Chip Comparison Chart
- Thread starter Dave Baumann
- Start date
I think that's wrong, it's 128x1 instead. I know there were some discussions of this already, but IIRC no final conclusion was reached. There are however some strong hints the memory controller is not as efficient as on the R200, the single texturing fillrate (measured with 3dmark2001 so it's mem bandwidth limited) of a Radeon 9000pro is about 15-20% slower than a Radeon 8500 (at the same core/mem clocks), so I tend to believe it's 128x1. Some reviews (digit-life for instance) claimed the RV250 has the same memory interface as theUttar said:
RV200.
OpenGL guy
Veteran
128x1? That sounds like a bad joke. If you think that RV250 is different from R200 (which it may be, I have no idea), then maybe you should rethink what R200 is? Just a thought.mczak said:
OK, fairly large update, with some new fields that I need some more assistance with:
http://www.beyond3d.com/misc/comp.php
BTW - formatting is horrid at the moment because I'm only populating it at the moment, the display will be very different once completed.
http://www.beyond3d.com/misc/comp.php
BTW - formatting is horrid at the moment because I'm only populating it at the moment, the display will be very different once completed.
A suggestion :
Making this a comprehensive chart on a single page would make life hard for you as well as the readers, Dave. Instead, for certain categories, link certain things to a separate (pop-up, fixed window size) page. An example would be expanding on the ps.
Also, I would personally appreciate linking each and every chip to actual product names (again, maybe pop-up window). With the way NVIDIA and ATi is trying to provide for every segment of the market, it's pretty easy to forget which is what.
Making this a comprehensive chart on a single page would make life hard for you as well as the readers, Dave. Instead, for certain categories, link certain things to a separate (pop-up, fixed window size) page. An example would be expanding on the ps.
Also, I would personally appreciate linking each and every chip to actual product names (again, maybe pop-up window). With the way NVIDIA and ATi is trying to provide for every segment of the market, it's pretty easy to forget which is what.
As I said, this is purely for information gathering purposes at the moment. I have a number of ideas for eventual display of the data.
Xbit-labs has a new review of the 9100. In it they say the reason 128MB 8500's were faster than their 64MB predecessors is because the former user 4x32-bit memory chips, which allowed them to use interleaving. Apparently the 9100's are slower than the 8500LE's because they use 8x16-bit chips, thus no interleaving. Is this true? Does it qualify for a table mention, or is this more of a board-level detail?
Its what each of the texture units are capable of sampling. i.e. Parhelia has 4 TMU's in each of its four pipes, but each unit is only Bilinear capable on its own, so two are required for Trilinear filtering; NV10 has one texture unit per pipeline, but this is a Trilinear unit so there will be no fillrate difference between Bi or Trlinear on this chip.
PVR_Extremist
Newcomer
The original Kyro did AGP4x Dave
It was the Kyro 2 crippled with AGP 2x
T
It was the Kyro 2 crippled with AGP 2x
T
ta
Why do you think this can't be? The Parhelia even has a 256x1 memory bus (not that I say it's efficient). Older Radeons also have 128x1 memory interface, and IIRC that 64x2 of the R200 was an official figure, so going back to 128x1 for the budget RV250 doesn't seem to be out of the question. I don't insist it must be 1x128, it might be possible it's the same as R200 but the cards don't use memory interleaving and way lower ram timings which maybe could explain why it's less efficient. But since a lot of reviews said the memory interface is different to R200 (but most didn't specifiy exactly what is different) I'm going to stick to the 128x1 theory (unless, of course, you can officially deny that...). Maybe I should take a DMM and try to figure it out on my R9000pro (the address signals should be separated on 4 chips from the other 4 ram chips if it's 64x2, but the same on all 8 chips if it's 128x1).OpenGL guy said:
OpenGL guy
Veteran
Says who?mczak said:
Says who?
That was assumed because the memory controller doesn't look terribly efficient - even with a relative modest theoretical single texture fillrate and a very high raw memory bandwidth, it stays quite a bit below in single texture fillrate in practice. However, it looks like I'm wrong on that. The quote here, http://www.tech-report.com/etc/2002q2/parhelia/index.x?pg=2 is a bit vague, but speeks about "granular access" which should in fact mean it's a multi-channel memory controller.OpenGL guy said:
Well, nobody said otherwise - you'd think ATI would have pointed that out to show the Radeon is superiour to the GeForce2. But if you insist, I have a Radeon (SDR) around here, so I can use the DMM on that too
chavvdarrr
Veteran
info for older chips
A similar attempt, although site is for sell
http://reactorcritical.com/chips/summary-full.shtml , also there are some very thorough(IMO) reviews for older chips there.
A similar attempt, although site is for sell
http://reactorcritical.com/chips/summary-full.shtml , also there are some very thorough(IMO) reviews for older chips there.
Similar threads
