Cannot seem to find this anywhere and need to know before ordering.
Do all of the RTX GPUs support Tesla Compute Cluster mode. Desperately need this to do Windows 10 development for deployment on Amazon EC2 P3 V100.
Has anyone found any information on this?
Thanks
GeForce GPUs generally do not support TCC
Titan, and most Quadro support TCC. The Quadro RTX GPUs announced so far (8000, 6000, 5000) should all support TCC.
Got it. I had only seen the announcement on the GeForce RTXs, not the Quadro RTX ones.
Ouch, NVidia are really making it difficult for developers. At $2,300 for the low end card that is a killer.
Welcome to the end of Moore’s Law. The time of deflationary pricing for computing devices is basically over, and component pricing in the computer industry will work like it works in other mature industries.
That also means that the expectation of getting better performance at roughly the same price as previous generation parts no longer holds. From here on out, more performance will cost more money, and generally performance will increase more slowly, except for niches where new accelerators come to bear.
The development costs for a new high-performance processor generation (CPU or GPU) are currently in excess of $1B, maybe more like $2B. At the same time, die sizes and thus manufacturing costs also exceed those of prior generation parts. In this situation, how would you price new products if you had to recoup the expenditures?
If some other vendor could offer better performance/price or better performance/watt for their compute solutions, NVIDIA would likely be forced to lower their prices (with a resulting drop in profit margins). The fact that this is not happening is a pretty good indication that NVIDIA offers best-in-class performance.
Since the announcement of the new RTX 3000 Product line, it became apparent that there will be no new Titan RTX, but the 3090. I’m wondering if for that card, TCC will be enabled? Following your GeForce Argument, thats not the case.
Also one was able to set Graphics Clocks for the Titan RTX with nvidia-smi, but not for the 2080 TI, so now I’m wondering where the line for those features is drawn now.
Or whether P2P Cuda Memcopy between two GPU’s have to pass through CPU, or can happen directly over NVLINK.
Thanks in advance!
I wonder how does the WDDM 2.7 “hardware accelerated GPU scheduling” compare to TCC in terms of CUDA efficiency?
Can you confirm TCC support for the RTX 3090? Since there are no Ampere cards for Quadro at the moment, looking for something in Ampere that will work with CUDA.
Hehe, yep I get it. But hurting developers by forcing them to buy incredibly expensive Quadro cards for development will ultimately hurt Nvidia’s sales. We have to get client applications running somewhere.
It should be noted that “TCC” is a unique feature / restriction relevant only to Windows. It’s sole feature is to allow peering in the Windows environment where that is otherwise not available because of Windows OS restrictions. GeForce cards have always been restricted in terms of peering in Windows.
In Linux however, any version of Linux, TCC is irrelevant. What is relevant is hardware peer-to-peer capability enabling UVA (unified virtual addressing). Up until the 2000 series, All 900 series and 1000 series GeForce cards could peer (share memory) over pcie in Linux - which is really what this is all about.
Beginning in the 2000 series of GeForce cards, peer-to-peer cannot be achieved over pcie. You could connect up to 2 cards over NVLink that had that interface (2070+). That restriction applies both to Windows, Linux, or any other OS environment, as well.
If you’re doing serious multi-gpu compute - you aren’t likely to be using Windows as your platform. Thus TCC is kind of a moot point. Peer-to-peer in Linux, however, is not.
Thank you for letting me know. I really appreciate it.
Dear Robert-Crovella,
thanks for your message. Do you know how to find out whether a specific model REALLY supports the TCC mode?
I am working with a Dell 7740 which has a Quadro RTX 5000 graphics card. I could install CUDA 11.1 but the nvidia-smi tool was not able to swith the card to TCC mode, just telling “not supported”. Does this message tell me an absolute no go or does the answer depend on the CUDA-Version installed?
Does anybody know whether a Quadra RTX 5000 (mobile) can support the TCC mode or not?
Kind regards
Thorsten Frenzel
