Kevin, I honestly have no idea. I guess it involves transistors b/w the ROPs and the memory controller. Maybe nano-leprechauns, too.
Greg,
Everyone wants pretty DX9 games, yes?
Oh God, yes.
all the [kids say] any precision less than FP24 isn't enough. They grudgingly accept FP16 for HDR. FX10 isn't even enough precision for ATi's DX8 gen SM1.4 (which requires FP12).
Er, wait ... what? You're mixing terms here, and so confusing me. (And no one's screaming anything. Leave "fanatics" out of this, FFS. If you think I'm a fanatic, then use it in singular form. Otherwise, let's stick to the ppl actively involved in this discussion, please.)
FP16, as I understand it (as s10e5), has less
precision than both FX12 (NV) and FX16 (ATI). It has greater precision and
range than FX8, though. The question of FP24 (s16e7?) or FP32 (s23e8?) precision in intermediate shader calculations seems unrelated to that of high (dynamic)
range buffers, where I believe we've been stuck at FX8 since roughly forever (TNT?). In the case of buffers, FX10 is an improvement, and FP16 a further one. It does seem fair to call FX10 more MDR and FP16 HDR, though even I'm not crazy enough to try and convince marketers what's proper.
That's news to me that PS1.4 requires FP12. I thought a) R300 was the first gen to offer FP in the fragment shaders, b) R200 was FX16, and c) NV2x was FX12. Either you've mixed a), b), and c) to come up with FP12, or I'm the one who's mixed up.
So, if you are going to use FX10 for HDR how are you going to hang on to your much vaunted precision for rendering effects correctly?
I dunno. FX10 seems to have the same precision as FP16 but not the same range. I don't know how the dynamic part of high dynamic range plays into this, but Source seems to be doing a decent job with what I believe are FX8 buffers. Then again, maybe what I'm seeing with my 9800 in CSS maps like Nuke and Militia is better classified as bloom. In any case, yeah, minimum FP16 straight through would be ideal, but is it currently feasible? That might be answered by the fact that ATI is "compromising" FP16 to allow for AA.
That's the question the [kids] need to answer - they pooh-pooh'd low precision in the past in favor of FP24, now suddenly FX10 is just peachy?
Apples and oranges? Eh, we'll see what future games do with "HDR." Maybe the upcoming Unreal Performance Test will shed some light on this.
---
Thanks, Apple, much appreciated. I'm not sure you can read too much into an instantaneous FRAPS grab, but you can see some FC HDR benches
here that show a X1800XT with 4xAA is basically twice as fast as a 6800GT without, a least at 1600x1200. The X1900XT/X are a good bit faster w/o AA, but it seems FP16 + AA is just throttled by bandwidth, as the gap b/w the X1900 and X1800 narrows. GDDR4 might blow this whole FP16+AA thing wide open, even for cards on 128-bit buses.
Cool. Just a few posts down in that NVN thread, 5'8 also compares "HDR" on a X1800 and 6800 in AoE3. Aside from some odd details (missing/reduced shadows, lighter static shadows in water, less pronounced bump mapping), it doesn't look like FX10 is noticably inferior to FP16, at least in the brightness intensity. Then again, I'm not sure if I'm seeing anything more than a touch o' bloom. The lighter static shadows in the water could be reduced precision at work.