Nvidia GT300 core: Speculation

SLi was Big Bang I AFAIR.

 

So, this would be a "Slii".
Sounds good... but only if it rimes with "Wii".

 

It has two quite different iterations though (SM2 and SM3). One could say that DX9 was like DX10+DX11 in terms of features...

 

Mm, my point was more along the lines that while it was the longest lasting, DX10 even though getting so much bad mouthing, is going to live pretty much as long as the "king of the hill"

 

Based on what MS revealed about DX11 at Gamefest it will probably have faster uptake than DX10. DX11 will run on DX10 and DX10.1 hardware and drivers, although some of the new DX11 features like tesselation will require DX11 hardware. It's not like DX10 where it's an all or nothing deal with feature support.

The new multithreading friendly features of DX11 will be supported on DX10 hardware with DX10 drivers (though they'll be more efficient with DX11 drivers and even more so with DX11 hardware) and I expect that functionality more than anything else will drive uptake of DX11. DX11 is currently expected to ship Q3 or Q4 09 with a tech preview coming this year.

 

Its weird suddenly reading about DX11 at this point in time.

DX10 seems like it was just yesterday.

 

And support in games still seems like tomorrow

 

There's a mistake, G98 is geforce 8400GS (8SP/64bit), the 9400GT is a G96 with half the shaders disabled (16SP/128bit). Maybe that doesn't change anything to the other information's veracity but let's be rigorous on the details .

 

Well, G92x2 isn't exactly 'Enthusiast' right now also.

 

True.

Am I the only one that find it strange NVidia isn't updating it's mainstream and valuesegment until 3rd quarter -09. Surely they must be hurting in those segments? And that is where the most money is?!

 

The graph states compiled by vr-zone underneath doesn't it? Bears the question above all why would NV start from the top with 40nm and then move down and NOT dip its toe in the water with something else as AMD is doing with RV740.

 

So it sounds like 'NV60', which would be Nvidia DX11 part, is coming in September

G80 = 'NV50' (new generation part, DX10)
GT200 = 'NV55' (major upgrade part, still DX10)
GT300 = 'NV60' (new generation part, DX11)

 

Yes, there will be several changes with GT300. Not the same ALUs and so on...

 

Assuming this GT300 will be on a 40 nm process, allows for some serious increase of transistors again This time Nvidia will not want to miss breaking the 2 TFlop barrier.
To that end they might want to increase the ALU:tex ratio from 3:1 to 4:1.

Hypnotically a GT300 could look like

- 512 shaders
- 128 texture units
- 512 bit GDDR 5
- 2.2 B transistors
- 650 Mhz core. 1.5 Ghz shaders
- 460 sqmm, on 40 nm
- 2.3 TFLOP (512*3*1.5)


The AMD R800 could be something like

- 1280 shaders
- 64 texture units
- 256 bit GDDR 5
- 1.3 B transistors
- 800 Mhz
- 270 sqmm, on 40 nm
- 2.05 TFLOP (1280 *2* 800)

 

I'd say that'd be more like 200m2. But that's for another thread.

Jawed

 

Even I would say under 200 sqmm with this specs.

 

I'm wondering if scaling works as good as it used to be.

The GT200b is rumored to be 480 sqmm, where it should be more like 412 sqmm.

 

IIRC someone once mentioned that memory controllers don't "scale aswell as most of the chip", so having that 512bit big fat bus might explain part of the "worse than expected scaling"

 

Some photos of an uncapped G200-B were leaked time ago and the chip area was measured to be within the narrow margin of 480~484 sq.mm.

That would be the I/O contact interface, not the control logic itself.