At CES 2021, NVIDIA launched a GPU that is going to interest many STH readers: the NVIDIA GeForce RTX 3060. While the “60-series” is not the most powerful GPU of a generation, it is also one of the lower-cost options. Even though NVIDIA needed a GPU to slot below the RTX 3060 Ti, we were a bit surprised by this launch for a few reasons.
NVIDIA GeForce RTX 3060 12GB
The overall CUDA core count is well below the GeForce RTX 3060 Ti and RTX 3070, but so is the price. The card is “starting at $329” and available “end of February”. We put these into quotations because both MSRP and availability of the GeForce RTX 3000 series has not been the same as with other launches. We have almost no chance to buy them where I live.
The memory situation is very interesting indeed. While we move down from a 256-bit to a 192-bit bus, we are still using GDDR6 but with 12GB on this model compared to 8GB on the Ti model. While we get less compute capacity and memory bandwidth, we get much more memory capacity. NVIDIA even explicitly calls out the fact that this is a big deal for many non-gaming workloads:
“The GeForce RTX 3060 is built with a 192-bit bus, which allows us to equip this latest addition to the 3060 family with 12 GB of fast GDDR6 memory. The 12GB of faster GDDR6 also makes it a great choice for creators working with large datasets, models, and detailed scenes.” (Source: NVIDIA PR)
For anyone who is more memory capacity constrained than bandwidth or compute capacity constrained, the RTX 3060 is going to offer an interesting value proposition. This is more memory than the RTX 3060 Ti, RTX 3070, RTX 3080, and even older higher-end cards such as the RTX 1080 Ti and RTX 2080 Ti. 12GB plus CUDA for $329 is absolutely going to be interesting for many in the STH community just because it is something different.
NVIDIA GeForce RTX 3060 12GB Specs
Here is the spec table for the NVIDIA GeForce RTX 3060 12GB compared to the 3060 Ti.
We are going to note here that while the RTX 3060 Ti is a 200W part, the RTX 3060 is a lower power part at 170W.
Final Words
Overall, we really like this direction, but it does raise a good question: memory on the rest of the line. With the RTX 3060 launch at 12GB and the RTX 3090 at 24GB, the other cards in the 8-10GB range seem like they are out-of-line in terms of memory capacity. As a result, we expect NVIDIA will have to bump memory capacity in future iterations of those cards. It is likely not in a rush to do so as it cannot make enough of what it has and adding more memory seems like a cost NVIDIA does not need to include at this point. Still, that makes it even more curious why we have a 12GB card in this segment while the RTX 3080 only has 10GB onboard.
Cool! Looks like NV is also reading Tim Dettmers and his advices on what GPU to use for DL-based AI. :-) https://timdettmers.com/2020/09/07/which-gpu-for-deep-learning/
Sure I’m excited, but I doubt they’ll be $329 in seven weeks and I’ll be able to buy one (or 20)
Think any manufacturers will put out blower models? Maybe with hot clocked memory chips?
But yes, this begs the question of what Nvidia is going do with models between the 3060 and the 3090.
Would it cannibalize pro cards? Are they feeling pressure from the high-capacity AMD cards and consoles?
EJ – I think these are very much in the realm of being able to be cooled by blower models.
Single slot blower 3060 12G with underclocked TGP of 150w and horizontally mounted 8pin connector would be awesome.