Next Generation nvidia GPU details revealed

[vmsein]

http://www.dailytech.com/article.aspx?newsid=11842

NVIDIA's upcoming Summer 2008 lineup gets some additional details

Later this week NVIDIA will enact an embargo on its upcoming next-generation graphics core, codenamed D10U. The launch schedule of this processor, verified by DailyTech, claims the GPU will make its debut as two separate graphics cards, currently named GeForce GTX 280 (D10U-30) and GeForce GTX 260 (D10U-20).

The GTX 280 enables all features of the D10U processor; the GTX 260 version will consist of a significantly cut-down version of the same GPU. The D10U-30 will enable all 240 unified stream processors designed into the processor. NVIDIA documentation claims these second-generation unified shaders perform 50 percent better than the shaders found on the D9 cards released earlier this year.

The main difference between the two new GeForce GTX variants revolves around the number of shaders and memory bus width. Most importantly, NVIDIA disables 48 stream processors on the GTX 260. GTX 280 ships with a 512-bit memory bus capable of supporting 1GB GDDR3 memory; the GTX 260 alternative has a 448-bit bus with support for 896MB.

GTX 280 and 260 add virtually all of the same features as GeForce 9800GTX: PCIe 2.0, OpenGL 2.1, SLI and PureVideoHD. The company also claims both cards will support two SLI-risers for 3-way SLI support.

Unlike the upcoming AMD Radeon 4000 series, currently scheduled to launch in early June, the D10U chipset does not support DirectX extentions above 10.0. Next-generation Radeon will also ship with GDDR5 while the June GeForce refresh is confined to just GDDR3.

The GTX series is NVIDIA's first attempt at incorporating the PhysX stream engine into the D10U shader engine. The press decks currently do not shed a lot of information on this support, and the company will likely not elaborate on this before the June 18 launch date.

After NVIDIA purchased PhysX developer AGEIA in February 2008, the company announced all CUDA-enabled processors would support PhysX. NVIDIA has not delivered on this promise yet, though D10U will support CUDA, and therefore PhysX, right out of the gate.

NVIDIA's documentation does not list an estimated street price for the new cards.


--------------------------------------------------------------------------------

[takamishanoku]

Wow. Looks like we will finally get a monster card. 512bit check. 240SP (supposively 50% faster so approx 480SP if one a 8/9 series?) check. 1 GB mem. Check. Looks like Crysis has met it's match.
Finally a new card and what a beast.

[RPGWiZaRD]

I bet $450~500 for GTX 260 and $550~600 for GTX 280 at launch depending on brand+bundle+oc or stock etc. At least judged by the specs NVIDIA could price it like that... going higher would be absurd though even if it would be very fast.

[Shintai]

Why use 512bit when you can go GDDR5. On the other hand 512bit and GDDR5 would be quite insane. Too insane.

[Mk]

i bet ~600$

[Hornet331]

Why use 512bit when you can go GDDR5. On the other hand 512bit and GDDR5 would be quite insane. Too insane.

imho nv only has gone 512bt, cause they couldn't get hold of enough gddr5. ATI/AMD snatched all the supply away, same as with gddr4.

[stayfrosty]

The main difference between the two new GeForce GTX variants revolves around the number of shaders and memory bus width. Most importantly, NVIDIA disables 48 stream processors on the GTX 260. GTX 280 ships with a 512-bit memory bus capable of supporting 1GB GDDR3 memory; the GTX 260 alternative has a 448-bit bus with support for 896MB.

Does this mean that we can just unlock those 48 stream processors that they've "disabled"

side note: Why does nvidia REFUSE to make dx 10.1 compatible video cards?

[w0mbat]

No, we cant.

[Shintai]

Does this mean that we can just unlock those 48 stream processors that they've "disabled"

No, its along time since you could enable something. Its all lasercut.

[Hornet331]

Does this mean that we can just unlock those 48 stream processors that they've "disabled"

yeah, the same way you could unlock the 112 Sp 8800GT to a 128 SP 8800GTS...

[RPGWiZaRD]

Seriously why nitpick 512bit bus, it's the logical next move what we wanted, price/performance ratio won't suffer that much if NVIDIA's sacrifising a bit profit per card for sales numbers and GDDR3 availability is good so go figure. This is just the finally significant bump in performance and that's how it will be priced after. GDDR5 + 256bit bus isn't a bad choice either but NVIDIA's decision doesn't seem any worse to me, pretty much equal, difficult to say when I don't have any numbers from GDDR3 and 5 industry but remember the G92 fiasco, I bet NVIDIA don't wanna risk the same happening with GT200 if GDDR5 availability would be rather limited at this point still and since ATI will use GDDR5 it would be even worse.

[cantankerous]

Wondering, would Physx on these new cards be just as fast as having the actual phsyx card in your system in hardware form now? I know it may be impossible to answer for sure but a theoretical answer is good enough for me. I am curious as to whether this will be pure hardware based physx or simply a software based 'emulation' using the shaders available on the GPU itself. If it truly is comparable than I can yank my hardware based physx card out of my system without having of a fear of losing anything (which isn't much).

[cantankerous]

I would rather have GDDR3 in a faster bus than have it the other way around due to much higher latencies on GDDR5 than that of aggressively binned GDDR3.

