Nah, thats BS, IMO. And I don't see any differences from X1600 XT on that board.
RV560/570 Gemini roadmap
- Thread starter kemosabe
- Start date
Dave Baumann said:
It sure looks different than the X1600XT you have in your X1600/X1300 review: http://www.beyond3d.com/reviews/ati/rv5xx/images/b1.jpg
Check upper right corner and I see m ore yellow stuff to the left but that could be for the Rage Theater chip? :S
Last edited by a moderator:
There are a number of reasons why I don't think the X1600 PCB won't be the right fit for RV570, which is also why I think their specs are wrong. I wonder if they are getting confused with an 80nm version of RV530 sampling.
cost of 256-bit memory bus
Dave, do you think you could ask around discreetly for a good estimate of the manufacturers cost of a 256-bit memory system? The reason I ask is that this cost has been used as an argument in countless discussions on these boards without ever being quantified, making all those posts, dating back to the Parhelia and early 9700 rumours pretty pointless.
The thing is, when I do my own back of the envelope calculations, the cost I come up with is quite small, taking into account incremental packaging costs and a PCB with a few additional layers. I might have missed something though.
Now, manufacturers rarely want to speak too loudly about actual hardware costs in any circumstances, since they always look low to uneducated consumers. And in this case, 256-bit buses are also used as one of the features that mark the divide between the $200 and down, and the $300 and up cards coupled of course with substantially more complex and expensive GPUs, so manufacturers may be hesitant to state an estimate for the 256-bit bus in and of itself since it is part of the more exclusive package.
Nevertheless several years of pretty baseless arguing is a bit pathetic, and since it seems perfectly possible to pin a ballpark figure on the cost of a 256-bit bus, it would really be a boon if you (or someone with personal industry knowledge) could put these discussions on a firmer financial footing.
Dave, do you think you could ask around discreetly for a good estimate of the manufacturers cost of a 256-bit memory system? The reason I ask is that this cost has been used as an argument in countless discussions on these boards without ever being quantified, making all those posts, dating back to the Parhelia and early 9700 rumours pretty pointless.
The thing is, when I do my own back of the envelope calculations, the cost I come up with is quite small, taking into account incremental packaging costs and a PCB with a few additional layers. I might have missed something though.
Now, manufacturers rarely want to speak too loudly about actual hardware costs in any circumstances, since they always look low to uneducated consumers. And in this case, 256-bit buses are also used as one of the features that mark the divide between the $200 and down, and the $300 and up cards coupled of course with substantially more complex and expensive GPUs, so manufacturers may be hesitant to state an estimate for the 256-bit bus in and of itself since it is part of the more exclusive package.
Nevertheless several years of pretty baseless arguing is a bit pathetic, and since it seems perfectly possible to pin a ballpark figure on the cost of a 256-bit bus, it would really be a boon if you (or someone with personal industry knowledge) could put these discussions on a firmer financial footing.
How do you define the "cost of a 256-bit bus"?
The primary cost of a 256-bit bus is the physical die size required to have enough pad space to support it, so that "cost" is intrinsically linked to the size, hence performance targets of the rest of the chip - i.e. they probably are never going to "fit" a 256-bit bus to an RV530 configuration chip as a bunch of "wasted" die would be required to facilitate the size of the bus, alternatively, if you have a chip thats big enough to support it, its likely that these days you'll include it. So, generally, I'm not sure you can purely pin a "cost" to tbe bus itself as it has to be linked to the die size.
There are of course other costs associated, such as packaging, components, PCB etc. The costs here aren't just limited to the component costs, but also fabrication time costs. For instance, IIRC I was told that a a single machine can manufacture about a 1000 high end 8/10 layer boards per month - the lower the bus size, the few the layers, the more boards can be produced by a single machine, etc.
The primary cost of a 256-bit bus is the physical die size required to have enough pad space to support it, so that "cost" is intrinsically linked to the size, hence performance targets of the rest of the chip - i.e. they probably are never going to "fit" a 256-bit bus to an RV530 configuration chip as a bunch of "wasted" die would be required to facilitate the size of the bus, alternatively, if you have a chip thats big enough to support it, its likely that these days you'll include it. So, generally, I'm not sure you can purely pin a "cost" to tbe bus itself as it has to be linked to the die size.
There are of course other costs associated, such as packaging, components, PCB etc. The costs here aren't just limited to the component costs, but also fabrication time costs. For instance, IIRC I was told that a a single machine can manufacture about a 1000 high end 8/10 layer boards per month - the lower the bus size, the few the layers, the more boards can be produced by a single machine, etc.
