Upper limits with cpu ??

[Guest]

For Elrond,Techie question....building an AMD XP2100 system any point of going with the DDR2700???? I know it will be obsolete by end of summer.ThanksTony

[Guest]

Hi Max,"transfer everything across the AGP bus and into the video card's local frame buffer"Thats where you're missing the point. Think of the approach like tile rendering, but instead of for polys and vertices, this method is for textures. With an approach like being implemented into the P10 (and many earlier yet un-built designs), you *don't* transfer everything across the AGP bus and into on-board texture storage... You *only* transfer the subset of textures (note: not even the full textures, but in actuality subsets of full textures) across the bus that are actually required to render a scene. As Anand has thoughtfully pointed out, this is very akin to what the 256KB or 512KB of cache on existing processors does - but in this instance its 40+MB of cache. The onboard highspeed memory is used for two things in this design, one the same and one drastically different: The buffer storage (which is the same) and extremely high speed cache (which is different). The third use is completely dropped - texture storage (well, dropped on the card and tasked to much more plentiful main memory instead).This approach doesn't rely upon or require an extremely fast AGP bus because in reality smaller amounts of sub-texture data are transferred instead of having to transfer *all* required full-size textures as today (realize that those full-size textures that load in 256MB of high speed DDR have to get there somehow today - and usually because of the limited amount of texture storage (even at 256MB), swapped quite often). With the different approach and re-tasking fast onboard memory as a high speed cache, AGP 8x will be more than enough headroom. This is in contrast to some arguments where AGP 8x may *not* have enough headroom in the standard on-board texture storage methodology as implemented in todays cards. And the new approach allows for drastically more texture storage than any existing design.Time will tell how affective this design is: but I've always held high hopes for such an implimentation.Take care,http://members.rogers.com/eelvish/elrondlogo.gifhttp://members.rogers.com/eelvish/flyurl.gif

[Guest]

Sorry... Double posted.

[Guest PaulL01]

>Sorry... Double posted. I'm sorry sir you'll still have to pay full price for admitance, O my Lord! Sir...:-lolP

[Guest Max Cowgill]

Elrond, when I made my previous statements about textures being transferred via the AGP bus yada yada yada, I was referring to previous/existing implementations, not the virtual memory/caching system of the P10. Unfortunately, the P10 is only one product and as far as we know, there are no other planned consumer products with a similar caching system (certainly no existing ones). This is for consumer-level hardware only, exclusive of professional-level hardware. Who knows what the Parhelia, NV30, R300 and future products will have in store for us though? There may indeed be a transition to caching systems for future products, and I agree that it would be much better than today's simple frame buffer implementation (there's that word again). Anyway, I'm off to work again in another 20 minutes or so, so I won't be able to continue this discussion further until tonight (if necessary) ;)Max Cowgill

[Guest]

Ouch! Full price for both or do I get a freebie on the second? This toll bridge is spensive man!"O my Lord! Sir..."Ah, but that makes the price a pitance my second. :-) http://members.rogers.com/eelvish/elrondlogo.gifhttp://members.rogers.com/eelvish/flyurl.gif

[Guest]

Hi Tony,Well, although I personally wouldn't go the DDR333 route quite yet, there can be a valid point as long as you invest in CL2 DDR333 instead of CL2.5. At todays prices, this can be quite expensive however. If you opt for the existing CL2.5 DDR333, you won't see more than a one or two percent improvement (which in practice would be more than un-noticeable). At CL2, you may see up to ten percent speed improvement which is just at the edge of becoming noticeable in most situations.The caveat here is what your intended uses are. If professional video editing, 3D design, photo editing and the like is on your plate, indeed PC2700 even at CL2.5 in some instances could come in handy. For standard business/development/sim/game use as is the case for most home users however, I wouldn't touch it. Unless of course it was the same price as PC2100 at CL2, then I would (but its not :-)).So to summarize: if you can afford CL2 PC2700 or better at a $100-$200 premium per 128MB over PC2100, it may be quite worth it - specially if you plan on bringing it forward to future platforms where the design is more suited (Athlon64 as a possible example). Waiting for DDR400 may be more optimal as Samsung has just started shipping in bulk and the KT400 chipset should be released this fall. For the sake of completeness however, be aware that there is now quite the argument over weather DDR400 should be bypassed this year completely for DDR-II at 400Mhz next year.I myself have chosen to stick with high-quality Corsair PC2100 CL2 for the time being for pricing reasons. The memory market right now is indeed extremely difficult to recommend from.Take care,http://members.rogers.com/eelvish/elrondlogo.gifhttp://members.rogers.com/eelvish/flyurl.gif

