB ram and my mbp 2.66 (6MB), 4GB ram, both unibodies running 10.5.6:

Note the STREAM (all about memory bandwidth) marks -

Macbook:
Memory Test 175.69
System 196.44
Allocate 254.23 933.60 Kalloc/sec
Fill 164.70 8007.99 MB/sec
Copy 189.87 3921.74 MB/sec
Stream 158.91
Copy 154.36 3188.22 MB/sec
Scale 148.57 3069.31 MB/sec
Add 169.70 3615.00 MB/sec
Triad 164.75 3524.50 MB/sec

Memory Test 181.43
System 207.51
Allocate 236.94 870.13 Kalloc/sec
Fill 186.60 9072.78 MB/sec
Copy 205.01 4234.43 MB/sec
Stream 161.18
Copy 149.43 3086.39 MB/sec
Scale 154.32 3188.22 MB/sec
Add 172.30 3670.29 MB/sec
Triad 171.20 3662.48 MB/sec

So, as it would would appear it doesn't mean much? However, when a application needs to maintain sustained memory bandwidth such as in some large, framed, vector math processing scenario's then computation time adds up and relative time will be saved with the larger cache, overall.

See this paper on Stream bench-marking: (Note in the article that other variables come into play such as 32-bit vs 64-bit vs shared ram, how an application is written, etc ...)

http://www.cs.virginia.edu/stream/ref.html

In my case I spoke with the company engineer that designed the application, I use, and got a recommendation for key considerations on hardware choice.
However, I exchanged my MB because, it was mis-behaving, and chose the mbp because I hear tell that Snow Leopard may off-load processes to the 'GT. As well, the extra cache would be of benefit for cutting down the time needed for running large simulations and processing routing and placement algorithms in the applications I run. Otherwise, the 2.4 MB is so close and in some case's faster in the benchmark(s), as well as equally performing when it came to memory bandwidth.

gtds - You might have read that paper - I don't want to sound presumptuous!

According to the benchmark, no. But, Xbench scales the test time and I do not know the battery of test used.

Not your app's but another benchmark: http://www.anandtech.com/mac/showdoc.aspx?i=3246&p=12

It look's like many people don't choose to spend the extra on the cache. Again the P8700 does a fantastic job.

You know how it is though, real life may be a different story. For example, the '8700 (just noticed, should have said '8700 in previous post too!) and '9550 can share the cache between the duo core's. So will the '8700 but it has half the amount. Right there, under worse case loading, could mean a 50% reduction in time spent processing each task whether from the same or parallel process. An extended (time consuming) tasks, especially, the same tasks, running sequentially, large blocks of data and operands, over and over, would be more predictable and therefore, if software has provided for, exploit cache scheduling further.

There's a person, on this forum, that deal's with those application's. His name is Guy Mancuso (as well as other's). Mr. Mancuso has found it beneficial to install 2 Intel SSD's and 8GB of ram in his 2.93, 15", Macbook Pro. I'll look around the site a see if I can dig up where he posted the percentage's in time (cost) saving's.

Oh, and your comment about Snow Leopard and Open GL is good planning, in my opinion. I figured the additional cache may be an advantage. I hope so!