I've always figured it was down to die size. Even the engineering skull work has already been done for up-the-scale parts, so there's no groundbreaking going on, or terra incognita risk.
I'll be interested to see what that die size is re RV570 tho. It's liable to be interesting either way, in either ATI's upper midrange part will be the size of NV's flagship. . .or it will break new ground in our understanding of how small a die size can support 256-bit. So from technogeek pov, interesting either way.
I'll be interested to see what that die size is re RV570 tho. It's liable to be interesting either way, in either ATI's upper midrange part will be the size of NV's flagship. . .or it will break new ground in our understanding of how small a die size can support 256-bit. So from technogeek pov, interesting either way.
Entropy,
Take another approach: do IHVs really need >128bit busses in the mainstream segment? What I mean is - final cost included or not - usually mainstream products are more or less half the performance of their high end brothers within the same product family.
Does anyone expect adequate 4xAA performance in high resolutions on a ~$200-250 mainstream GPU? Even if what good would the added bandwidth of a hypothetical 256bit bus there really do if you consider multitexturing fillrates, ALU performance and what not?
Example:
7900GTX:
15.6 GTexels/sec
10.4 GPixels/sec
51.2 GB/s
250 GFLOPs/s
512MB ram
7600GT:
6.72 GTexels/sec
4.48 GPixels/sec
22.4 GB/s
108 GFLOPs/s
256MB ram
While most numbers are =/< 1/2 of a high end GPU (final street price included), why would a 256bit bus make sense in that segment when the added bandwidth wouldn't make such a big difference after all because it would bottleneck elsewhere. So yes under that reasoning and taking ALL factors under account I can see a 256bit bus here adding only in cost.
Take another approach: do IHVs really need >128bit busses in the mainstream segment? What I mean is - final cost included or not - usually mainstream products are more or less half the performance of their high end brothers within the same product family.
Does anyone expect adequate 4xAA performance in high resolutions on a ~$200-250 mainstream GPU? Even if what good would the added bandwidth of a hypothetical 256bit bus there really do if you consider multitexturing fillrates, ALU performance and what not?
Example:
7900GTX:
15.6 GTexels/sec
10.4 GPixels/sec
51.2 GB/s
250 GFLOPs/s
512MB ram
7600GT:
6.72 GTexels/sec
4.48 GPixels/sec
22.4 GB/s
108 GFLOPs/s
256MB ram
While most numbers are =/< 1/2 of a high end GPU (final street price included), why would a 256bit bus make sense in that segment when the added bandwidth wouldn't make such a big difference after all because it would bottleneck elsewhere. So yes under that reasoning and taking ALL factors under account I can see a 256bit bus here adding only in cost.
You potentially save on memory costs tho. Anyone have a feel for what that amounts to? I don't think anyone is assuming that we're going to get top-of-the-line memory on these boards. It seems to me the potential is there to get more bandwidth for cheaper than before (but not X1900XTX or 7800GTX512-class bandwidth --I mean compared to, say, 7600GT or X1600XT).
geo said:
What do you mean exactly? What I meant in my former post is that for the performance threshold G73 is targetted for, it can't really take a significant benefit out of more bandwidth or let's say a 512MB framebuffer.
The 7600GT has 700MHz GDDR3(128bits) while the 7900GT has 600MHz GDDR3(256bits).
True. However, the top-of-the-line card may be severely constrained by its memory system in some cases. If so, giving a middle range card a higher bandwidth to processing power ratio than the top offering might be useful.Ailuros said:
It is interesting to compare the x1600xt with the x1900xt. The x1600xt can be described as one fourth of the x1900xt, yet, in many applications, it performs at a level that almost perfectly matches its bandwidth ratio to its higher end brethren, rather than its ALU processing capabilities, implying that in many real life applications the x1900xt is bandwidth constrained rather than constrained by its internal processing capabilities (or constrained by some other factor that happens to match the bandwidth differences).
Bringing us back to the subject of this thread, for a hypothetical RV5X0 GPU that has twice the processing capabilities of the RV530, this would imply that it would stand to benefit quite substantionally from having a correspondingly faster memory subsystem, rather than adopting the ratio of the x1900xt.
The x1600xt vs x1900xt performance characteristics is the best data point I can see in the market at this point, particularly when cross referenced with x1800xt to x1900xt data.
