Introducing the Nvidia GeForce RTX 5070 Ti
The Nvidia GeForce RTX 5070 Ti marks the third entry for the Blackwell RTX 50-series GPUs, which Nvidia officially unveiled at CES 2025 in early January. It takes over from both the original RTX 4070 Ti as well as the newer RTX 4070 Ti Super, with the latter having replaced the 4070 Ti last January. The 5070 Ti looks better against the 4070 Ti, while it’s more of an incremental upgrade from the 4070 Ti Super. The good news is that its MSRP is also $50 cheaper than the launch prices on the 4070 Ti / Super cards, coming in at $749. Faster, cheaper, and new features is a great way to make it onto our list of the best graphics cards — assuming there’s sufficient supply, which remains to be seen.
Nvidia is allowing reviews of the base-MSRP 5070 Ti models today, with higher-priced variants tomorrow (and after). The official launch date for all RTX 5070 Ti cards is February 20, 2025, as well, so you can’t buy one until tomorrow. We were told that there should be quite a lot more RTX 5070 Ti cards at launch than either the RTX 5090 or RTX 5080, but even so, we anticipate the first batch at least will sell out quickly, and prices will likely head north in the short term before coming back down (we hope).
You might think that a month after the RTX 5090 launch, things would have quieted down, but that’s not really true. In addition to trying to test some third-party AIB (add-in board) cards, we’ve had several new drivers, the 5080 launches, and unfortunately, one of our sample AIB cards has been finicky (to say the least). We’re still looking for a solution, which may involve sending the card back for a different sample. The 50-series launch hasn’t gone smoothly, in other words.
We did a lengthier deep-dive into DLSS 4 and MFG, using the 5080 and 5090, to get a better understanding of what the tech does and doesn’t offer. The short summary is that MFG is a lot like framegen, with even more marketing hype. When it gives a straight doubling in performance, from 60 to 120 fps, or from 120 to 240 fps in the case of MFG, you can make a good argument that it looks and feels “better.” But typically, it’s more like a 50~80 percent improvement, making comparisons more difficult, and if it’s less than a 50% increase, it can end up feeling worse.
We haven’t had sufficient time to do additional MFG testing on the 5070 Ti (yet), but we’ll flesh out that section of this review in the coming days. While the official Nvidia party line is that MFG makes the 5070 Ti “twice as fast” as the previous generation, that’s a gross exaggeration. It might be able to spit out twice as many frames — 1x rendered and 3x generated — compared to the 4070 Ti, but the actual feel of games with MFG doesn’t improve nearly as much as those inflated numbers would suggest.
For additional information about Nvidia’s latest Blackwell GPUs, check the links in the boxout. The RTX 5070 Ti continues the same pattern, just with fewer GPU cores and less performance than the 5080 and 5090. It offers the same feature set as the other Blackwell GPUs, like FP4 number format support (for AI) and MFG (Multi Frame Generation) for gaming. It also has the same 16GB of GDDR7 memory as the 5080, though the memory is clocked slightly lower. Let’s start with the specs table and discuss how it stacks up.
| Graphics Card | RTX 5070 Ti | RTX 4070 Ti Super | RTX 4070 Ti | RTX 3080 | RTX 3070 Ti | RTX 2070 Super |
|---|---|---|---|---|---|---|
| Architecture | GB203 | AD103 | AD104 | GA102 | GA104 | TU104 |
| Process Technology | TSMC 4N | TSMC 4N | TSMC 4N | Samsung 8N | Samsung 8N | TSMC 12FFN |
| Transistors (Billion) | 45.6 | 45.9 | 35.8 | 28.3 | 17.4 | 13.6 |
| Die size (mm^2) | 378 | 378.6 | 294.5 | 628.4 | 392.5 | 545 |
| SMs | 70 | 66 | 60 | 68 | 48 | 40 |
| GPU Shaders (ALUs) | 8960 | 8448 | 7680 | 8704 | 6144 | 2560 |
| Tensor Cores | 280 | 264 | 240 | 272 | 192 | 320 |
| Ray Tracing Cores | 70 | 66 | 60 | 68 | 48 | 40 |
| Boost Clock (MHz) | 2452 | 2610 | 2610 | 1710 | 1765 | 1770 |
| VRAM Speed (Gbps) | 28 | 21 | 21 | 19 | 19 | 14 |
| VRAM (GB) | 16 | 16 | 12 | 10 | 8 | 8 |
| VRAM Bus Width | 256 | 256 | 192 | 320 | 256 | 256 |
| L2 Cache | 48 | 64 | 48 | 5 | 4 | 4 |
| Render Output Units | 96 | 96 | 80 | 96 | 96 | 64 |
| Texture Mapping Units | 280 | 264 | 240 | 272 | 192 | 160 |
| TFLOPS FP32 (Boost) | 43.9 | 44.1 | 40.1 | 29.8 | 21.7 | 9.1 |
| TFLOPS FP16 (FP4/FP8 TFLOPS) | 352 (1406) | 353 (706) | 321 (641) | 238 | 174 | 72 |
| Bandwidth (GB/s) | 896 | 672 | 504 | 760 | 608 | 448 |
| TGP (watts) | 300 | 285 | 285 | 320 | 290 | 215 |
| Launch Date | Feb 2025 | Jan 2024 | Jan 2023 | Sep 2020 | Jun 2021 | Jul 2019 |
| Launch Price | $749 | $799 | $799 | $699 | $599 | $499 |
In terms of raw specs, the big changes are in memory speed (and capacity versus the 4070 Ti) plus the FP4 tensor support. There are other architectural differences that we’ve discussed elsewhere, like how the CUDA cores are now all “full citizens” and support both FP32 and INT32 (only half the cores in Ada and Ampere supported INT32), and there are other changes to support new neural rendering techniques. But forget all that for a moment and just look at the die sizes and transistor counts.
