Titan Z now only $1499

[Safety]

Re-read Filter's original post in context. What he said is correct. The usable pool of memory for either Gaming or CUDA applications is 6GB.

the 12GB VRAM is the only thing drool worthy about the TitanZ...I hope Nvidia doesn't continue to restrict huge VRAM to the Titan cards

you mean 6gb got to love marketing.

Polynyc2 forgets that you 1) divide the memory by the number of GPUs 2) there was a 6GB variant of the 780 available.

 

[wonderfield]

Re-read Filter's original post in context. What he said is correct.

No, it is not correct. The Titan Z has 12GB of memory. It is not just "marketing", though obviously the available amount of memory is a marketing point for NVIDIA — the Titan Z physically has 12GB of video memory, distributed across how ever many chips with however many gigabits per chip. Each GPU has direct access to 6GB, but both GPUs need not necessarily be working with mirrored data sets. That is necessarily the case with SLI, but not in general.

 

[xorbe]

I believe @wonderfield has it. When doing graphics, the 2x6GB are mirrored (aka SLI), but for computational, it's like two independent gpus and vrams. It's not so surprising.

 

[Safety]

I believe we're all saying the same thing but in different ways. What matters is that VRAM is not pooled and no single application can utilize all 12GB (SLI or non-SLI) of the TitanZ. A CUDA application that uses more than one GPU (non-SLI'd) is limited to the part with the lowest amount of VRAM. If my system had 4 x 980's with 4GB, I wouldn't claim to have 16GB of VRAM. Furthermore if one of those 980s were devoted to driving the GUI while the other 3 were used for compute, it would be inaccurate to say that I have 8GB of VRAM.

 

[Unknown-One]

Titan is a gaming card and competes with the GTX 980.
My point is that the price should reflects this fact.

The Titan sits in the niche between gaming cards and workstation cards. The price reflects this position.

You would need a rack full of GTX 980's to match the double-precision performance of a Titan. These cards are not even close to comparable when you look at all possible workloads.

If you want a card that's actually designed for gaming, the GTX 780 Ti and/or GTX 980 is what you're after. Can get gaming performance better than a Titan with a lower pricetag, and you wont miss the removal of workstation-oriented features if all you're doing is gaming.

The usable pool of memory for either Gaming or CUDA applications is 6GB.

They're trying to tell you that it's possible to take the Titan Z out of SLI mode, thus making all 12GB usable. This could be handy for GPU-compute applications that can be coordinated across both GPUs.

 

[Safety]

They're trying to tell you that it's possible to take the Titan Z out of SLI mode, thus making all 12GB usable. This could be handy for GPU-compute applications that can be coordinated across both GPUs.

I'm not a gamer but someone who uses CUDA applications. VRAM does not pool for non-SLI scenarios. You could launch multiple instances of an application but then your data set would have to be divided. In terms of usage the TitanZ is nothing more than two normal Titans on the same PCB.

 

[wonderfield]

I understand the confusion now. I responded to someone using the term "pool" to the effect that applications could address the entire 12GB pool. "Pool" in that context was ambiguous and I assumed it was being used in a general sense rather than the more strict technical sense of a memory pool.

I don't believe a single context can access the entire 12GB. Of course, all 12GB is certainly addressable and useable by applications. Even in SLI it's "useable", just to no direct benefit.

 

[RadXge]

The Titan sits in the niche between gaming cards and workstation cards. The price reflects this position.
You would need a rack full of GTX 980's to match the double-precision performance of a Titan. These cards are not even close to comparable when you look at all possible workloads.

No. The Titan is a gaming card not a workstation card. NVIDIA clearly markets the Titan as a gaming card. See for yourself, http://www.nvidia.com/gtx-700-graphics-cards/gtx-titan-black/

Also,

However, we wouldn't recommend the Titan (and we don't think Nvidia would either) for everyday work in a design studio. Quadro and FirePro boards are the right tools for that job, primarily due to their properly optimized and certified drivers.

http://www.tomshardware.com/reviews/geforce-gtx-titan-opencl-cuda-workstation,3474-23.html

Your confusion arises form the fact that the Titan is for the most part a rebranded Quadro.
The titan simply doesn't have the proper driver support for workstation tasks (even though the double-precision power is there).

 

[Safety]

There are too many gamers in this thread making wildly incorrect statements about GPGPU uses for the Titan and TitanZ.

Your confusion arises form the fact that the Titan is for the most part a rebranded Quadro. The titan simply doesn't have the proper driver support for workstation tasks (even though the double-precision power is there).

The Titan is a rebranded Tesla K20X without ECC memory. The closest analog in the Quadro line is the single gpu K6000 with 12GB of ECC capable VRAM. For single-precision compute purposes there is no difference between the Titan, the Tesla and the Quadro when it comes to drivers. Drivers only come into play with the Quadro line if you need accurate on screen rendering or if the software vendor will only offer technical support with qualified hardware. These days most application are not as finicky when it comes to using a gamer card to drive the display. If you have an application that utilizes double-precision then you would most likely need ECC capable VRAM as well (scientific computing where the results must be accurate).

