I hope Tom (of Toms hardware) will decap a 9700 pro and a 9500 to at least shed some light on whether or not they're the same die.
ATI to launch Radeon 9500 and 9700 on the 18th
- Thread starter Sabastian
- Start date
RussSchultz said:
Nope ...
Humus said:don't know anything about the R9700 yields, so let's just pull a number out of my ass, I think it's realistic that say 40% of the chips are fully functional, and 30% are flawed in such a way that they can be used as a R9500
... that would mean half the failing chips would work for a 9500.
I'm of course speculating as wildly as everyone else, but I think it makes a lot of sense to reuse failing 9700 chips.
Entropy said:
I don't think this statement is accurate considering
1) the 9700 Pro would spend quite a bit of time waiting for data from the system for most of the systems where people might cut corners and get the 9500.
2) it ignores the part of the functionality, I'd think geometry handling for example, that is not directly tied to the number of pixel pipelines and not primarily limited by the memory bandwidth.
Well, why would the 9000 Pro have to drop in price? For the other features the 9500 brings to the table, we know pretty solidly that it should always be faster than the 9000 Pro...enhanced "bandwidth" usage should help even when the similar pipeline situation is limiting both, but we know that the number of TMUs isn't an absolute indicator of relative performance, correct?
Slower than the 4200? Where do you get the certainty of that? I've asked the question before about where the 9500 would fall based on comparisons to the 9000 Pro, and my own opinion is it would compete well with the 4200 atleast...why do you think otherwise?
demalion said:
True, but equally true for most higher end cards, i.e. on host limited tasks, they will all twiddle their thumbs and perform similarly. That's the "host vs. gfx-card" discussion though.
Raw fillrate.
'nuff said really, but if this wasn't the case, why is the 8500 faster than the 9000 Pro in just about every benchmark under the sun? And the nVidia 4200 trounces the 9000 Pro even more solidly in all reviews I've seen.
If the leaked specs are correct I'd predict that 9500 will perform below the 4200 in most benchmarks/applications, but will entice prospective buyers with its "future tech" rendering features. IF that prediction is correct, I can't really see how ATI can make a major market impact with the 9500 unless they price it pretty much at the same price as the 4200, preferably lower. If they price it higher, people will percieve that they pay more for less performance - never a happy recipy for market success.
Notice the big, bold IF though.
Entropy
Hmm...that snip seemed pretty pertinent to your comments about this necessitating the lowering of the price of the 9000/9000 Pro. Do you retract that part of your comment? They are competing with the 4200 already right now, and I still don't see why the introduction of the 9500 would change the way they are competing any more than the passage of time would (i.e., they are newer than the 4200 so the price has had less of a chance to drop).
As for fillrate...ok, 1) in which situations are fillrate the most limitng factor, and 2) on what are you basing your belief that it is likely the 9500 would suffer in this circumstance? Since they have the same number of pipelines, I presume your belief is based on the number of TMUs. To reintroduce the snipped text with my question I think is related:
Ok, could you share your thoughts on how this relates to fill rate?
I note looking at the digit-life/xbit 9700 Pro review that in 3dmark multitexturing fillrate comparison (please highlight the flaws in this test if you know and if they might confuse the issue) the Parhelia dominates at 1024x768, yet loses at higher resolutions. Also, in single texturing fill rate the 9700 Pro is approaching twice the fill rate of the 4600, as well as in all the situations where anisotropic filtering is used. To me, this combined with the rest of the benchmarks of the Parhelia, consideration of the impact of anisotropic filtering, looking at how the 9700 Pro performs in actual games in relation to even a 4600 when the 4600 dominates it in some fill rate tests, seems to indicate that without further factors specified it doesn't seem reasonable to believe the 9500 will not be able to compete well with the 4200 at the least, and that simply stating fill rate is not "'nuff said" on this issue (atleast for what I know so far).
As for fillrate...ok, 1) in which situations are fillrate the most limitng factor, and 2) on what are you basing your belief that it is likely the 9500 would suffer in this circumstance? Since they have the same number of pipelines, I presume your belief is based on the number of TMUs. To reintroduce the snipped text with my question I think is related:
Ok, could you share your thoughts on how this relates to fill rate?