While clearly true to some extent this brings up another unquantified issue in these discussions. Just how significant is this pad space limitation at this point in time? Looking at past history, I'm not sure it holds water given the size of todays mid-range GPUs, and when you look at the I/O capabilities vs. die area of other than GPUs - well, I'm not sure that an argument that may or may not have been valid back in R300 vs RV350 discussions are valid in todays GPU marketplace. While there is obviously a chip cost associated with driving these external connections, it is not clear that a chip like the RV530 would have been somehow incapable of fitting a 256-bit interface without incurring disproportionate costs.Dave Baumann said:
Like Dave said, pinning a hard cost on these issues is difficult. If it hadn't been I would have done it myself. But looking at it from the outside, I still feel that the products we see offered are due to market and product differentiation concerns as much as technical ones. Actually I hope that's the case, because then 256-bit interfaces in the mid-range is a card the IHVs can play if the competitive landscape makes it justifiable.
And consumers might then enjoy the benefits that Ailuros says they shouldn't expect.
Last edited by a moderator:
Ailuros said:
So, theoretically, you could put 350mhz @256bit on a 7600GT-class card and have the same bandwidth, right? There's got to a savings involved in that, tho I imagine memory prices are more asymptotic in the way they rise (i.e. the cost difference between 300mhz and 400mhz memory is much less than between 600mhz and 700mhz memory). But what you just pointed at there is that 7600GT memory is more expensive per-unit than 7900GT memory, and that can't be something you'd enjoy as an IHV or AIB.
But I'd also wind back to what Wavey said as being important --if you're already there on die-size anyway for other reasons, is probably when 256-bit makes sense. Apparently ATI got to that point, and if we want to criticize them for that underlying fact is one conversation. . .but having got there anyway, the wisdom of having decided to leverage it into 256-bit, it seems to me, is a different conversation. Maybe a "lemons and lemonade" kind of thing.
Entropy said:
While transitors shrink, allowing for more transistors for a given die area, the pads don't shrink (or at least, move at anything like the rate transistors can, if at all), so the number of pads is a function of the die size, irrespective of the process. Take a look at chips such as RV410, RV530, NV43 and G73 (the upper bounds of 128-bit performance) and you'll see they are all the same; NV42 and R430 are the lower bounds of 256-bit chippery, and there is a reasonable jump in size to them - these sizes are likely to be firstly fixed at around these levels.
I went through this with Uttar on IRC not long ago - bear in mind that the memory costs is a juggle with the framebuffer sizes and witdth you want as well.geo said:
For instance, if there was a hypothetical 256-bit 7600 GT bear in mind that you would actually need 8 256Mb 350MHz chips to get the same size and bandwidth as 4 512Mb 700MHz chips on the 128-bit version as each chip is 32-bit wide (also, the 256-bit solution would pobably be lower performance in this case as well, because more data would be required per channel). So, in this case, which would actually end up cheaper? I don't know, because I don't have the figures (and they are likely to fluctuate) but its another factor to consider.
trumphsiao
Regular
geo said:Yes, we do need a memory broker on staff at times. Get on that, will ya?
GDDR4 1.2GHz at 128bit Bus
Well, I think it´s pretty fair to assume that RV570 will "easily" be larger or at least about the same size as R430 for more reasons than just blind guessing. Both cores are in their newly created "performance"-segment, both cores (are) should be pretty large and rather complex. Adding something to that notion: There could´ve been a change of plans, considering ATis lower than expected yields and "technology issues" on 80nm, so if there is any truth about that digitimes-story posted some pages back, 256-bit is entirely feasible, pad size shouldn´t be a noteworthy limitation then.
The only thing really bugging me is the codename itself: RV570. Why RV and not R. If they haven´t thrown their nomenclature overboard over night, there has to be something that is gonna save them costs, while at the same time performance is very good.
GDDR4 is certainly doable by Q2/06, but like geo just mentioned, those starting costs are a factor of uncertainty. I´m very skeptic about an introduction this early in the game. (Samsung will use their 80nm process for GDDR4, but i haven´t heard or read anything about yields, yet.)
The only thing really bugging me is the codename itself: RV570. Why RV and not R. If they haven´t thrown their nomenclature overboard over night, there has to be something that is gonna save them costs, while at the same time performance is very good.
GDDR4 is certainly doable by Q2/06, but like geo just mentioned, those starting costs are a factor of uncertainty. I´m very skeptic about an introduction this early in the game. (Samsung will use their 80nm process for GDDR4, but i haven´t heard or read anything about yields, yet.)
Last edited by a moderator: