When Tuesday does the G70 NDA expire?

Of course I can

 

With something like HT you'll probably need faster cache to feed the wider design.
To have faster cache it must be smaller.
If its smaller it won't be able to feed 3 pipes. So you'll have to make'em 2
To compensate this, you'll need to increase GHz.
To increase it you need to make whole pipeline deeper/longer
making it longer you'll loose efficiency, and lower IPC
mmmmm let me see... i see some similarities with Netbust....

Remember mp6-Rise? http://www.cpu-world.com/CPUs/CPU.html
It was first x86 CPu with 3 FPu pipes. Shined on benchmarks. Sucked in real life. It had small 16KB cache and was unable to feed 3 pipes....

 

I don't see why you'd need faster cache. It's not like you're actually adding any new pipelines to the architecture. If the cache is fast enough to feed all current functional units in the K8 right now, how would Hyperthreading make things worse?

 

point is.
netburst has 2 fpu units and fast 16KB L1 instruction cache, directly taking data from fast L2
K7/K8 have 3 units and bigger, slower 64KB data/instruction caches

I do believe AMD and Intel had reasons to choose such designs. All i can say is possible explanation, not the right one.

 

I'm gonna repeat myself here, but could someone clarify this:

Can 7800 GPU pixelshader units do TWO times 4-component vector op + MADD (+ extra MADD from miniALU ?). Or is it two times 3 + 1 ?

 

It can do two individually splittable (3:1 and 2:2) 4-component MADD/DP ops. The Mini-ALUs do simple operations like scaling (divide/multiply by power of two) invert and saturate (clamp to [0, 1]).

 

As Xmas mentioned, I don't believe the extra MADD power comes from the mini ALUs as they are for pixel modifier ops. From what I read in Dave's recent article on G70, each unit ALU unit, SU0 and SU1, was provided with a MADD or a MUL depending on the type of unit that was lacking in shader previously in NV40 . If this is the case, SU0 recieved a MADD (since it was previously composed of only MUL+RCP/RSQ) and SU1 recieved a MUL (since it was previously composed of only MADD+Special function). Xmas may correct me if I'm wrong.

 

MUL is a part of MADD (=MUL+ADD), so there was no need to extend SU1. SU0 now has ADD ability, too.

 

There is a lot of confusion around what has changed. Nvidia has been quite confusing around the MAD stuff because they just wanted people to understand that they doubled the floating point power.

Here is a basic pipeline illustration that shows the difference between 6800 and 7800 math units: http://www.hardware.fr/medias/photos_news/00/13/IMG0013260.gif

 

ROTFLMAO!!!!!! ^^^^^^

 

Ok I removed about 5/6 duplicate post here... sorry for the troubles as the board was having some issues

 

Howard, is that you?

"The way of the future. . .the way of the future. . ."



I've been having "now it's there; now it's gone" issues with B3D the last couple days as well.

 

Lol I don't know what happened. Maybe the forum went mad about the mad stuff

 

MAD is MUL + ADD. There was already a multiplier in the first unit. Basically Nvidia just added an adder in that unit.

 

Thanks for the clarification Trident and Xmas. After reading the B3D G70 review I was given the impression that Nvidia made both units in G70 MADD+MUL rather than just MADD or MUL alone, as with NV40. It is now clear that Nvidia's additions just made it a MADD unit in each shader unit (SU0 and SU1) plus the special function goodies that were in SU0 and SU1 previously.

 

No, nothing wrong with an extra ADD, as the benchmarks show. People keep saying "MADD + MUL," though. Tridam was pointing out that they're misusing the terms.

MADD + MUL = MUL + ADD + MUL.

G70 pipe = 2 * (MUL + ADD) = 2 * MADD. There's no extra "+ ADD" after MADD. MADD stands for MUL + ADD, not, say, matrix addition (what else are people guessing? ).