I note looking at the digit-life/xbit 9700 Pro review that in 3dmark multitexturing fillrate comparison (please highlight the flaws in this test if you know and if they might confuse the issue) the Parhelia dominates at 1024x768, yet loses at higher resolutions. Also, in single texturing fill rate the 9700 Pro is approaching twice the fill rate of the 4600, as well as in all the situations where anisotropic filtering is used. To me, this combined with the rest of the benchmarks of the Parhelia, consideration of the impact of anisotropic filtering, looking at how the 9700 Pro performs in actual games in relation to even a 4600 when the 4600 dominates it in some fill rate tests, seems to indicate that without further factors specified it doesn't seem reasonable to believe the 9500 will not be able to compete well with the 4200 at the least, and that simply stating fill rate is not "'nuff said" on this issue (atleast for what I know so far).
Comparing with the Matrox Parhelia is irrelevant, both from technical and market perspectives.
I don't particularly care much about the fate of the 9000 prices. It was a side comment (in parenthesis, even) so I condensed my quotation, that's all.
The 9500 should be roughly equivalent to the 9000 in therms of fillrate, with some advantages here and there. The main differentiator between the two will be the more advanced 3D-features of the 9500. It would be reasonable to assume that ATI would charge a decent premium for that. On the other hand, I doubt they will find many buyers if they place the 9500 significantly above the 4200 in price. Positioning vs. the 9000Pro is an ATI problem, not a consumer problem, since consumers are more likely to spring for the 4200 anyway, and that's really the relevant part to compare to.
Fillrate, while not a sexy parameter in these circles at the moment, remains the most important performance determinant for existing games, and upcoming games where a mid-level card could be expected to serve for the next year or so. Pixel fillrate being more important than texel fillrate (according to my own, now outdated multivariate analysis even), but dual texturing is still a quite useful performance feature and is likely to remain so, again for the games these cards are targeted at. I doubt many (any?) of these games would benefit significantly from DX9 features. Ergo...?
The above discussion requires that the leaked specs are accurate.
They weren't particularly for the 9000, and price positioning is even more treacherous ground. I think we should tread carefully.
Entropy
I don't particularly care much about the fate of the 9000 prices. It was a side comment (in parenthesis, even) so I condensed my quotation, that's all.
The 9500 should be roughly equivalent to the 9000 in therms of fillrate, with some advantages here and there. The main differentiator between the two will be the more advanced 3D-features of the 9500. It would be reasonable to assume that ATI would charge a decent premium for that. On the other hand, I doubt they will find many buyers if they place the 9500 significantly above the 4200 in price. Positioning vs. the 9000Pro is an ATI problem, not a consumer problem, since consumers are more likely to spring for the 4200 anyway, and that's really the relevant part to compare to.
Fillrate, while not a sexy parameter in these circles at the moment, remains the most important performance determinant for existing games, and upcoming games where a mid-level card could be expected to serve for the next year or so. Pixel fillrate being more important than texel fillrate (according to my own, now outdated multivariate analysis even), but dual texturing is still a quite useful performance feature and is likely to remain so, again for the games these cards are targeted at. I doubt many (any?) of these games would benefit significantly from DX9 features. Ergo...?
The above discussion requires that the leaked specs are accurate.
They weren't particularly for the 9000, and price positioning is even more treacherous ground. I think we should tread carefully.
Entropy
Entropy,
For starters, if you haven't already I'd recommend looking over our 9000 PRO review as it directly compares against a 8500. Its surprising how close the 9000 is in many cases, however in SS and Quake3 based games, that rely heavily on multitexturing it does drop back a little. However, 9500 does have better Z buffer opimisations - I'm going to be interested to compare 9000 and 9500 at equal clocks if they are similar configurations.
However, the real importance here is the effect of AA and Aniso filtering 9500 will have. While GF4 4600 will scale quite nicely at high resolutions, its AA and Aniso is nowhere near as useful in gaming situations. The fact that R300's AA is fillrate free and compressed (meaning it much less bandwidth dependant) will be a big factor.
