Updated 3dfx Rampage Specs!
- Thread starter utopia
- Start date
Anyone think 3dfx could come through on a product like this?
Maybe they will come out on top again.
Maybe they will come out on top again.
We'll NEVER see the 8 chip version.
Lemme get this straight:
4 .18 micron chips and 128mb of 225Mhz DDR RAM on a midrange card? I'm sure that'll have a low low MSRP of $4-500. Get real. The NV20 is gonna be single chip in likely hood, and probably competitive with the midrange version. nVidia needs to make 1/4 as many chips and only needs half as much memory, it just doesn't make sense.
That's just what it looks like on the surface though, 3dfx is way ahead of nVidia and has figured out how to avoid any future supply problems. If you make you're cards expensive enough to produce, the demand will plumet, and will easily be met by the available chips.
Brilliant, absolutely brilliant.
The fact that the T&L chip is seperate means to me that the rendering chip is probably a derivitave off the vsa100, which was a souped up version of the voodoo3, which was the banshee with multitexturing, which was a 2d/3d crippled voodoo2, which was the voodoo 1 with multitexturing.
Lemme get this straight:
4 .18 micron chips and 128mb of 225Mhz DDR RAM on a midrange card? I'm sure that'll have a low low MSRP of $4-500. Get real. The NV20 is gonna be single chip in likely hood, and probably competitive with the midrange version. nVidia needs to make 1/4 as many chips and only needs half as much memory, it just doesn't make sense.
That's just what it looks like on the surface though, 3dfx is way ahead of nVidia and has figured out how to avoid any future supply problems. If you make you're cards expensive enough to produce, the demand will plumet, and will easily be met by the available chips.
Brilliant, absolutely brilliant.
The fact that the T&L chip is seperate means to me that the rendering chip is probably a derivitave off the vsa100, which was a souped up version of the voodoo3, which was the banshee with multitexturing, which was a 2d/3d crippled voodoo2, which was the voodoo 1 with multitexturing.
yeah they better have bandwidth saving measures, as I've said since the NV15 and Voodoo 5 came out.
"The fact that the T&L chip is seperate means to me that the rendering chip is probably a derivitave off the vsa100, which was a souped up version of the voodoo3, which was the banshee with multitexturing, which was a 2d/3d crippled voodoo2, which was the voodoo 1 with multitexturing."
you seem to be missing the fact that the RAMPAGE IS a new architecture that of course includes everything good about the Voodoo architecture. IE the upgradeability by using multiple chips.
It appears 3dfx is counting on DDR SDRAM, and multiple chips to get around the bandwidth issue. the DDR SDRAM will surely help them a fair amount, but having something like ATi's Z-Buffer compression would help them out a lot.
In any case, it appears to me that the individual RAMPAGE chip is pretty much a NV15 chip without T&L built in. (ie 4 pipes, 200mhz, no word on extra texturing units, though I think one extra would be nice... T&L would seemingly make anything more then dual texturing obsolete, and even dual texturing wouldn't be that usefull)
Still this is all rumors... If it is true, then the dual Rampage seemingly would be the one to get. becuase the single one would be about the same as the NV15.
"The fact that the T&L chip is seperate means to me that the rendering chip is probably a derivitave off the vsa100, which was a souped up version of the voodoo3, which was the banshee with multitexturing, which was a 2d/3d crippled voodoo2, which was the voodoo 1 with multitexturing."
you seem to be missing the fact that the RAMPAGE IS a new architecture that of course includes everything good about the Voodoo architecture. IE the upgradeability by using multiple chips.
It appears 3dfx is counting on DDR SDRAM, and multiple chips to get around the bandwidth issue. the DDR SDRAM will surely help them a fair amount, but having something like ATi's Z-Buffer compression would help them out a lot.
In any case, it appears to me that the individual RAMPAGE chip is pretty much a NV15 chip without T&L built in. (ie 4 pipes, 200mhz, no word on extra texturing units, though I think one extra would be nice... T&L would seemingly make anything more then dual texturing obsolete, and even dual texturing wouldn't be that usefull)
Still this is all rumors... If it is true, then the dual Rampage seemingly would be the one to get. becuase the single one would be about the same as the NV15.
225 MHz DDR....ooooo wow. That's not terribly impressive. That's only 7.2 GB/s. They didn't mention the memory bus width so I'm assuming it's still 128 bit. It's going to be bandwidth limited. I won't be impressed until a card comes out with over 10 GB/s.