[kromosto]

512bit is good at least they can go gddr4 or 5 with 512bit if bandwidth is not enough.

[Shintai]

I would rather have GDDR3 in a faster bus than have it the other way around due to much higher latencies on GDDR5 than that of aggressively binned GDDR3.

GPUs dont care about latency. And you already got some CL17-CL19 on a 9800GTX.

[Hornet331]

gddr5 + 512 bit is insane... noone needs that much bandwidth... with 4ghz gddr5 it would be 250gb/s+....

[Calmatory]

I would rather have GDDR3 in a faster bus than have it the other way around due to much higher latencies on GDDR5 than that of aggressively binned GDDR3.

Aggressively as in CAS latency? Which one is faster, CAS 2 and 235MHz or CAS 8 and 1050 MHz? Yes, many would say that the first ones have better latency and are tighter, but in reality the latter has lower latency and is "tighter" as in CAS latency.

...and you want to pay more for the chip due to 512-bit bus?

...the reason they don't want to go 512-bit is that the chip will be more complex and bigger, thus making them expensive to produce. Sure we have had our 256-bit busses for like a decade, but the RAM chips are getting faster and faster. remember the das of 100 MHz SDRAM cards? Compare them to 4 GHz GDDR5 cards. If the bus width stays same, the bandwidth is being increased by 4000 %.

And the 4 GHz is just the beginning.

[bowman]

Does this mean that we can just unlock those 48 stream processors that they've "disabled"

side note: Why does nvidia REFUSE to make dx 10.1 compatible video cards?

No, they're dead. Same as the cache in a Celeron or the fourth core in a Phenom X3.

[Calmatory]

No, they're dead. Same as the cache in a Celeron or the fourth core in a Phenom X3.

Not always it is dead. Remember the 6200 and 6800 cards for example? Or Applebred durons?

[nn_step]

Why use 512bit when you can go GDDR5. On the other hand 512bit and GDDR5 would be quite insane. Too insane.

Too insane and setting benchmarks records are not mutually exclusive...

[DeathReborn]

imho nv only has gone 512bt, cause they couldn't get hold of enough gddr5. ATI/AMD snatched all the supply away, same as with gddr4.

Yea, nothing at all to do with very few companies actually following through with mass production & timing issues with GDDR4 in comparison to GDDR3/5. nVidia may go GDDR5 in a future refresh but right now there's no need for it, unless you stick to 256bit.

The only thing that's been a bigger flop then GDDR4 is RAMBUS, or is that the other way around...

[perkam]

Why use 512bit when you can go GDDR5. On the other hand 512bit and GDDR5 would be quite insane. Too insane.

I imagine Nvidia has to make up for significant R&D costs and, in order to keep the price competitive, has settled for the lower priced GDDR3. ATI, on the other hand, has had its architecture out for a year and can afford to use the GDDR5 while keeping prices competitive. It is a marketing reality whenever a new product design comes out of a company and it attempts to break even and with its R&D investment as quickly as possible in order to better tweak pricing without offending the CFO. However, ATI using GDDR5 and Nvidia not using it can mean two things

ATI may NEED GDDR5 to compete with Nvidia
Nvidia may only need GDDR3 to compete with ATI.

Either way, if Nvidia turns around and offers a GTX 280 with 1GB GDDR5 for $450, it will retain the crown even if no one buys it and quantities will be limited, which is what it has done in the past and I doubt it will not do so if it becomes necessary to trump the speedy little 55nm 4870 core on GDDR5.

I just hope Nvidia knows that there is a price to pay, even for shooting too high above your competitor:



Nvidia on the left with a $400 GDDR3 GTX-280 with ATI on the right with $300 GDDR5 4870 after price cut

Perkam

[Shintai]

Too insane and setting benchmarks records are not mutually exclusive...

At what benefit? 0.1%? We already seen the R600->RV670 with 512bit to 256bit and basicly 0 difference. GT200 would have about zero benefit aswell from even twice of the R600.

So the only record to be set would be the darwin award nomination time.

[Luka_Aveiro]

I always thought bus width in nVidia's current GPU architecture determines the amount of ROPs

If you look at numbers:

G80 GTX: 384bit -> 24 ROPs
G80 GTS: 320bit -> 20 ROPs
G92 GTX: 256bit -> 16 ROPs
G92 GS: 192bit -> 12 ROPs

If nVidia decided to write GT200 with GDDR5 and 256bit bus it would only have 16 ROPs, which could be a bottleneck when using AA and high resolutions...

Am I wrong?

[RPGWiZaRD]

At what benefit? 0.1%? We already seen the R600->RV670 with 512bit to 256bit and basicly 0 difference. GT200 would have about zero benefit aswell from even twice of the R600.

^ I'd love to see some good comparision. Somehow I think your statement is a bit inaccurate. Or did you look into your magical crystal ball again? (don't take it too seriously)

Nah but I can't say if there's a big impact with bus size either but at least I have one simple good point, a company like NVIDIA doesn't go for 512bit if there's no reason in this case performance related as I fail to see what other things it could help with or are you doubting their $100.000 or whatever worth test equipment or test procedures too perhaps?