It has to be more/different than RV670+RV670 on same card, card with 2 chips wouldn't get it's chips their own chip codename
So, do we know anything about RV670 yet?
- Thread starter nicolasb
- Start date
It has to be more/different than RV670+RV670 on same card, card with 2 chips wouldn't get it's chips their own chip codename
vertex_shader
Banned
Because R600 weak as highend GPU so ATi can't make any cutdown version from R600 in the performance segment.
HD2900XT in the reality a performance GPU with highend BOM cost and MSRP price.
Silence big about rv670 in the last few days, anyone still optimistic performance wise after latest G92 rumors (112SP, 600/1500/1800, 229-249$)?
Last edited by a moderator:
So you think R680 will be something different?
Yes R600 is weak, but it confused me with RV name for RV670.
I think it will be different from 2 chips on 1 board ala GX2 - maybe 2 dies on 1 package, maybe 1 huge chip, but at least something different than just 2 chips on 1 board
vertex_shader
Banned
If R670/680 really turns out to be fast, quiet and with stable drivers from launch, I'm quite tempted to buy the thing.
It may be the last oportunity for AMD to turn standalone graphics into profit - they've given up on "highest end" and I fear "high end" would be next and "discrete" after that. After all, a year after merger and launch of the Vista, they are running out of the benefit of the doubt
It may be the last oportunity for AMD to turn standalone graphics into profit - they've given up on "highest end" and I fear "high end" would be next and "discrete" after that. After all, a year after merger and launch of the Vista, they are running out of the benefit of the doubt
If R670/680 really turns out to be fast, quiet and with stable drivers from launch, I'm quite tempted to buy the thing.
It may be the last oportunity for AMD to turn standalone graphics into profit - they've given up on "highest end" and I fear "high end" would be next and "discrete" after that. After all, a year after merger and launch of the Vista, they are running out of the benefit of the doubt
I know many people think this, but I don't think so.
I think R670/R680 (2xRv670) will be the way of the future, in the fact we will see multiple dies, or rather cores working in tandom on the same package, or at least pcb. We've heard R700 may be three or four, this is two.
If ATi somehow finds a way to fit two dies on one package for R670/R680, as I still think they will, who's to say we won't see a 4 die on one package solution on 45nm?
If you think about it that way, with a one chip/core card being the low-end replacement (2400), a two-chip/core card being the value part (2600), a three-chip/core being the performance part (pro/GT), and a full-fledged four-chip/core part being top-of-the-line, it actually makes sense. We're already seeing it with Rv670/G92 on the midrange/highend, all that's left is to fit in the lower-end segments into the same equation.
The hard part, obviously, is designing a chip/core that can scale that well, with performance being there at the bottom, and power consumption/heat/size being controllable at the top; essentially cores that are 40-50W or less, as memory would put them around the 225W marker. I think this is shown as possible with both 2400/2600 and their small size (82/153mm2)/power consumption (35/75W). Now imagine what's possible on 45nm instead of 65nm. Heck, a Rv670 might be darn near the size/power usage of RV610 on 45nm. Imagine the possibilities. A solution that at the top could be 3.x better than RV670, with good performance spacing for low to high-end products (per segment), but with a die that is used for multiple products in essence saying money on yields compared to larger chips.
I think they can figure it out, as RV670 seemed to have been designed with a two-chip solution in mind, that being it can stay within power/heat/size envelopes with two on a package or card (~225W).
Who's to say R600 wasn't designed with the shrink to 55nm, and essentially the dual-chip solution to compliment a mid-range single-chip solution in mind? Sure, R600 may have been a fiasco if you look at it as a single product destined for the high-end, but what if that's the best/most efficient use of R&D they could squeeze out of their tech with the knowing they would be shooting for essentially a shrink that would be consisting of a 55nm dual chip solution, and perhaps even farther ahead to R700 solutions?
This is not to say that is the reality, as it's probably not, but saying AMD/ATi is out of the high-end because of R600 is a slight exaggeration, as they may be shooting from a different angle with long term goals in mind. Or, quite simply, it could be a massive fuck-up of amazing proportions that they are working to rectify.