RTX 5070 Ti offers a decent step up from the original 4070 Ti, but compared to the 4070 Ti Super, it looks extremely similar. They both have the same number of transistors and die size, using the same TSMC 4N process node. The transistor counts aren’t necessarily 100% accurate, as they’re more of a rough estimate, but it does suggest that the architectural changes may not be that significant. It’s also possible some things were dropped or simplified to make room for new features. What those things might be is more difficult to pin down.
Getting back to the specs, total compute performance looks basically the same as the 4070 Ti Super: 44 TFLOPS FP32 and 352–353 TFLOPS FP16 on the tensor cores. Clock speeds are also lower, on paper, with the 5070 Ti. As usual, however, Nvidia’s stated boost clocks are quite conservative estimates, and most of the games we tested ran at much higher speeds.
Memory speed and bandwidth are both quite a bit higher with the 5070 Ti. It’s 33% faster than the 4070 Ti Super in speed and bandwidth and 78% higher in bandwidth compared to the 4070 Ti, thanks to having a 33% wider interface. And, of course, it also offers 33% more VRAM capacity than the 4070 Ti.
We tossed a few earlier GPUs from the 30- and 20-series into the table for reference. If you’re still using something like a 3070 Ti or 2070 Super, the 5070 Ti should offer a pretty massive improvement in performance. Is it enough to get people to upgrade? For some, we’re sure the answer will be yes, but if you’re already sporting something like an RTX 4070 Ti Super or even an RTX 4070, this will likely be an easy generation to sit out.
Basically, you get improved AI features, including MFG support. You also get upgraded display outputs (DisplayPort 2.1 UHBR20) and a PCIe 5.0 x16 interface, though neither of those is likely to matter much for most consumer workloads. Power requirements are a bit higher than the 4070 Ti Super, with a TGP of 300W compared to 285W on the prior generation. (TGP stands for “Total Graphics Power,” and it’s what Nvidia calls the power of the entire graphics card, including all components. AMD uses the term TBP for Total Board Power, which ends up being mostly the same thing with a few minor nuances.)
You also get the much-disparaged 16-pin power connector, with a 3x 8-pin adapter, should you need it. This is the newer 12V-2×6 standard, but even that hasn’t proven 100% reliable with the 50-series cards. The 5070 Ti shouldn’t have any meltdown issues, considering it’s only rated for 300W. That’s better than the 5090, which seems to be following at least partially in the footsteps of the 4090. Pulling 600W through such a design just seems ill-advised, and the lack of more advanced monitoring functions hasn’t helped matters.
Nvidia takes a staggered approach to GPU launches every generation, typically starting at the top and working its way down. The 5070 Ti is the third card from the Blackwell series, using the same GB203 GPU as the 5080 but with only 83% of the GPU cores enabled. Building redundancy and the ability to down-bin chips into a GPU design is common practice, so while the 5080 requires a basically perfect chip that’s fully enabled, anything that can’t meet the criteria can potentially be used in a 5070 Ti — and in the future, we’ll undoubtedly see some 5070 or even 5060 Ti variants that use “leftover” GB203 chips.
Performance on paper may be up to 17% slower than the 5080, give or take, for 25% less money. And for games and tasks that are more memory dependent, the gap may only be 5~10 percent. At lower resolutions and settings, everything starts to become CPU and platform-limited, but we don’t think most people buying a $750 to $1,000 GPU will want to run such low settings.
Let’s check out the Asus RTX 5070 Ti Prime in more detail before hitting the benchmarks.