I don't believe a single context can access the entire 12GB. Of course, all 12GB is certainly addressable and useable by applications. Even in SLI it's "useable", just to no direct benefit.

Correct. If I had an 11GB data set with a TitanZ installed I would have to split it into two or more smaller parts to fit into the 6GB ceiling. On the other hand a single gpu Quadro K6000 with 12GB of VRAM would allow me to fit the entire data set into one application instance.

 

[Unknown-One]

No. The Titan is a gaming card not a workstation card. NVIDIA clearly markets the Titan as a gaming card.

Incorrect. Like I said previously, the Titan sits in the niche between gaming cards and workstation cards.

It is neither, and both at once. The price and feature-set reflect this middle-ground position.
Titan products are more expensive than comparable GeForce products, but less expensive than comparable Quadro products.
Titan products have more features than comparable GeForce products, but less features than comparable Quadro products.
Titan products are restricted to the Nvidia reference cooler (like Quadro cards), while GeForce cards can be customized by partners.

Titan products do not deliver any additional gaming performance over equivalent GeForce counterparts. GeForce products are dedicated consumer-level gaming hardware, Titan products are decidedly more than that.

I'm not a gamer but someone who uses CUDA applications. VRAM does not pool for non-SLI scenarios.

I never said it did, I said you'd have to use a CUDA application that could be coordinated across both GPU's.

This would require a client-server architecture with each GPU (and its own memory pool) running an independent instance + a front-end application keeping them in sync. A basic cluster.

large-scale deployments used for scientific applications already employ similar architectures, making a LOT more GPU RAM useable.

 

[Safety]

I never said it did, I said you'd have to use a CUDA application that could be coordinated across both GPU's.

This would require a client-server architecture with each GPU (and its own memory pool) running an independent instance + a front-end application keeping them in sync. A basic cluster.

large-scale deployments used for scientific applications already employ similar architectures, making a LOT more GPU RAM useable.

Not all data sets can be segmented. In these scenarios, the size your data set is limited to the lowest amount of VRAM available in the cluster. For instance the Nvidia Grid is composed of what are essentially GTX 680s with 4GB of VRAM. In my case I cannot process a data set larger than 4GB through Amazon's EC2 platform. The number of CUDA processors is cumulative but the VRAM available is 4GB. Launching multiple independent instances is not practical for this situation.

 

[Armenius]

Incorrect. Like I said previously, the Titan sits in the niche between gaming cards and workstation cards.

It is neither, and both at once. The price and feature-set reflect this middle-ground position.
Titan products are more expensive than comparable GeForce products, but less expensive than comparable Quadro products.
Titan products have more features than comparable GeForce products, but less features than comparable Quadro products.
Titan products are restricted to the Nvidia reference cooler (like Quadro cards), while GeForce cards can be customized by partners.

Titan products do not deliver any additional gaming performance over equivalent GeForce counterparts. GeForce products are dedicated consumer-level gaming hardware, Titan products are decidedly more than that.

It doesn't matter what the card actually is... Yes, it's hardware sits in between GeForce and Quadro. But NVIDIA markets the TITAN line as gaming products, meaning their intended use is for gaming as defined by NVIDIA. Look at what drivers all the TITANS use: They use GeForce drivers instead of Quadro drivers, which is why you can't do all you can with a TITAN when compared to a Quadro.

It's a good product for people who need compute performance, but can't afford a good workstation graphics setup. But let's not try to explain away how NVIDIA is marketing the TITAN product line.

 

[Unknown-One]

It doesn't matter what the card actually is... Yes, it's hardware sits in between GeForce and Quadro. But NVIDIA markets the TITAN line as gaming products, meaning their intended use is for gaming as defined by NVIDIA.

Their intended use is as a middle-ground product. Part of that intended use is, of course, gaming workloads.

The price and feature-set continue to reflect this middle-ground position between gaming and workstation cards.

Look at what drivers all the TITANS use: They use GeForce drivers instead of Quadro drivers, which is why you can't do all you can with a TITAN when compared to a Quadro.

Which is why Quadro cards command an additional premium over Titan cards.

Once again, the middle-ground position of the Titan line is very-obviously being displayed.

It's a good product for people who need compute performance, but can't afford a good workstation graphics setup. But let's not try to explain away how NVIDIA is marketing the TITAN product line.

I'm not trying to explain anything away,

It's a card that can handle ALL the workloads of gaming hardware and SOME of the workloads of workstation hardware. The fact that Nvidia is marketing the Titan's ability to play games is not surprising, considering that's a major part of its overall feature set.