For starters, if you haven't already I'd recommend looking over our 9000 PRO review as it directly compares against a 8500. Its surprising how close the 9000 is in many cases, however in SS and Quake3 based games, that rely heavily on multitexturing it does drop back a little. However, 9500 does have better Z buffer opimisations - I'm going to be interested to compare 9000 and 9500 at equal clocks if they are similar configurations.
However, the real importance here is the effect of AA and Aniso filtering 9500 will have. While GF4 4600 will scale quite nicely at high resolutions, its AA and Aniso is nowhere near as useful in gaming situations. The fact that R300's AA is fillrate free and compressed (meaning it much less bandwidth dependant) will be a big factor.
Entropy said:
Entropy said:
Hrm as is now the price of the Radeon 9000 series cards is pretty much in stroke with the Geforce 4 MX 440. IIRC the Radeon 9000 smacks the heck out of that card and to boot it is DX8 while the MX is still only DX7. The market segment that the Radeon 9000 is aimed at is the low end rather then the mid stream occupied by the Geforce 4 Ti series ...... look here... price wise the Radeon 9000 is as low as the Geforce 4 MX 440.
http://www.newegg.com/app/viewproduct.asp?DEPA=&submit=Go&description=Radeon+9000
http://www.newegg.com/app/viewproduct.asp?DEPA=&submit=Go&description=geforce4+mx440
BTW does anyone have a clue about what happened to the Geforce 4 MX 460 that nvidia sent around to most every review site to have benchmarked against the Radeon 9000? Seems as though the card appeared solely to be benched against the 9000 to prop up the appearance that the MX series is not so far behind as they really are. BTW does anyone know where I can find a review that compares the Geforce 4 MX440 vs the Radeon 9000 without the Geforce 4 MX 460 benches being in the review?
I think that the worst problem with your argument is that ATi would be pretty foolish to release the Radeon 9500 performing slower then the Geforce 4 ti 4200. Consider the head room the Radeon 9700 has even when compared with the Geforce 4 ti 4600 you don't think that ATi couldn't find the room in there for its midstream part to hammer nvidias mid stream part? IMO that would be pretty stupid on ATIs behalf. At any rate I think that your conclusion is mis-founded somehow.
Keep in mind that the Radeon 9000 performance differences to the 8500 have a lot less to do with the lack of a 2nd TMU on each pipe that it does with the fact that it it only has 1 vertex shader where the 8500 has two. The 9000 did get a faster triangle setup engine than the 8500 has (from the 9700 again) to help offset the cost of loosing a vertex shader, but it can only help reduce the loss, not replace the lost power.
It isn't just fill-rate that differentiates the 9000 from the 8500, in fact I think in general that has less impact on the performance that some folks suggest. With the 9000s much improved loop-back capability it isn't adding any *passes* to the rendering with its 1 TMU instead of 2, though it does add *cycles*. The 8500 can do 6 textures through 3 cycles, the 9000 needs 6 cycles to do that (inherited from 9700) but that still keeps everything in a single "pass". I think sometimes people don't take into account that TMUs are a lot smarter than they were back in the days of single vs. dual textureing. Back then 2 TMUs was 2 TMUs... you couldn't do multitexturing without 2. That isn't the case with chips today.
The 9000 also generally seems to come closer to its *theoretical* fill-rate numbers than the 8500 ever has, so even though there is a large difference in the look of the theoretical numbers, the difference in fill-rate is less in real world cases than the theoretical numbers suggest.
It isn't just fill-rate that differentiates the 9000 from the 8500, in fact I think in general that has less impact on the performance that some folks suggest. With the 9000s much improved loop-back capability it isn't adding any *passes* to the rendering with its 1 TMU instead of 2, though it does add *cycles*. The 8500 can do 6 textures through 3 cycles, the 9000 needs 6 cycles to do that (inherited from 9700) but that still keeps everything in a single "pass". I think sometimes people don't take into account that TMUs are a lot smarter than they were back in the days of single vs. dual textureing. Back then 2 TMUs was 2 TMUs... you couldn't do multitexturing without 2. That isn't the case with chips today.
The 9000 also generally seems to come closer to its *theoretical* fill-rate numbers than the 8500 ever has, so even though there is a large difference in the look of the theoretical numbers, the difference in fill-rate is less in real world cases than the theoretical numbers suggest.
