Hacker News Reader

Comment by causality0 | Hacker News Reader

The stand-out thing to me is that the RTX 3070 only comes with 8GB of RAM. The GTX 1070 I bought five years ago with an MSRP $100 less than the 3070 (and that I was able to purchase at MSRP) has 8GB of RAM. The GPU I bought five years before that only had 1.28GB and before that 256MB. What the hell has the industry been doing for the last half decade?

View on HN

Limiting the RAM so AI researchers buy Quadro/Titan cards.

Just like Canon that downgraded their DSLRs to protect their cinema cameras, NVIDIA will loose shares with enthusiasts while AMD (which got solid TF/PyTorch support recently) can finally catch up.

> Just like Canon that downgraded their DSLRs to protect their cinema cameras, NVIDIA will loose shares with enthusiasts while AMD (which got solid TF/PyTorch support recently) can finally catch up.

Is this for real? As in: I'm thinking about starting to do model training with TensorFlow. I can buy an AMD GPU to use for this and not regret it later?

Here's a tutorial : https://medium.com/analytics-vidhya/install-tensorflow-2-for...

I haven't tested myself, but since I have PTSD from Linux CUDA installations I'm glad competition is finally here.

whether or not the drivers work for quick demos/benchmarks is not really what I care about with AMD, tbh, their ML pipeline sw QA is bad, historically the major problem was long-term stability. Segfaulting 12 hours into a run is not good for an ML practitioner's sanity.

I always thought that the marketing jargon "cannibalize your market" just meant "create a kick ass product"

I hate marketing.

You will be happy to know that the RTX 3060 has 12GB of VRAM. I try BERT-Large training on it here:

https://www.naut.ca/blog/2021/03/16/rtx-3060-vs-jetson-agx-f...

True but without comparable offerings like CUDA, it's not really looking bright tbh. Open CL and open vino are years behind virtually speaking with respect to support.

The 1070 had GDDR5 RAM and the 3070 has GDDR6X RAM. Instead of greater capacity, we get greater bandwidth.

https://graphicscardhub.com/wp-content/uploads/2017/01/gddr6...

Thirteen years later and it's less than three times as fast. This is just sad.

I believe the non-RTX (important, confusing with nvidia, as usual) 3060 has 12 GB of ram, which makes it a nice buy for many ML tasks.

non-RTX 3060? I don't think that exists. Got a link to that?

whoops. Thanks for correcting me *non-TI. Can't edit parent post. And also note some non-TIs have 10GB, not 12 (still more than 8GB). Nvidia has successfully confused me =D.

Reply on news.ycombinator.com