Originally posted by: apoppin
Who
really knows?
i see two out of five of your examples with a *?* ... with many 'holes' in-between
what i am saying IS ridiculous is
128-bit interface getting 12,000
RV630 [like x1700] is the midrange ... 128-bit - 12,000
R130 [like x1300] is the lowend ... 64-bit - 1,200
there would be a LOT in between ...
i think they picked that figure
out of the air
The two parts with the "?" are parts that we're pretty much ceirtain will be there but don't know the names to yet.
I don't see why everyone makes such a big deal about the memory widths. Having a thinner pipe at higher clock speeds IS better than having a wider pipe at lower clocks, and the clocks on these parts is pretty damn high. Yes, it would be better if the widths were higher (without sacrificing clock speed, of course) but having insanely complex PCBs for mid-range parts is a no-no, since margins on these parts are quite a bit lower than on high-end parts.
Just for comparison's sake: 2GHz (effective), 128bit memory can yield an aggregate bandwidth of around 32GB/sec; 1.55GHz (effective), 256bit memory can yield around 49.5GB/sec. Do remember, however, that you need to have all the memory channels active at full load to reach these numbers, 4 on the 128bit system, 8 on a 256bit system.
Considering that these parts will likely have better buffers, caches, fetchers and the like, not to mention that the new architecture itself is likely to be more efficient, the overall usable bandwidth may actually be even closer.
The numbers for the 64bit parts do seem a bit low, in my opinion, but not by as much as you'd expect. Bandwidth for 800Mhz (once again, effective), 64bit memory should be around 6.4GB/s, which is about 1/5th of what's available on the 2GHz, 128bit parts. So about 1/5th of the bandwidth available to the higher-end part to only achieve about 1/6th of the performance. There are many possible reasons for this, though. Core clock for the low-end parts could simply be too low to utilize even this relatively small amount of bandwidth; the low-end parts may just have a lot less hardware than the higher-end parts (this is the equivalent of the clock being too low, since the effective compute power per second is the product of the number of units and their speed); or, like the RV610LE article mentions, the drivers used for that test could be to blame. Even so, I still think 2k for 3DMark '05 is pretty damn good for the lowest-of-the-low, expect the RV610Pro (with its 1.4GHz-effecive memory) to yield close to 2x that performance.
EDIT: Fixed a stupid sentence, hopefully before anyone saw it