And 256 MB of RAM? How much do they expect us to pay? The V5 6000 is going at $600 with 128 MB. What will a card with 256 MB and 8 chips cost us? Get it together 3dfx.
Will it be fast? Hell yes. Will it be worth it? Hell no.
And 256 MB of RAM? How much do they expect us to pay? The V5 6000 is going at $600 with 128 MB. What will a card with 256 MB and 8 chips cost us? Get it together 3dfx.
Will it be fast? Hell yes. Will it be worth it? Hell no.
Um, 4 x SLI = 4 x 7.2GB/s = 28.8 GB/s
You pay for what you get. If 3dfx will have this speed, so will nVidia, but only in a single-chip config, so only 7.2GB/s. A dual-SLI Rampage will still be double anything nVidia can throw out. And you'd think 3dfx has developed some sort of HyperZ of their own. I mean two years, sheesh...
You pay for what you get. If 3dfx will have this speed, so will nVidia, but only in a single-chip config, so only 7.2GB/s. A dual-SLI Rampage will still be double anything nVidia can throw out. And you'd think 3dfx has developed some sort of HyperZ of their own. I mean two years, sheesh...
All I know is there is no way I will be able to afford that card. My wife would kill me. You spent how much for what?
Those specs are fabricated, just like the ones from a little while ago. I don't have any info, but what has been posted is simply unreasonable. 8 chips? 256MB DDRSDRAM? That isn't believable at all.
My guess is that there will once again be one, two, and four chip versions. T&L could come in the form of a seperate chip, but due to all this talk about modularity, I'm beginning to suspect it may just be a part of the chip itself. This is all speculation, though.
When we come to the pixel processor, however, things start getting sketchy. We know its a completely new architecture, with absolutely no connections to the Voodoo line. We've heard rumblings about allocatable texturing units. Basically, what it comes down to is the fact that we don't really know any specifics. Or at least I don't. But, I can make what I think are educated guesses of its performance relative to today's architectures. Here they are:
As for the T&L unit... We know it's called S.A.G.E., which most interperate to mean Scalable Architecture - Geometry Engine. Once again, the details are fuzzy - it could be a module on the Rampage chip itself, or a seperate chip. Since it has been said that the new chip will be produced on the .15/.18 process, my guess is that S.A.G.E. is just a module on the chip, unless both the Rampage chip and the geometry part are very large. All indications up till now, however, are that it will offer sowhere near 75 Mtriangles/sec.
These are, I think, good estimates of Rampage's performance. The specs I wrote are those that are relevant in todays games. Those specs relevant in the future are not known. However, indications are that there will be extensive support for features never before seen in consumer products. I came to this conclusion reading the statements of a 3dfx employee who said the company will focus on OpenGL in the future, because it is the API that will allow developers to expose new hardware features "at will." Take from that what you will. Apparently not even DX8 will support all of Rampage's features...
Marty
My guess is that there will once again be one, two, and four chip versions. T&L could come in the form of a seperate chip, but due to all this talk about modularity, I'm beginning to suspect it may just be a part of the chip itself. This is all speculation, though.
When we come to the pixel processor, however, things start getting sketchy. We know its a completely new architecture, with absolutely no connections to the Voodoo line. We've heard rumblings about allocatable texturing units. Basically, what it comes down to is the fact that we don't really know any specifics. Or at least I don't. But, I can make what I think are educated guesses of its performance relative to today's architectures. Here they are:
- single chip, .18 micron process
- 250MHz clock
- 4 pixel pipelines, 1 texel/pixel
- support for ddr memory
As for the T&L unit... We know it's called S.A.G.E., which most interperate to mean Scalable Architecture - Geometry Engine. Once again, the details are fuzzy - it could be a module on the Rampage chip itself, or a seperate chip. Since it has been said that the new chip will be produced on the .15/.18 process, my guess is that S.A.G.E. is just a module on the chip, unless both the Rampage chip and the geometry part are very large. All indications up till now, however, are that it will offer sowhere near 75 Mtriangles/sec.
These are, I think, good estimates of Rampage's performance. The specs I wrote are those that are relevant in todays games. Those specs relevant in the future are not known. However, indications are that there will be extensive support for features never before seen in consumer products. I came to this conclusion reading the statements of a 3dfx employee who said the company will focus on OpenGL in the future, because it is the API that will allow developers to expose new hardware features "at will." Take from that what you will. Apparently not even DX8 will support all of Rampage's features...
