Titan V announced, $3000 volta gpu

[Bouowmx]

I read from PC Perspective, and see from Gamers Nexus that TITAN V has 96 ROPs, the same as the previous TITAN Xp.

So, with unbalanced architecture, TITAN V is ROP-bottlenecked in some cases.

 

[HurleyBird]

So there are probably 128 ROPs in total on the die, but 1/4 of them are disabled to go along with the disabled portion of the HBM memory interface.

GamersNexus recently posted a video about Titan V overclocking where they were getting something like a 6-12% performance uplift at 4K with a 200 MHz memory overclock. I'm guessing that a theoretically fully enabled Volta could see gains around 20% at 4K.

 

[tamz_msc]

As before on what? A Titan? That was in the last previous generations are much smaller hybrid HPC/rendering chip and not the high end HPC chip?

Only with Pascal did you have different chips for Tesla and Quadro/Geforce/Titan. Kepler Titan was GK110, which was the same chip which featured in the K40. Maxwell Titan was GM200, which also featured in the M40. The only difference was Pascal, and since NVIDIA is showing no desire to spin off the GV100 at the moment, it means they're obviously going to market the card like it was during Kepler Titan and Maxwell Titan.

Hence they feature the same type of design on the cooler as before.

 

[tamz_msc]

So there are probably 128 ROPs in total on the die, but 1/4 of them are disabled to go along with the disabled portion of the HBM memory interface.

GamersNexus recently posted a video about Titan V overclocking where they were getting something like a 6-12% performance uplift at 4K with a 200 MHz memory overclock. I'm guessing that a theoretically fully enabled Volta could see gains around 20% at 4K.

This is true. Titan V has 4.5MB L2, whereas GV100 has 6MB L2. Since L2 is tied to the ROPs it means that the full chip has 128 ROPs.

 

[Yotsugi]

128 ROPs? How do you even split them between 6 GPCs?

 

[BryanC]

128 ROPs? How do you even split them between 6 GPCs?

There is a crossbar between the GPCs and the L2 (where the ROPs are).

 

[Glo.]

It also means that we should expect per clock increase in performance from consumer Volta chips, compared to Pascal GPUs .

 

[Muhammed]

Yep, the increased ROPs and clocks from consumer Volta will result in an even more gains over Pascal. Adding 16 Gbps GDDR6 and removing Tensor units and FP64 will help with that tremendously.

 

[tential]

We had two titans last gen.
Anyone ready for the Titan VIP Volta?

 

[Bouowmx]

For real, no way NVIDIA will pass the opportunity for graphics-optimized G?102 TITAN, focusing on single-precision and featuring a full complement of ROPs. I suppose that is where "Ampere" comes into play.

 

[Guru]

Being double the size of 1080ti and costing $3000 its not impressive in games at all, even when OC'ed. We are looking at 20% to 40% performance increase in most games. There are few outliers, but I'll put that off to driver issues and possibly the games having weird optimizations.

I don't see them selling a 800+ die size GPU. So we are probably looking at a 500 chip the Titan XXL for $1200, likely with GDDR6 at 15gbps, much higher clocks and probably up to 5000 cores at best.

 

[NTMBK]

Only with Pascal did you have different chips for Tesla and Quadro/Geforce/Titan. Kepler Titan was GK110, which was the same chip which featured in the K40. Maxwell Titan was GM200, which also featured in the M40. The only difference was Pascal, and since NVIDIA is showing no desire to spin off the GV100 at the moment, it means they're obviously going to market the card like it was during Kepler Titan and Maxwell Titan.

Hence they feature the same type of design on the cooler as before.

But during the Maxwell generation, NVidia also created the GK210 to satisfy the demand of double-precision heavy compute users with complex kernels. The split has been happening for quite some time.

 

[StrangerGuy]

Nvidia are brilliant in their branding over the years, they know how and when to milk the shit out of consumers. Coming out with a Titan-branded card based on their big HPC die (something they haven't done for many years) and throwing it up for $3k brings out the whales who simply can't have anything less than TIIIITTAAAANNNNN.

As opposed to what, AMD selling Vega at a loss for brownie cred points? Besides you really think AMD wouldn't want to be in NV shoes right now?

 

[Malogeek]

As opposed to what, AMD selling Vega at a loss for brownie cred points? Besides you really think AMD wouldn't want to be in NV shoes right now?

As opposed to nothing? I didn't say they shouldn't be doing it, nor comparing to AMD. It's an ideal business decision for a product at the right time, that's all.