NVIDIA Maxwell Speculation Thread

We can already see the GTX 970 struggling in some games with its 3.5GB pool of fast memory, and the R290X 8GB is sometimes significantly faster than the 4GB variant. So it's not unlikely that games would need more than 6GB within a couple of years, especially in usual circumstances (4K, supersampling, VR, stereoscopy, multi-monitor setups, or any combination of the above).

GM200 has a 384-bit bus, so barring a hybrid memory configuration, it was either 6GB or 12GB (3GB is clearly not enough).

As if framebuffers take up nearly that much memory, and the 970 struggles because it's memory is divided into full bus bandwidth and slow bus. The game using it inevitably doesn't know there's 2 memory pools, so it's up to the drivers to keep resources out of the 512mb of slow memory if possible. When that fails the bandwidth access times spike and framerate drops.

You simply won't need 12gb. It's questionable how much you'd need 8, with about 5gb of ram reserved for apps on the consoles you'd need less than a gig extra even to run a 5k framebuffer. Hell making even a crazy 256bit per pixel framebuffer at 5k gives just a 480mb framebuffer. And 8 full 32bit render targets stored at once is supremely high. Good luck getting a 384bit bus to supersample that. No, just like last time the new Titan is clearly dedicated to compute, and at the PCB size and sheer cost of the ram will no doubt have a cost to suit.

The upcoming Nvidia and AMD launches look like a repeat of the Titan/780ti v 290x launches. Similar performance goals (probably +/-15% of each other) with Nvidia clearly aiming at high profit margin compute customers with a huge expensive card with a lot of ram and AMD clearly aiming at lower profit margin/higher volume gamers with a lower price and less ram (4-8gb?).
 
Titan is clearly dedicated to compute
I don't think so, I think the card is like a gamer version of a Ferrari. Sure a Ferrari is powerful and expensive but its not a F1 by any stretch of the imagination. Nobody needs a Ferrari for driving on the roads but it is not a dedicated race vehicle. The Titan X is overbuild for gaming but it is still market towards gamers. It is still in essence a gamer card. Much like people who buy $1000 CPUs for gaming, people will pay $1000 for a GPU with the top of the line specs to game on. They may not need 8c/16t and 12GB of vram in any games but that doesn't mean they didn't buy the system to game on.
 
When I tried to speculate about big GPU dedicated to compute or dedicated to not compute I was wrong every time, so I'll say it does both.
Though, no double precision benchmark or specs yet?
 
What do you expect from DP given respective (rumored) die sizes and DP rates from other Maxwell chips?
 
Without a new process (which I think it's universally agreed-upon we won't be seeing this time around), wouldn't it be impossible to hit any significant 3D rendering performance improvement without sacrificing DP?

Add to this, the confirmation (or maybe only unofficial/rumor?) that there won't be any professional/tesla boards this time around either, and you would expect DP to be seriously cut-down compared to current titans, yes?
 
It's been said[who?] that area/cost of DP is considerably more than 2x of SP. As a layperson, that's all I know...

Yes, for the multiplier itself that's true, but what does that mean for the size of the total die? If it amounts to 10mm², it's probably worth it for GM200 even if it's already very big. If it's more like 100mm², it's probably not.
 
For a single precision multiply, you need a 23x23 multiplier for the mantissa. For a double precision a 52x52. That's followed by normalization. An add or a multiply-accumulate can be even more problematic, because both terms have to normalized to a common exponent.

My guess the hardware that's quarter rate double precision generates the 52x52 product with 4 passes through a multiplier a single precision multiplier.
 
"Designed for single precision, for those wanting double precision we still have the Titan Z"
 
Back
Top