Marty
- Mar 20, 2000
- 102,402
- 8,574
- 126
also, 3dfx just bought gigapixel... if one of the companies has a good batch of engineers we just might see tiling in the rampage chip. might push it back a little, but if they can still release it in jan it will be great. assume, as always, double the previous' product's processing power, and then the boost a chip gets from tiles, thats a lot of useable fill rate. 3dfx doesn't have to have a part out to compete against nv20, since people are always buying graphics cards, not simply the day that nvidia releases a part.
These previews are really starting to get me concerned over the performance of the next generation parts. Memory bandwidth is the limiting factor in most current generation video cards - Radeon being the exception. The particular idea that ATI has pushed forward is that 300MPixel/s fillrate requires 200MHz DDR with 32 bit everything. Judging from the performance of the Radeon relative to the GTS and V5 I'd have to say they seem to be correct. Assuming that the relationship between fillrate and memory bandwidth is linear we can get some disturbing ideas that the next generation parts are going to be bandwidth limited unless some serious bandwidth savings are implemented, a la the Kyro.
Going by ATIs numbers we can say the factor to convert from MPixels/s to MB/s is approximately 20. 320MPixel*x=6400MB/s.
Now, looking at the bandwidth provided by 128 bit memory running at 225MHz DDR, 7.2GB/s. This corresponds to a maximum fillrate of only 360MPixels/s! Even using 256 bit DDR at 225MHzthe 14.4GB/s only corresponds to a fillrate of 720MPixels/s. Both of those numbers are well under the 1GPixel/s that the chip can perform.
I've got a feeling that this next generation will be won by saving the most bandwidth, not by the most advanced core.
Going by ATIs numbers we can say the factor to convert from MPixels/s to MB/s is approximately 20. 320MPixel*x=6400MB/s.
Now, looking at the bandwidth provided by 128 bit memory running at 225MHz DDR, 7.2GB/s. This corresponds to a maximum fillrate of only 360MPixels/s! Even using 256 bit DDR at 225MHzthe 14.4GB/s only corresponds to a fillrate of 720MPixels/s. Both of those numbers are well under the 1GPixel/s that the chip can perform.
I've got a feeling that this next generation will be won by saving the most bandwidth, not by the most advanced core.
DOH!! I forgot about the bandwidth increasing with multiple processors, sorry. And with eight chips that would be 57.6 GB/s, not bad at all.
But still, how much is it going to cost?
And with eight chips that would be 57.6 GB/s, not bad at all. But still, how much is it going to cost?
no 8 chips I think they were talking about 4 normal ones and 4 for T&L. which is 4X the bandwidth that 200mhz DDR SDRAM can give u.
for me, it seems that the specs are going to be at 200mhz/200mhz core/mem, with DDR SDRAM, and 128 bit interface. that means the same bandwidth as the Geforce 2 GTS, except that it's possible they might do some sort of memory bandwidth reducing trick (probably not tile rendering, I think that's going into the video card that Gigapixel will help in, possibly a year after? we will see.)
So, if these specs are true, each chip will be a little bit more powerful then todays Geforce GTS (core 200, mem 166) becuase it has more bandwidth. having a dual chip config like that would be killer IMHO.
the NV20 IMHO again is going to be an NV15 with tile based rendering to take out much of the bottleneck. this is going to be what the NV15 should have been.
As you can see, the NV15 (and probably the NV20) will have 4 pipes with 2 texture units per pipe, yeilding 1.6 gigapixels at 200mhz. (it's quite possible they'll be increasing the mhz).
with 3dfx's solution, they'll have one low end part to beat out the NV15, it will compete well with the NV20. then you will have the dual chip Rampage, which will be quite interesting to see against the NV20. If the quad chip (8 chips including sage, the T&L unit) ever comes out, it will be one big mofo..
for me, it seems that the specs are going to be at 200mhz/200mhz core/mem, with DDR SDRAM, and 128 bit interface. that means the same bandwidth as the Geforce 2 GTS, except that it's possible they might do some sort of memory bandwidth reducing trick (probably not tile rendering, I think that's going into the video card that Gigapixel will help in, possibly a year after? we will see.)
So, if these specs are true, each chip will be a little bit more powerful then todays Geforce GTS (core 200, mem 166) becuase it has more bandwidth. having a dual chip config like that would be killer IMHO.
the NV20 IMHO again is going to be an NV15 with tile based rendering to take out much of the bottleneck. this is going to be what the NV15 should have been.
