NVIDIA Maxwell Speculation Thread

[rpg.314]

http://www.google.com/patents/US20140118393
http://www.google.com/patents/US20140118380
http://www.google.com/patents/US20140118362

2nd gen is tile based rendering ?

I haven't read the patents either, but going by Arun's statement, imho their tiles are screen space tiles for raster/tri setup, not sort middle rendering.

 

[itaru]

add this
Managing per-tile event count reports in a tile-based architecture
http://www.google.com/patents/US20140118370

 

[LiXiangyang]

PCB of GM210? Has a 8-pin and 2 6-pin power connectors and 16GB GDDR5 memory.

http://news.mydrivers.com/1/310/310672.htm

 

[McHuj]

By the part number, those memories look like 4Gb parts, so with 16 chips it would only be 8GB. They 7.0 Gbps speed, if these are on a 512-bit bus, then I would guess 448 GB/sec.

Now, we need someone to measure the die size relative to the memory chips.

 

[Picao84]

What certainty to we have that this is a Maxwell at all? For all we know it can be the elusive GK210.

 

[Alexko]

Sounds like GM204 to me.

 

[tviceman]

By the part number, those memories look like 4Gb parts, so with 16 chips it would only be 8GB. They 7.0 Gbps speed, if these are on a 512-bit bus, then I would guess 448 GB/sec.

Now, we need someone to measure the die size relative to the memory chips.

It's a pointless exercise at this point because the measurement would be of IHS, not the die itself.

 

[Alexko]

It's a pointless exercise at this point because the measurement would be of IHS, not the die itself.

There's no IHS on GPUs, just the packaging. So you can remove something like 0.5mm on each side and you've got a pretty decent estimate of the die's size.

 

[Ailuros]

Either way those pictures are highly misleading; as a reminder way too many before the GK104 launch swore on over 350 square millimeters for it based on die pictures.

 

[AnarchX]

What certainty to we have that this is a Maxwell at all? For all we know it can be the elusive GK210.

GK210 has a GK110 sized package. The die size of this chip seems to be ~430mm², comparing with GK104. According to Zauba shipping data GK104 and GM204 have the same 40x40mm package size.

 

[Picao84]

GK210 has a GK110 sized package. The die size of this chip seems to be ~430mm², comparing with GK104. According to Zauba shipping data GK104 and GM204 have the same 40x40mm package size.

If you say so, I believe. The chip looked larger than that on the pictures to me. The power supply looks also OP. I would expect a Maxwell chip to be more energy efficient and not need even more juice than Kepler, even on 28 nm. As to this being GM204, there were rumors, now mysteriously removed that it would show up first in Mobile form...

 

[Ailuros]

No idea what their plans are, but GM206 sounds like a better candidate for a mobile first debut.

 

[psurge]

The power supply looks also OP. I would expect a Maxwell chip to be more energy efficient and not need even more juice than Kepler, even on 28 nm.

For a desktop Maxwell part, I think they need to have non-trivially higher performance than GK104, or new features, otherwise why buy it? To me, lower power consumption is not a compelling reason to buy in this particular segment (though I'm certainly no market expert). Given the lack of new features, I would expect to see the increased power efficiency used to deliver more performance than Kepler, with equal or even slightly higher target power consumption.

 

[Megadrive1988]

I'm hoping I'll be able to buy a graphics card based on the mid-range GM204 by the time (or soon after) the PC version of GTA5 comes out.

 

[Alexko]

I'm hoping I'll be able to buy a graphics card based on the mid-range GM204 by the time (or soon after) the PC version of GTA5 comes out.

I don't think GM204 will be a mid-range chip, not if it's meant to replace the 780 Ti.

 

[Megadrive1988]

I don't think GM204 will be a mid-range chip, not if it's meant to replace the 780 Ti.

Then if GM204 is high-end, what would "Big Maxwell" be?

 

[Alexko]

Then if GM204 is high-end, what would "Big Maxwell" be?

Probably a comparable chip in terms of die size and power consumption, but on a more advanced node, therefore "bigger" by transistor count (and performance, of course).

 

[fellix]

There's definitely room for a beefier GPU substrate, with 384-bit memory interface and a slightly larger die.

 

[Picao84]

I don't know how reliable these are, but they are certainly interesting bits. From the discussion/comments here: http://wccftech.com/nvidia-geforce-...es-8-gb-gddr5-memory-2nd-generation-gm2-core/

Before any other source:

They are keeping the SLI connector because the Denver CPU(s) now control SLI. Using PCI instead of the connector adds pointless latency since they no longer are using the main CPU for SLI, it would also add some stress to the CPU which kind of eliminates the point of using the Denver CPUs for SLI.

Nope, the top 8pin + 6pin are for the GPU and the other 6 pin is dedicated to the quad core Denver CPU. They will probably change it on the final though since the denver CPU doesn't use very much power, the prototype 880's Denver CPU is clocked at 3ghz.

Notice how the sli pins are more towards the middle? That entire area on the smaller side is basically dedicated realesate to the CPU, the holes near the top are so the heatsink can make better contact with the CPU.

I would have expected the ARM cores to be on the same die or package as the GPU, but maybe it was not feasible yet on 28nm?

Unfortunately he also says this, which sounds too good to be true:

225w TDP is the rumor, ~50% faster than the 780ti if it is a full GM204 chip.

 

[Ailuros]

Given the "source" it just brought tears in my eyes....