DaveBaumann said:
I have read it! How can you doubt me so?
The cases where the 8500 outperform the 9000 is those where multitexturing ability is relevant. (Again, where the host is limiting, any halfway decent cards will perform very similar, so those cases just aren't meaningful in terms of video card comparisons, no matter how relevant they are from a total systems perspective.)
I still stand by that in many (most?) fillrate limited cases, the new 4200s may well edge ahead of the 9500. Assuming bla bla bla as above
By judicious choice of benchmarks anything can be proven, obviously, but look at high resolutions in just about any benchmark where the 9000 and the 4200 are compared. Will the 9500 have sufficient enhancements to generally overtake the 4200?
I'm not convinced that an R9500 will be faster than a 4200 in AA. There is little doubt about the 9500 outperforming the 4200 in Anisotropic filtering performance however. Whether either of these cards can be considered fast enough to actually use much IQ enhancing goodies is a matter of personal preference, and the game in question. But trying to generalize, I'd say that these features matter more on the highest end cards which are more likely to have "headroom" in terms of performance.
If priced below the 4200, the 9500 should sell like hotcakes. If priced the same as the 4200, it should sell quite well. If priced significantly above, sales should be sluggish (which, if the "defective chips" hypthesis holds any truth whatsoever, could be desireable until yields improve on the 9700 and they make a dedicated 9500 chip.) Generally though, manufacturers want their products to sell, and ATI is interested in market share as well as immediate profit. To me, it would seem more reasonable to assume that ATI means for the 9500 to eventually take the place of the 9000. Why continue to offer a mid end DX8 part that competes against your own mid end DX9 part, at a small difference in cost? Does it make sense for ATI to offer the 9000, 9000Pro, 9500, (9500 Pro?), 9700, 9700Pro + 7500s and 8500s that are still being cleared + the rumoured new higher end part.
Something is bound to give in that product line, and my prediction would be for it to be the 9000 series being either squeezed down or out. If ATI won't do it directly, the market will do it for them.
Enough of this marketing crap already! Beancounter positioning games, and market presence and perception are way to influential for any rational predictions to be of much use anyway.
Entropy
Ichneumon said:
Ich, I’m not so sure about that. If you look at the review the 3 benches that has most difference are SS:SE, and the two Q3A based titles. Now, SS:SE is heavily multitexture based. I’m not convinved the Q3A titles are even using the vertex shaders (in which case would negate 8500’s apparent advantage) and they will be making much use of multitexturing.
Conversely, we know from 9700 that having plenty of vertex processing ability actually helps a title like Dungeon Siege plenty (look at the difference between 8500 and 9700 – and normally this is CPU limited), however 9000 actually appears to be a little faster than 8500 according to my testing.
IMO, the differences in most cases are purely down to the second TMU.
Entropy said:I have read it! How can you doubt me so?
heheh – I didn’t really!
Entropy said:
Oh (assuming 275MHz + 4 pipes is the case) I’m certain there probably will be cases as well, although it is certainly going to be interesting to see how these two Z buffer optimizations/tricks play against each other as well.
Entropy said:
Well, dependant on resolution and AA depth I am. 4X AA on GF4 Ti 4200’s are totally bandwidth dependant and the second TMU really isn’t doing anything here. The fact that the colour buffer is compressed with R300 makes for a massive difference.
DaveBaumann said:
You should be right there.
A clearer statement of my position would have been that the difference in AA performance should be much smaller between the R9500 and the 4200, than between the R9500 and the R9000, and that I'm not sure that it won't be a close enough call that the impact on percieved value relative to the 4200 will be negligeable. We'll see about that.
Entropy
heheh - shall we try clocking the 9700 core and memory down to 125MHz and see!
Entropy said:
Oh? I think we may see that same architecture recycled again and again much in the same way the MX architecture was. I think that this particular design based on the 8500 architecture may appeal to OEMs much in the same way the MX has as well.
Consider the cheaper 40 million transistor design borrowing from the R300 while retaining the basic R200 core. This appears to be the same sort of design method used by nvidia for its MX series cards. IMO that architecture could very well be around for a while and it will be the OEMs that keep it in the market because of its price/performance much in the same way the MX was. IMO the MX architecture will be the one pushed out of the market.