As you can see, the NV15 (and probably the NV20) will have 4 pipes with 2 texture units per pipe, yeilding 1.6 gigapixels at 200mhz. (it's quite possible they'll be increasing the mhz).
with 3dfx's solution, they'll have one low end part to beat out the NV15, it will compete well with the NV20. then you will have the dual chip Rampage, which will be quite interesting to see against the NV20. If the quad chip (8 chips including sage, the T&L unit) ever comes out, it will be one big mofo..
256mbs of DDR ram alone will cost more than the NV20. This thing will kick ass in games though. The fact that no one can possilby afford it might hurt sales.
The specs seem to outrageous to be real. 8 chips!!!! Come on does anyone need that kinda power. In 6 months an even faster card one 1 chip will come out that will put that to shame. Its not worth buying and its not worth making.
The specs seem to outrageous to be real. 8 chips!!!! Come on does anyone need that kinda power. In 6 months an even faster card one 1 chip will come out that will put that to shame. Its not worth buying and its not worth making.
soccerman,
Yeah 4 processors and 4 T&L chips makes more sense. You would have 28.8 GB/s of bandwidth with the specs that utopis provided, so I don't think we have to worry about it being bandwidth limited.
One question though, how powerful will the T&L be? Isn't the reason that the GeForce's T&L is so good is because it is built into the GPU and not on a seperate chip? I know the 3DLabs Oxygen GVX1 has a seperate T&L chip and it sucks compared to the GeForce.
Yeah 4 processors and 4 T&L chips makes more sense. You would have 28.8 GB/s of bandwidth with the specs that utopis provided, so I don't think we have to worry about it being bandwidth limited.
One question though, how powerful will the T&L be? Isn't the reason that the GeForce's T&L is so good is because it is built into the GPU and not on a seperate chip? I know the 3DLabs Oxygen GVX1 has a seperate T&L chip and it sucks compared to the GeForce.
You would have 28.8 GB/s of bandwidth with the specs that utopis provided, so I don't think we have to worry about it being bandwidth limited.
I've got to disagree. With no bandwidth saving techniques this amount of bandwidth is nowhere near where it needs to be to support a 4 GPixel/s fillrate. Once again, going by what ATI claims 28.8 GB/s would correspond to approximately 1.44 GPixel/s.
I've got to disagree. With no bandwidth saving techniques this amount of bandwidth is nowhere near where it needs to be to support a 4 GPixel/s fillrate. Once again, going by what ATI claims 28.8 GB/s would correspond to approximately 1.44 GPixel/s.
(300Mpixels/sec) x (4bytes/pixel) = (1200Mbytes/sec)
The amount of bandwidth neccesary to draw 300Mpixels/sec.
(6400Mbytes/sec) - (1200Mbytes/sec) = (5200Mbytes/sec)
Assuming 200MHz DDR memory, the amount of bandwidth remaining after taking into account the amount needed for actually drawing the pixels.
(4buffers/frame) x (1000000pixels/buffer) x (4bytes/pixel) = (16Mbytes/frame)
The amount of data needed for every frame that is drawn, assuming a resolution of 1000x1000, or 1 million pixels.
(5200Mbytes/sec) / (16Mbytes/frame) = (325frames/sec)
The amount of frames that can theoretically be drawn if all remaining bandwidth is used up. Actually, in this case, there is an excess of bandwidth available, since 300 Mpixels/sec at a resolution of 1000x1000 is only enough for 300fps. Conclusion: ATI's numbers are wrong.
Marty
The amount of bandwidth neccesary to draw 300Mpixels/sec.
(6400Mbytes/sec) - (1200Mbytes/sec) = (5200Mbytes/sec)
Assuming 200MHz DDR memory, the amount of bandwidth remaining after taking into account the amount needed for actually drawing the pixels.
(4buffers/frame) x (1000000pixels/buffer) x (4bytes/pixel) = (16Mbytes/frame)
The amount of data needed for every frame that is drawn, assuming a resolution of 1000x1000, or 1 million pixels.
(5200Mbytes/sec) / (16Mbytes/frame) = (325frames/sec)
The amount of frames that can theoretically be drawn if all remaining bandwidth is used up. Actually, in this case, there is an excess of bandwidth available, since 300 Mpixels/sec at a resolution of 1000x1000 is only enough for 300fps. Conclusion: ATI's numbers are wrong.
Marty
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 22K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
News NVIDIA and Intel to Develop AI Infrastructure and Personal Computing Products- Started by poke01
- Replies: 407
-
Facebook
Twitter