Nvidia confirms GT-200 with 1 billion transistors

[v_rr]

512bit bus with GDDR3 at 2000-2200 would be faster than any available ram on a 256bit memory bus....

256bits + GDDR5 gives you more bandwidth.

Look to R600 at 512bits. The formule didn´t work.
The PCB is much more complex, the core is bigger, you have more bus stops inside the chip.

[railmeat]

Finally something worth upgrading to!

When are they out?

lol i so agree.

i scrapped the 9800gtx from 8800gtx upgrade idea when i read lesser specs


@ v rr "The PCB is much more complex'

the pcb is merely for heat i thought being thicker\thinner,nothing more.

[keiths]

Compare to SLI: 256, 48/32, 1.28/1gb.

[DilTech]

256bits + GDDR5 gives you more bandwidth.

Look to R600 at 512bits. The formule didn´t work.
The PCB is much more complex, the core is bigger, you have more bus stops inside the chip.

R600's formula didn't work because it couldn't USE the bandwidth. Too few rops, inefficient AA, too few tmus for AF... The rest of the architecture held it back.

Also, you realize you'd need roughly 4ghz GDDR5 at 256bit to match 2ghz GDDR3 at 512bit, right? Also, GDDR3 is likely to be lower latency than GDDR5 as well.

[BenchZowner]

Awww only GDDR3?

No need for more expensive RAM...
GDDR3-2100/2200 + 512BIT BUS = 134.4Gbps / 140.8Gbps

Already more than good enough

512bit bus with GDDR3 at 2000-2200 would be faster than any available ram on a 256bit memory bus....

GDDR5-4035 + 256BIT BUS = GDDR3-2200 + 512BIT BUS.
GDDR5-6000 + 256BIT BUS = 192Gbps
GDDR3-2600 + 512BIT BUS = 166.4Gbps

{ Edit: quoting your message, not disagreeing }
Those numbers are just insanely high, and current GPUs won't benefit from a point... and that point is under those figures

And plain bandwidth wont give you a thing... if you haven't equipped the card with efficient sub-units ( ROPs/TMUs/TFUs ).

Compare to SLI: 256, 48/32, 1.28/1gb.

Care to elaborate ?
256 you mean SPs.
that 1.28/1GB ???

[hennyo]

Time to get a second job. My 8800 gtx is looking to retire, its getting tired.

[STaRGaZeR]

R600's formula didn't work because it couldn't USE the bandwidth. Too few rops, inefficient AA, too few tmus for AF... The rest of the architecture held it back.

Also, you realize you'd need roughly 4ghz GDDR5 at 256bit to match 2ghz GDDR3 at 512bit, right? Also, GDDR3 is likely to be lower latency than GDDR5 as well.

I think his point is that if the two options can provide the same bandwidth (256+GDDR5 vs 512+GDDR3) it's cheaper and way less complex to design and use the first. And I agree with that. I loved the R600 PCB's visual, but looking at it from a practical POV, 256+GDDR5 is better. Also, 512bits will force NVIDIA to use 16 chips in one PCB, just like the 2900XT. If they plan to do a GX2 with the GT-200, I just can't imagine the monster PCB (two of them more than likely) it'll be with 32 GDDR ICs, that huge chip and all the power circuitry.

Of couse it'll be fast, but I don't want a second 2900XT talking about heat and power consumption.

[MuffinFlavored]

BenchZowner really gave out a lot of information here, thanks.

More ROPs? Bigger memory bus?
Did they finally listen to the thousands of shouts?
Before or after they said "we are opening a can of whoop ass on intel"?

[AliG]

This going to be awesome, only thing I'm worried about is die size. The yields of this will be even worse than that of g80's, pricing will definitely be steep here. I have no doubt it'll definitely be a thorn in the r700's ass because hardly any of ati's xfire drivers are offering consistant above 75% scaling of the 3870x2 vs 3870. Not to mention I can't think of any games except for maybe CoJ that's coded to make use of all the shaders the r700 will have, let alone a single rv770. Thing is, I'd be willing to bet there's going to be a huge difference in price.

This battle is definitely going to be hot. I wouldn't be surprised if either side took the performance crown here (but I'm leaning more towards gt200 due to how inefficient the r600 design is in current games), but this will be more than just pure performance. Its going to be close, and pricing will mean a lot this time around.

