Dima Posted July 26, 2002 About that exciting thing I promised, I'll make it short, so here are a couple of things everyone should be intrested in: 1. John Carmack On NV30: Nvidia is the first of the consumer graphics companies to firmly understand what is going to be happening with the convergence of consumer realtime and professional offline rendering. The architectural decision in the NV30 to allow full floating point precision all the way to the framebuffer and texture fetch, instead of just in internal paths, is a good example of far sighted planning. It has been obvious to me for some time how things are going to come together, but Nvidia has made moves on both the technical and company strategic fronts that are going to accelerate my timetable over my original estimations My current work on Doom is designed around what was possible on the original Geforce, and reaches an optimal impliementation on the NV30. My next generation of work is designed around what is made possible on the NV30 2. During Siggraph 2002, NVIDIA revealed their new, revolutionary CineFX architecture that NV30 will be based on, instead of writing tons of information about it, click here to download a paper describing everything. Another important paper about this next-gen architecture can be downloaded here. Also, a very intresting article on the architecture can be found here: http://firingsquad.gamers.com/hardware/cinefx/default.asp To summarize it all quickly, here are the main advantages of the architecture: Vertex Shaders beyond DirectX 9 with up to 1024 static instructions, up to 65536 instructions executed in loops, branches and subroutines Pixel Shaders beyond DirectX 9 - up to 1024 instructions Long programs for both pixel shading and vertex shading Support for OpenGL and DirectX HLSL via NVIDIA's Cg New focus on computational efficiency rather than memory efficiency (The 'focus on computational efficiency' suggests that there are new overdraw reduction routines in the pipeline as doing this level of computations per pixel can end up as being very expensive for pixels that are overdrawn, however it appears that no details of these routines are in the public domain as yet. Likewise there is also currently no mention of what FSAA methods or Anisotropic filtering is supported - seeing as most of these details are coming about because of Siggraph its not surprising as these are more features that are important to users rather than developers; no doubt developers are more eager to be able to play with the vertex and pixel shader functionality!) OpenGL extensions that support long pixel and vertex shaders programs 64 & 128-bit floating point color precision High bandwidth Support for high-speed DDR2 memories 8X AGP Advanced programmability and high-level shading language support Other chip-specific details remain under NDA until later this year. As you can see, the architecture goes BEYOND the specifications of DirectX 9 and it's all accomplished using NVIDIAs CG language (more on that in the papers above). Other important things about the architecture, from the mouth of a very trusted source who was in contact with Nvidia during the show: 1. It will be marketed as the first Cinematic Shading GPU 2. Pixel shaders can be 1024 instructions in a PASS . Of course you can multipass, but the program instruction length is considered a pass ALU functions for the Pixel Shader: ADD,DP3, DP4, LRP, MOV, MAD, SUB, X2D (add and multiply) ,TEX, TXD, TXP (texturing), COS, EX2, FLR, FRC, LG2, POW< RCP, RSQ, SIN (math instructions), SEQ, SLF, SGR, SGT, SLE, SNE, STR (set "on" instructions), DST, LIT, RFL (graphics), MIN, MAX (minimum , maximum) PK2H , PK2US, PK4B, PK4UB, PK4UBG (pack) UP2H, UP2US, UP4B, UP4UB, UP4UBG(unpack) , Kill (kill) 3. Vertex Shaders programs can be 65,536 instructions in length with the loops branches etc , it's also considered one pass . Vertex Shader Instructions: ARL (4 component A0 and A1, ARR rounding instruction instead of truncating like ARL, BRA , CAL, RET (branchching instructions) COS , SIN (high precision trigonometric functions), FLR, FRC (floor and fractions of floating point values), EX2 , LG2 (high precision exponentiation and logarithym function), ARAm SEQ, SFL, SGT, SLE, SNE, STR, SSG 4. .13 micron 5. NVIDIA said fall part in June , and I've heard nothing from them that changes that outlook . despite the rumors. 6. NVIDIA will be using a form of DDR2 memory . This is CONFIRMED information. What clockspeeds , what bit interface, or memory bandwidth I cannot confirm at this time 7. 12bit fixed point (sorry for confusion), 16bit 32bit float formats supported 8.16 textures per pixel ala DX9. 9. 16 texture units I cannot at this time confirm the configuration of those 16 units but they are there 10. NV30 and DX9 schedules are now aligned. This may change if Microsoft delays DX9 , but not the other way around. I think it's time to get excited! :) P.S If there are any questions about the architecture or various features, forward them to me and i'll do my best to answer them! :) 0 Share this post Link to post
gatewatcher Posted July 26, 2002 so what carmack will do next is build arround gf3's and above? 0 Share this post Link to post
Dima Posted July 26, 2002 gatewatcher said:so what carmack will do next is build arround gf3's and above? Carmack's next engine will be built around the R300/NV30 architectures, with full floating point 64 & 128 bit precisions (switchable I suppose, like 16 & 32 bit are now). As was noted on many websites, NV30's architecture is VERY flexible (much more flexible than R300's), it's a programmers dream come true! GF3's architecture is very limited, especially when it comes to the lengh of the pixel & vertex shaders instructions, it lacks important features such as 64 & 128 bit color precisions, flow control in VS, among many other things that are WAY too technical to discuss here! :) 0 Share this post Link to post
Lord FlatHead Posted July 27, 2002 Firing Squad has a pretty interesting article on the NV30 as well. It's mostly a comparison of between the raw number-crunching power of the GeForce3&4 and this new chip. 0 Share this post Link to post
Dima Posted August 25, 2002 Disorder said:Any idea how much the NV30 will cost? Yes, around 400-450$ 0 Share this post Link to post
Zaldron Posted August 25, 2002 Dima said:Yes, around 400-450$ Not surprising considering the GF3's initial price. 0 Share this post Link to post