GPU for Deep Learning and AI

[ElMachoBurg]

Hi there,

I've recently been working on some convolutional neural networks and they take a massive amount of vram (16GB+) so I was thinking about purchasing a new GPU, I was thinking:

 

-Tesla K80 24GB, for 400€ second hand, which is very good value however I'm worried that as it is an older Kepler GPU and uses and older version of CUDA it may give me trouble or incompatibilities or some sort of error.

- 2 x RTX 2080ti as I've heard I might be able to add their vram together using NVLINK to pool the memory together but I am unsure about that.

 

I am aware that the performance of the 2 RTX 2080ti is much greater than that of a single K80 but im mostly interested in the vram rather that the speed.

 

So what would you recomend?

Has anyone managed to add vram using NVLINK?

Has anyone used a Tesla K80 card?

[Firewrath9]

nope, can't add ram toghether.

If not very particular about speed, then just use it on CPU, one of my friends who uses tensorflow for work has a server with 384GB of ram (apparently DDR3 ECC is cheap)

[Fasauceome]

1 minute ago, Firewrath9 said:

nope, can't add ram toghether.

Not with old SLI, but new NVLink can

Quote

On the GeForce RTX 2080 Ti, NVLink allows for up to 100GB/s bidirectional bandwidth between the cards (the GeForce RTX 2080 offers 50GB/s). When fully utilized, NVLink will minimize inter-GPU traffic over the PCI Express interface and also allows thememory on each card to behave more as a single, shared resource.

 

[Lord Mirdalan]

Kepler Quadro ~= GTX 800 series... I'd go with the RTX cards.

[Firewrath9]

2 hours ago, fasauceome said:

Not with old SLI, but new NVLink can

 

Quote

 

Do GeForce RTX Cards Support Memory Pooling in Windows?

Not directly. While NVLink can be enabled and peer-to-peer communication is functional, accessing memory across video cards depends on software support. If an application is written to be aware of NVLink and take advantage of that feature, then two GeForce RTX cards (or any others that support NVLink) could work together on a larger data set than they could individually.

 

depends on the software

AFAIK doesnt work with tensorflow

https://www.pugetsystems.com/labs/hpc/RTX-2080Ti-with-NVLINK---TensorFlow-Performance-Includes-Comparison-with-GTX-1080Ti-RTX-2070-2080-2080Ti-and-Titan-V-1267/#should-you-get-an-rtx-2080ti-or-two-or-more-for-machine-learning-work