The mid-range enigma
While considering ways to start this article, it occurred to me that the card on test possibly ought not to exist. With benchmarking done and a set of performance numbers which with to reference, along with an overview of the feature set and price, it seemed strange to me that ATI authorised Hercules to produce it (if indeed ATI could control that). You'll see why as the review progresses.
Whatever the reason for Hercules deciding to produce it, the card most certainly does exist and what a card it is. Hercules' 3D Prophet All-In-Wonder 9800SE seeks to do what ATI are doing officially with their own, brand new, 9600 Pro All-In-Wonder. Bringing the All-In-Wonder feature set to the masses is something almost all technology commentators would like to see, especially with the convergence of the PC and regular entertainment devices. With versions of the All-In-Wonder sat at both ends of the performance spectrum already, it's a mid-range refresh AIW product that's been missing.
With the mid-range being the main money maker for graphics card manufacturers, ATI have been keen to get their 9600 Pro All-In-Wonder just right. However, with them missing the boat in terms of the promised new RemoteWonder upgrade and with just a couple of extra features that the market demanded (such as dual head output) since the last round of AIW's, you can see why Hercules decided a 9800SE based part might be worth a shot. Along with the 9600 Pro design being slightly tardy to market, things are beginning to make a little more sense.
However, we need to take a closer look at Hercules' implementation of 9800SE. My concerns around ATI authorising its production should become clear by looking at the following specification table. Missing out All-In-Wonder features is desirable for brevity, I'm simply seeking to compare 9800SE performance with 9600 Pro and the more recent 9600XT. Here goes.
|Radeon 9800SE||Radeon 9600 Pro||Radeon 9600XT|
|Manufacturing process||150 nanometre||130 nanometre||130 nanometre|
|Pixel shader units||1 per pipe||1 per pipe||1 per pipe|
|Memory bus width||256-bit/32-byte||128-bit/16-byte||128-bit/16-byte|
|Texturing units||1 per pipe||1 per pipe||1 per pipe|
|Memory clock||680MHz DDR||600MHz DDR||600MHz DDR|
|Pixel fillrate||1520 Mpixels/sec||1600 Mpixels/sec||2000 Mpixels/sec|
|Texture fillrate||1520 Mtexels/sec||1600 Mtexels/sec||2000 Mtexels/sec|
While 9800SE should really be paired with a cut down 16-byte memory bus, there is a version of 9800SE available with double that bus width. A 9800SE with 256-bit memory interface is no more than a full Radeon 9800 Pro with 4 of its pixel pipelines disabled.
Given that GPU cores are made in batches on a silicon wafer, along with the fact that testing every core on the wafer for defects is costly and time consuming, it's fair to say that something along the following lines happens.
Given a wafer of say 50 R350 cores, a random selection of those (maybe 10% say) will be tested for defects. Should all the tested cores check out fine, that wafer of cores is marked as fine and all the cores are then used to produce full Radeon 9800 Pro products. Should any of the random test sample fail the testing, the wafer is to be discarded completely (if all selected cores fail initial testing), or used in 9800SE products with 4 pipelines disabled. The second scenario is more desirable, since it means that ATI get to make use of cores that may be slightly defective, but not entirely so.
It also has the statistical advantage that some cores on the defective wafer may be perfectly fine, with all 8 pixel pipelines able to work correctly, essentially giving rise to fully working R350 cores sold as cut down 9800SE's. With final QA testing done by the board manufacturer, after it purchases packaged wafers from ATI (produced by TSMC), any completely defective cores are discarded at that point in production (a statistically small amount). This lets fully working R350 cores slip through the cracks as 9800SE's, which are then sold accordingly, at an attractive price point befitting a mid-range entrant, to consumers.
We then have the softmod scenario, where fully working cores can then have their 4 'defective' pixel pipes enabled by the card driver, giving you free performance.
It's Hercules choice to use the 256-bit memory interface version of 9800SE, sold to them by ATI, that gives rise to their curious All-In-Wonder. With the same memory clock as a full 9800 Pro and the aforementioned memory interface, the 3D Prophet All-In-Wonder 9800SE is able to offer a monstrous memory bandwidth advantage compared to 9600 Pro and 9600XT, while still being a mid-range product. Especially compared to the 9600 Pro, where core clocks are very similar, Hercules card would definitely be the faster of the two in terms of performance, eclipsing the official middle markey ATI All-In-Wonder product for very similar money. While in pixel pushing scenarios the 9600XT's core clock advantage may well offer more performance than the Hercules, that memory interface advantage is very hard to ignore.
We'll see just how hard in upcoming pages. Just be mindful of the potential performance of the little enigma as we take a look at the card bundle and feature set, before we look at performance explicitly.