Sxynix Posted January 3, 2021 Posted January 3, 2021 The Nvidia GeForce RTX 3060 Ti brings the entry price for Ampere GPUs down to $400, because apparently the GeForce RTX 3070, GeForce RTX 3080, and GeForce RTX 3090 weren't selling out fast enough. Maybe that's not fair, but round four of the Nvidia Ampere architecture seems destined to repeat recent history and sell out within minutes of launch. The fact that this is now the best price to performance ratio in GPU land makes it one of the best graphics cards, and it currently occupies seventh place on our GPU Benchmarks hierarchy, not counting any Titan cards. Maybe if you've been really good (or you're really lucky), you'll be able to procure one this side of 2021. [Update: Yup, the RTX 3060 Ti sold out fast. Here's our advice on how to buy an RTX 3060 Ti (or other new Nvidia GPU).] Not a whole lot has changed relative to the RTX 3070. Both the RTX 3060 Ti and 3070 use the same GA104 GPU, but the 3060 Ti has fewer functional units enabled. You get the same 8GB of GDDR6 memory, clocked at the same 14Gbps. Meanwhile, the GPU core has 38 SMs enabled, versus 46 SMs on the 3070, making the 3060 Ti theoretically around 17 percent slower. Clock speeds also factor in a bit, as the boost clock on the 3060 Ti is 3.5 percent lower, so in general, the largest performance gap will be about 20 percent — and as we'll see, it's often quite a bit less than that. The specs tell the same story relative to the previous generation Nvidia Turing GPUs. The 3060 Ti has quite a bit more theoretical compute performance than the RTX 2080 Super, and more than doubles the compute performance relative to the RTX 2060 Super. We know GPU prices are all kinds of screwed up right now, but it's worth remembering that the 2060 Super had the same official $399 price as the 3060 Ti. Power requirements are slightly higher than the 2060 Super, but lower than the 2080 Super. Of course, theoretical specs don't always jive with real-world performance. One of the big changes with Ampere is that the INT32 cores can now do FP32 calculations as well. However, there's a decent amount of INT32 work in most games, so a good chunk of the potential FP32 performance is used up elsewhere. Still, with the other architectural changes, we expect the 3060 Ti to perform roughly on par with the 2080 Super, which means it will also be significantly faster than the 2060 Super. I didn't have good luck with the MSI Afterburner frequency scanner this round — it kept failing after 10-15 minutes or so of testing. That put us back to manually overclocking, which in practice usually means similar results, just with a bit more fiddling. We start by looping the Unigine Heaven 4.0 benchmark in a window at 1440p extreme, then bump up RAM clocks in 250 MHz increments until the GPU crashes or visual anomalies appear. Then we go back to the previous 'safe' level and repeat the same process for GPU clocks, except with 25 MHz increments. We were nearly able to max out the GDDR6 overclock at 16.5-17Gbps, but the EDR feature (Error Detection and Retry) meant performance was actually slower than a lesser RAM OC. Eventually, we stuck with a 1000 MHz overclock on the RAM (the best-case maximum would be in the 1000-1250 MHz range, based on our testing), giving an effective data rate of 16Gbps. For the GPU, 150 MHz was nearly stable but crashed in two tests, so we dropped down to a 140 MHz overclock that completed the full test suite. Besides changing the memory and GPU clocks, we also increased the power limit to the maximum 110 percent and adjusted the fan curve. That last one is critical for maintaining a stable overclock, as sticking with a default fan curve tends to lead to periodic crashes. Basically, we set the fan to a 50 percent minimum and gave it a steady ramp from 50 percent at 40C up to 100 percent at 80C. We could certainly do more tuning, but that was sufficient to keep temperatures in check and still wasn't overly loud. Advertisement Overclocked results are present in our standard benchmark charts, but not in the bonus extended tests or DirectX Raytracing (DXR) suites. We have several AIB 3060 Ti cards we'll be looking at in the future that have factory overclocks, so if you want something a bit more potent than the reference design, stay tuned
Recommended Posts