[Guest]

Graci...mon ami!!!As always thanks for your tips and knowledge....send me the bill and charge me double.RegardsTony

[Guest]

Max, as we see with todays machines, even with the best nVidia card around, the system is still completely constrained by the agp port speed and it's ability to move data back and forth between the local ram and system memory.Given a higher bandwidth at this critical point, the requirement for local video memory is greatly reduced. If textures and data can be moved into and out of the local video ram at a reasonable rate, they do not have to all be stored locally to make the card function like we are stuck at right now.a 24 bit colour screen requires assuming cranked resolution of lets say 1920 x 1440 is roughly 11 Megs of data. This resolution would give plenty of room for the card to alias the image down to 1280 x 1024. Assuming some advanced buffering required of lets say 3 frames, we end up needing 40'ish megs of onboard video.To sustain a acceptable rate of frames (30 - standard dvd quality), you are moving 11 Megs x 30 frames or a 324 Megs/second tranfer rate.Assuming the video card is kept busy with texturing and lighting due to effects, you STILL need to move the data ... Additionally, keep in mind, that with the continued silliness with use of large textures, even a 128 Meg card can only store 2 4096 bit textures or 32 1024 bit ones.We are driving up requirements exponentially yet only doubling resource availability over the last year or so. This can not continue unchecked as we have all seen that the addition on one silly little addon using high res textures can wipe out the most powerful PC on the planet.Funny that the real world, that deals in super high end data has no problems with these issues, yet us PC guys are wiped out by a simple game ;-)Ray

[Guest Max Cowgill]

>Max, as we see with todays machines, even with the best >nVidia card around, the system is still completely >constrained by the agp port speed and it's ability to move >data back and forth between the local ram and system memory. >>Given a higher bandwidth at this critical point, the >requirement for local video memory is greatly reduced. If >textures and data can be moved into and out of the local >video ram at a reasonable rate, they do not have to all be >stored locally to make the card function like we are stuck >at right now. >>a 24 bit colour screen requires assuming cranked resolution >of lets say 1920 x 1440 is roughly 11 Megs of data. This >resolution would give plenty of room for the card to alias >the image down to 1280 x 1024. Assuming some advanced >buffering required of lets say 3 frames, we end up needing >40'ish megs of onboard video. >>To sustain a acceptable rate of frames (30 - standard dvd >quality), you are moving 11 Megs x 30 frames or a 324 >Megs/second tranfer rate. >>Assuming the video card is kept busy with texturing and >lighting due to effects, you STILL need to move the data ... >Additionally, keep in mind, that with the continued >silliness with use of large textures, even a 128 Meg card >can only store 2 4096 bit textures or 32 1024 >bit ones. >>We are driving up requirements exponentially yet only >doubling resource availability over the last year or so. >This can not continue unchecked as we have all seen that the >addition on one silly little addon using high res textures >can wipe out the most powerful PC on the planet. >>Funny that the real world, that deals in super high end data >has no problems with these issues, yet us PC guys are wiped >out by a simple game ;-) Ray, I don't see why you addressed your post to me as though I don't know we need a faster AGP bus... I've already acknowledged this several times in this thread... Anyway, what formula(e) did you use to calculate your example of storing textures? I've got the formula for calculating necessary bandwidth but not the formula for calculating the memory necessary to store a texture (I know it's something along the lines of (xres*yres*bitdepth)/something)). Sad thing is I used to know this equation like the back of my hand :(Max Cowgill

[michal 44]

Hi Max - waiting for any correspondence with you regarding this surplous RAM. Cheers. :-)Michael J.michaljast@frys.com

[Guest Max Cowgill]

Sorry about that Michael, I've been working late nights recently and haven't had much time for anything else. I'm going to bump my "upgrade heaven" thread one last time to see if anyone else badly needs some more ram.Max Cowgill