[adamsleath]

Whenever AMD is near or the G9x cores dont sell well more.

ha. qft

[DeathReborn]

R600's formula didn't work because it couldn't USE the bandwidth. Too few rops, inefficient AA, too few tmus for AF... The rest of the architecture held it back.

Also, you realize you'd need roughly 4ghz GDDR5 at 256bit to match 2ghz GDDR3 at 512bit, right? Also, GDDR3 is likely to be lower latency than GDDR5 as well.

You forgot that GDDR5 is still not "available" so the original statement stands correct, for the time being.

[BenchZowner]

^_^

65nm
240 SPs [ MADD + MUL more than likely ]
120 TMUs
120 TFUs (?) [ not so sure about that ]

[LordEC911]

1100 MHz on a 512 bit bus would be 140 GB / second of memory bandwidth. 200 shaders, 32 ROPs, and 1 GB or Ram, damm. Sounds amazing!

192...

[BenchZowner]

196...

240...

[LordEC911]

240...

So "close to 200" becomes "256 with a cluster disabled."
This just shows(if true) what was obvious about the rumors, huge, hot and expensive due to poor yields.

Edit- Sorry I didn't see your post.

^_^
65nm
240 SPs [ MADD + MUL more than likely ]
60 TMUs
120 TFUs (?) [ not so sure about that ]

Let me work through the numbers but they don't really make much sense at first glance.
Edit- Hmmm... So going back to the original and just buffing up G80 numbers.

[BenchZowner]

So "close to 200" becomes "256 with a cluster disabled."
This just shows(if true) what was obvious about the rumors, huge, hot and expensive due to poor yields.

Edit- Sorry I didn't see your post.



Let me work through the numbers but they don't really make much sense at first glance.

Oh come on! Don't make me do that
10 clusters...

[dinos22]

^_^

65nm
240 SPs [ MADD + MUL more than likely ]
60 TMUs
120 TFUs (?) [ not so sure about that ]

how many STFUs? j/k

wow if that is true and this this is a single card, single core solutions
this will be crazy fast lol

[BenchZowner]

how many STFUs? j/k

One, a major one [ sssh, be quiet! ]

wow if that is true and this this is a single card, single core solutions
this will be crazy fast lol

Supposed to be single card, single die.

[dinos22]

wow man
if your source is reliable AMD should be packing
this is looking like a proper Ultra replacement that will murder it heheh

[BenchZowner]

GT200 performance should be higher than 2* G92GTS in many cases.

*Clocks are estimated, theoritically easy to reach*

G92GTS - GT200
648MHz/1620MHz/GDDR3-1944 - 600MHz/1350MHz/GDDR3-2200
256BIT BUS Width - 512BIT BUS Width
512MB FrameBuffer - 1024MB FrameBuffer
64 TFUs - 120 TFUs
64 TMUs - 120 TMUs
16 ROPs - 32 ROPs
128 SPs - 240 SPs

[Zytek_Fan]

http://www.youtube.com/watch?v=TBtpyeLxVkI

[LordEC911]

Oh come on! Don't make me do that
10 clusters...

Interesting approach...

GT200 performance should be higher than 2* G92GTS in many cases.

*Clocks are estimated, theoritically easy to reach*

G92GTS - GT200
648MHz/1620MHz/GDDR3-1944 - 600MHz/1350MHz/GDDR3-2200
256BIT BUS Width - 512BIT BUS Width
512MB FrameBuffer - 1024MB FrameBuffer
64 TFUs - 120 TFUs
64 TMUs - 120 TMUs
16 ROPs - 32 ROPs
128 SPs - 240 SPs

You know they are gunning for at least 1500mhz on the shader domain, they can't stop just shy of 1TFlop.

[Lightning_Rider]

All I can say is it better come out within 3 months. I just bought an eVGA 9800GTX since my 320MB GTS wasn't hacking it on my 24" monitor.

If this comes out in June or early early July then I'll be very happy.

[Nicksterr]

All I can say is it better come out within 3 months. I just bought an eVGA 9800GTX since my 320MB GTS wasn't hacking it on my 24" monitor.

If this comes out in June or early early July then I'll be very happy.

don't hold ya horses.

[adamsleath]

just add 5% on to whatever ati brings out. (leaves office to play golf)