EDIT: The Radeon 8500 we know is be phased out while living on through the Radeon 9000. The 7500 still has appeal to OEMs to some degree. but that will go after the 8500....
Entropy said:
Why?
You said simply and exclusively because of fillrate (with no further clarification or support) that one graphics archtecture (GF4 Ti 4200) will likely be faster than another different one (Radeon 9500).
I show an example of yet another graphics architecture (Parhelia) that seems to tend to have impressive fillrate, in fact in certain cases clearly exceeding faster versions of those same architectures.
I then point out that even though the Parhelia can perform faster than these faster versions in this criteria you specified was the determing factor for expectations on 9500 performance relative to the 4200, it is outperformed in actual games by a significant margin by these faster versions of these architectures.
I then connected this to the observation that in many of these fill rate tests the faster version of the 9500 (9700 Pro) approaches (and sometimes exceeds) twice the performance of a the faster version of the 4200 (the 4600). This is in relation to my guess as to how you mean the "less than half of the performance of the 9700 Pro" in connection with your statements on fill rate.
To me this is directly relevant to your proposed expectation on the 9500 compared to the 4200, and I wish you wouldn't dismiss this relationship as "irrelevant, both from technical and market perspectives.", with no further comment supporting the assertion, as if that answers anything.
I have a lot of questions in the post you didn't address, and if you have the answers connected to your opinion in this I'd appreciate it if you could supply them.I didn't say you "cared" about the fate of the 9000 prices, I addressed the assertions you made concerning those prices in relation to the 9500 launch and gave my reasoning as to why it seemed flawed to me.
Like the geometry horsepower, anisotropic filtering and FSAA implementations, and "shader" functionality, yes.
Aside from the performance expectations you still haven't seemed to clarify your reasoning for, what is your criteria for "significantly above"? The prices have been hinted at around 179 list for the 9500...is this what you have in mind? I can understand not putting faith in that number, but your argument as I responded to initially seemed to be based on putting faith in SOME price range...was it this one?
You do realize how low the 9000 prices are and that it seems it is cheaper than the 8500 to produce?
Why did you dismiss my Parhelia example? It is this sentiment that it seems to me that the example directly addresses. I still haven't gotten the clarification on this that is responsible for your outlook here.
Is your consideration of the "advanced features" ATi will charge a "premium" for simply as "DX 9 features" and not as features that will enhance the performance of non-DX9 games? That is the only thing that seems consistent (to me, with what you've written so far) with this part of your paragraph, and if it is the case I'd be interested in why that is your conclusion.
In any case, granting the suppositions I don't understand above as true, at worst it seems likely we'll have a card in the 9500 that is as fast as or faster than the Ti 4200 with anisotropic filtering and FSAA turned on. I'd call this worst case competitive, wouldn't you?
The idea that the 9500 will not be significantly faster than the 9000 strikes me as totally ridiculous. Even after losing the second TMU the 9000 is still equal to or faster than the 8500 in some situations, slower in others but still roughly the same speed. Giving up the second TMU in comparison to the Ti4200 isn't going to guarantee the card will be slower. I'd say the R9500 will definitely be faster than the Ti4200 due to better core optimizations and such. Maybe not a lot, but some... Either way I doubt it will be slower than a Ti4200 which, correct me if I'm wrong, will is clocked much lower than 275/275.
Ichneumon said:
I agree and will just add a point or two:
Although a second TMU cuts the numbers of cycles it takes to fetch and apply a texel in half (when doing multitexture), it isn't a given that the architecture is optimized - or even able - to deliever twice the amount of texels. (This is of course even more true when you're doing FSAA and AF or going to use Pixel Shaders with a lot of operations).
I have no idea how many transitors a second TMU in itself demand, but if it is going to be effecient you will probably have to add to the cache and other things to sustain a good texel throughput. Maybe it's about time beyond3d takes a critical view on TMU in the review: Multiple TMU's are soooo 90's!
Oh, and one last thing: Remember that the 9500 should have the vastly improved memory controller from the 9700. It's fairly important to the 9700's success, me thinks.