Or, I suppose, you could be right.
Time will tell, as always...
Last edited by a moderator:
I know this discussion has been held on this board multiple times before, but considering recent data loss maybe it is not out of place to held it yet another time
I can see and understand the move towards independent building blocks and scaling *inside* the chip/package. Take a look at any AMD/NVidia/Intel upcoming, or even current product, especially gfx. The multichip/package solutions, on the other hand, are costly means of providing performance where absolutely needed (multi-socket motherboards, multi-board gfx solutions), as it breaks the cost/performance line(curve) so sharply it's neither funny nor acceptable for consumer space.
I know that AMD's official advice for developers is: (improved) CorssFire is the future, you are strongly suggested to code the support in. But I fail to see the benefits for regular folks, apart from more spread die area which could help the cooling, which is limiting factor of a raw power. Maybe instead of one burning chip and large vent it really is more sensible to have two hot chips and two smaller vents, but I doubt it.
I believe the physical chip is not a really good building block for scaling. Yes, in CPU space we have the situation where manufacturers solder chips together into one package, but in my eyes it's more due to inertia and backwards compatibility than anything: in CPU there really aren't so much independent functional units you could scale (parallelism you could extract from input) without running into unsolvable bottlenecks.
I'm certainly not hardware engineer, but looking at scaling of multi"package" solutions (Crossfire, SLI, multi-CPU) it seems that from purely scaling point of view it is desirable to cut down a die (as done in gfx and CPUs now) than to connect together multipe dies.
I can see and understand the move towards independent building blocks and scaling *inside* the chip/package. Take a look at any AMD/NVidia/Intel upcoming, or even current product, especially gfx. The multichip/package solutions, on the other hand, are costly means of providing performance where absolutely needed (multi-socket motherboards, multi-board gfx solutions), as it breaks the cost/performance line(curve) so sharply it's neither funny nor acceptable for consumer space.
I know that AMD's official advice for developers is: (improved) CorssFire is the future, you are strongly suggested to code the support in. But I fail to see the benefits for regular folks, apart from more spread die area which could help the cooling, which is limiting factor of a raw power. Maybe instead of one burning chip and large vent it really is more sensible to have two hot chips and two smaller vents, but I doubt it.
Why not *inside* the chip then, as now? OK, I believe that 2-chip could really be the solution for "higher end" and multiboard for "ultrahigh end", but 4?in the fact we will see multiple dies, or rather cores working in tandom on the same package, or at least pcb....We've heard R700 may be three or four, this is two.
I believe the physical chip is not a really good building block for scaling. Yes, in CPU space we have the situation where manufacturers solder chips together into one package, but in my eyes it's more due to inertia and backwards compatibility than anything: in CPU there really aren't so much independent functional units you could scale (parallelism you could extract from input) without running into unsolvable bottlenecks.
Agreed
Jawed
Legend
CJ chimes in at R3D. HD3000 series. No more suffixes.
http://www.rage3d.com/board/showpost.php?p=1335186855&postcount=310
http://www.xtremesystems.org/forums/...1&d=1192629284
http://www.rage3d.com/board/showpost.php?p=1335186855&postcount=310
http://www.xtremesystems.org/forums/...1&d=1192629284
Yay for new naming convention! 3850 & 3870 eh? Will be interesting to see if the dual-chip rumors hold up...
Mintmaster
Veteran
That's certainly my guess, especially considering that they put UVD in there. The scaling ATI got is way better than I expected, as we rarely see perfect density gains. Throw in fixed AA resolve (even without dedicated hardware, it never should have been that slow), and ATI is back in the game.Very skinny die for a 55nm R600 redux, indeed!
Some alien tech at TSMC's.
I wonder how many redundant circuities are being thrown out of the original R600 design for this one.
I hope the new low end part (assuming ATI doesn't stick with RV630 in the long haul) doesn't scale as crappily this time. Half of RV630 really should have half the performance.
