Welcome to our community

Be a part of something great, join today!

Log in
  • Hey all, just changed over the backend after 15 years I figured time to give it a bit of an update, its probably gonna be a bit weird for most of you and i am sure there is a few bugs to work out but it should kinda work the same as before... hopefully :)

Titan Z. Thoughts?

From what I've read, it is less bang for your buck than simply paying the same price and getting 3 Titan Blacks into one computer, but really it all depends on the formfactor of your computer, how many available PCI-E slots you have, etc.

The Titan-z uses 2 lesser clocked Titan Black chips on the same card, with the same amount of ram for each GPU. If you have the space (and power) to accomodate 2 or 3 Titan Blacks, you'd be better off going that way.
 
I'm somewhat tempted to put one in my Z820 since it's difficult to put more than one real GPU in there. The price is a killer. If you have the slots, it's best to go with multiple Titans, like Mike says. And the GTX 790 will be shipping soon, next month or so. Sure it has less RAM and only handles single precision, but I bet it's within 80% as fast as the Titan Z in apps like Resolve and Premiere and at 1/3 the price. Dual 780 SC's right now are about 15% slower than dual Titan SC's in Resolve. Although on memory-intensive tests like NR, the Titans fare a whole lot better because they have double the RAM onboard.
 
What about 780 SC with 6GB? Those run a hair over $1K for a pair giving you 12GB total.
 
Jason Osterday said:
What about 780 SC with 6GB? Those run a hair over $1K for a pair giving you 12GB total.
Click to expand...

I think that the RAM per GPU is more important than the total amount of RAM. Multiple GPUs increase speed. More RAM per GPU increases what you can actually do (number of nodes in resolve for example). Somebody correct me if I am wrong.
 
Jeff, what is your take on system RAM to match GPU RAM? A friend, at NVIDIA, suggested that I have double the RAM in my system that is on the GPU to avoid memory copies. e.g. I have a K6000 with 12 GB of RAM. He said I should have at least 24 GB in my system just for the GPU. The Grantley system I will build later this year will probably have 128 GBs of RAM in it. 2-4 GB per core + 24 GB for the K6000. I may add a 2nd K6000 so another 24 GB. :)
 
It's incredible....but like Jeff stated, price is somewhat prohibitive.

I would love to see what these guys can do in quad SLi..... that would be 23,040 cores, with 48 GB of DDR5....
 
Stacey Spears said:
Jeff, what is your take on system RAM to match GPU RAM? A friend, at NVIDIA, suggested that I have double the RAM in my system that is on the GPU to avoid memory copies. e.g. I have a K6000 with 12 GB of RAM. He said I should have at least 24 GB in my system just for the GPU. The Grantley system I will build later this year will probably have 128 GBs of RAM in it. 2-4 GB per core + 24 GB for the K6000. I may add a 2nd K6000 so another 24 GB. :)
Click to expand...

Double the GPU RAM would be the absolute minimum, IMO. If you're running software that can capitalize on 12GB of GPU RAM, then I would think 24GB otherwise would be totally insufficient, unless the software primarily runs in GPU space.

I usually go with a relation of RAM per cores in the system and for working with 5K~6K footage, 4GB per core is a bit on the thin side now and I aim for a rule of 4GB *PER THREAD*. That's not a bad place to start. In my 16-core Z820 workstation, I have 128GB. At times it's too much, but I won't say it's overkill. There are times I can use near all of it -- could use all of it if I really tried to. Keeping a properly even quad-channel configuration means I have to drop to 64GB as the next proper configuration below 128GB. That's a pain in the ass and I've overrun 64GB in that system a number of times, hence why I now have 128GB. This is one of the biggest downfalls of the new Mac Pro. In a 12-core, dual GPU 6+6GB, cylinder, that 64GB is a bit restrictive.

I maxed out my 20-core E5-V2 build with 256GB and don't regret it. Figuring 4~8GB aside for OS and general application space, and 8GB of space to mirror what my dual 4GB GTX780 cards have, the remaining 240GB works out to 6GB of RAM per thread. Not at all out of line if I'm doing serious compositing or rendering frames out at 4K+ resolution from Modo/ Maya, Nuke, etc.. And 128GB wouldn't have been enough to feed it if I'm firing on all cores for rendering or other like tasks.

When the V3 Xeons come out in 6~8 months, I'll see if it's worth building another box. I may just wait for Broadwell core Xeons in 2016 unless a sudden need arises. At that point, I can totally see 32-core systems with 512GB RAM being a common setup within the high-end niche of the market.
 
Michael Vera said:
It's incredible....but like Jeff stated, price is somewhat prohibitive.

I would love to see what these guys can do in quad SLi..... that would be 23,040 cores, with 48 GB of DDR5....
Click to expand...

If I am not mistaken still, you'd only have 2 working display cards for SLI as the Nvidia maximum number of GPU for SLI is 4, whether four individual GPU or a pair of Dual GPU cards... Hence, you are capped at Maximum 2 of these. However, that's the reason for using the K6000 for display with with multiple K40's which have no graphic output concern, simply compute....
 
Last edited:
So you guys are really pushing for 32GB modules? The cost bump is tough to measure on a nMP considering the 2690v2 was the same price and considerably faster (vs 4/6 core).

Has anyone tested multi gpu ram vs single? Say two 3GB 780 SC vs 1 Titan? For the sake of understanding how well the RAM is distributed and utilized across SLI.
 
SLI is different than scaling with GPU acceleration. SLI is just distributing drawing out data to the displays across more than 1 video card. Scaling GPU acceleration data is specific to CUDA or Open CL and a completely different process. The GPU's don't have to be in SLI for scaling GPU acceleration data across multiple GPU's.
 
A few months ago I Googled GTX 790" and I thought it was called the "Titan Z"

Did I miss something. There isn't much info on the GTX 790.

I personally have a GTX 690 on my QUO Hackintosh takes about 40min for 10Min of Epic MX 5K HD 5:1 to ProRes 422HQ.

It runs pretty hot so I needed some fans and it takes two 8-Pin Power.

Ever since Red updated Cine-X to Dual GPU compatibility it's improved my speeds drastically.

I'm very grateful Thx to all the Cine-X coders I owe you one.

I've looked at tests I don't know if the Titan outperforms my 690 in many cases that I've seen online my card is faster.

If the GTX 790 does launch I'll be super tempted the Titan Z is overpriced for the performance and it takes up 3 SLOTS!
 
Unfortunately application code and efficiency along with memory management and caching models don't update at the pace of hardware. There are also other factors involved such a CPU processing and memory management that limit the load on the GPU's at any point in time. This is why you see previous gen cards at time show greater efficiency or performance at some tasks. However raw specs of the Titan cards are far greater than previous 600 series cards. So the processing potential is far greater. The ram bandwidth on the cards can significantly effect the GPU acceleration performance as well. This often wont show itself because the CPU's and ram are the current bottleneck with the GPU's far beyond what those can handle.
 
You must log in or register to reply here.
Back
Top