Hacker News new | ask | show | jobs
by popcorncolonel 3102 days ago
Quick question: Why? What does NVIDIA have to gain by forbidding people from using GeForce for ML in datacenters?
4 comments

Price discrimination. Because they aren't getting effective competition from AMD right now they can sell to multiple tiers of customer for different prices. AMD is competitive in crypto, so they exempted crypto from the ban.
No, it's only because the gp106/gp104 purpose built mining cards use geforce drivers.
Seems like there would just be an exemption for them otherwise.
The Titan V is $3000 (GeForce)

The Tesla V100 is $17,000: http://www.nextwarehouse.com/item/?2782569_g10e

We're talking the difference of ~5.5x the price here for otherwise similar cards.

The V100 is $8000 retail. The site you linked is incorrect.
It probably depends a lot. There is the SXM vs. PCIe distinction plus discounts for bulk and all kinds of negotiations.
No, it's $8000 if you want to buy one now. https://www.thinkmate.com/product/nvidia/900-2g500-0000-000
When I was running my p-2-p startup I wanted to try capitalizing on this. It's the reason why GPU instances cost an arm and a leg.
But you can already not buy a datacenter's worth of GeForce; they have limitations on the quantity per customer. That seems like a far more effective measure than this clause.
And yet, many datacenters run hundreds or thousands of GeForce cards. Hetzner, for example, rents out an i7-6700, GTX 1080, 64GBRAM for 117€/month
I've spent some time training tensorflow cnns on nvidia 1080gtx.

It works pretty well, but i couldn't reliably train production models for work on it. it's just too flaky. I mentioned this to our 'trustworthy' dell rep who suggested that a good v100 suite would surely solve all the reliability problems, what with it's greatly increased memory bus, or something...

You mean your algorithm is too flaky. The hardware is fine. A V100 won't fix your convergence problems.
yes. salesman bullshit. i think it was tensorflow that was flaky.
CNN models being "flaky" on GeForce hardware isn't something I've heard of? NVIDIA has made some deliberate decisions to make GeForce cards less attractive for deep learning in terms of performance, but I don't think making them produce incorrect results is in their best interest. What hardware did you test this against?
NVIDIA is planning on standing up its own deep learning datacenters to compete with amazon, GCP, Azure, because instead of selling a commodity, they'd rather sell a service (better margins).
They want to push their expensive enterprise cards (Tesla series) which are exempt from this rule (applies only to "GeForc" cards) and they are probably getting ready to push their own ML service with their own datacenters.