Nvidia's GeForce RTX 4070 Ti may rank as the worst kept "secret" of the past few decades of graphics cards. Originally revealed back in October as the RTX 4080 12GB that was later "unlaunched," the core specs and hardware didn't change — only the price and name are new. Following in the footsteps of the more potent RTX 4090 and RTX 4080, the cards officially go on sale tomorrow, January 5, 2023. Except reviews can go up a day early, so potential buyers at least get some advice before taking the plunge.
Will the RTX 4070 Ti rank among the best graphics cards? That's a tough call, which we'll get into more during this review. What we can say is that, based on our GPU benchmarks hierarchy testing, it's faster than most of the 30-series offerings and basically ties the RTX 3090, at a lower price point. That's the good news. The bad news is that this is by far the most expensive xx70-class graphics card Nvidia has ever released. We've gone from $329 on a GTX 970 to $379 on the GTX 1070, then $499 for the RTX 2070 and 2070 Super, $499 for the RTX 3070 as well with $599 for the 3070 Ti, and now we're looking at $799 for an RTX 4070 Ti.
Put that way, what we're really talking about feels a lot more like a replacement for the RTX 3080 12GB. Nvidia never gave that an official MSRP, but we briefly saw prices dip into the $700–$800 range before supply apparently dried up. Absent the cryptocurrency boom of 2020–2022, the 3080 12GB and 3080 Ti probably would have landed in the $800–$900 range. Now Ethereum has gone proof of stake and cryptocurrency mining profits are in the toilet, but it seems we're still looking at mining-inflated graphics card prices.
There are alternative ways of looking at things as well. We don't have hard numbers on how many cards have been sold, but all of the RTX 4090 cards at retail are still selling for over $2,000. RTX 4080 cards meanwhile seem to be relatively available at close to their base $1,200 price, but none of them are actually selling at or below that price point. The same goes for AMD's new RX 7900 XTX and RX 7900 XT, the latter of which can be found for $899 while the former is mostly sold out and often costs $1,200 or more.
In other words, there's clearly enough demand from people with deep pockets that newer, faster graphics cards at lower prices is a fantasy we're not going to experience. We've heard some speculation that RTX 4090 prices in particular are greatly inflated due to demand from the professional sector — if all you want is a number crunching monster for AI research, the 4090 at $2,000 is actually a great deal compared to $5,000–$10,000 for actual professional cards that may not even be as fast! But how far down the stack will that go before such professionals simply opt for a faster card?
Hopefully not down to the $800 range, and equally hopefully we won't get scalpers trying to buy up every card and then resell them on eBay. But we probably need a lot more than hope for either of those to be likely.
We've previously covered Nvidia's Ada Lovelace architecture, so the only real question now is how performance scales down to the AD104 GPU with fewer shaders, less memory, less cache, a narrower memory interface, etc. Leaked details from the rebadged RTX 4080 12GB suggest it will be around 20–25 percent slower than the RTX 4080… which puts it in RTX 3090 territory. Here's a look at the specifications for the new and previous generation Nvidia cards, alongside AMD's new RX 7900 offerings.
|Graphics Card||RTX 4070 Ti||RTX 4090||RTX 4080||RTX 3090 Ti||RTX 3080 Ti||RTX 3080||RTX 3070 Ti||RX 7900 XTX||RX 7900 XT|
|Architecture||AD104||AD102||AD103||GA102||GA102||GA102||GA104||Navi 31||Navi 31|
|Process Technology||TSMC 4N||TSMC 4N||TSMC 4N||Samsung 8N||Samsung 8N||Samsung 8N||Samsung 8N||TSMC N5 + N6||TSMC N5 + N6|
|Transistors (Billion)||35.8||76.3||45.9||28.3||28.3||28.3||17.4||45.6 + 6x 2.05||45.6 + 5x 2.05|
|Die size (mm^2)||294.5||608.4||378.6||628.4||628.4||628.4||392.5||300 + 222||300 + 185|
|Ray Tracing "Cores"||60||128||76||84||80||68||48||96||84|
|Boost Clock (MHz)||2610||2520||2505||1860||1665||1710||1765||2500||2400|
|VRAM Speed (Gbps)||21||21||22.4||21||19||19||19||20||20|
|VRAM Bus Width||192||384||256||384||384||320||256||384||320|
|TFLOPS FP16 (FP8/INT8)||321 (641)||661 (1321)||390 (780)||160 (320)||136 (273)||119 (238)||87 (174)||123 (123)||103 (103)|
We've mentioned before that Nvidia appears to be attempting to stretch out the range of performance offered with the RTX 40-series, and the same thing happened with the RTX 30-series. From the fastest to slowest RTX 20-series GPU, the RTX 2080 Ti was about double the performance of the RTX 2060. Meanwhile, last generation's RTX 3090 Ti was three times as fast as the RTX 3050. At present, we don't even have a vanilla RTX 4070, never mind a future RTX 4050, and we're already looking at potentially double the performance from the bigger and more expensive RTX 4090.
Things of course won't scale perfectly, but the 4090 does have more than twice as many GPU shaders, Tensor cores, RT cores, etc. It also has exactly twice as much memory and memory bandwidth, but only 50% more L2 cache. There will be workloads — stuff like AI training and inference comes to mind — where the card that costs twice as much will deliver twice the performance while consuming twice as much power. But we'll get to the benchmarks soon enough.
What about AMD's RX 7900 series cards? The RX 7900 XT costs $100 more than the 4070 Ti, and at least in several areas it looks to have superior specs. The teraflops figures favor AMD's GPU by nearly 30%, for example, and AMD gives users 67% more memory with 59% more memory bandwidth. There's little doubt that Nvidia will still pull ahead when it comes to ray tracing performance — and deep learning workloads — but for rasterization AMD might come out ahead. But then we also have to ask: Do you want to spend nearly $1,000 for a graphics card that relegates ray tracing to second tier status?
Sure, we get it: There still aren't many games where ray tracing is required, and quite a few games that support ray tracing only seem to use it for modest improvements in image fidelity while dropping performance quite a bit. But there are games like Control and Cyberpunk 2077 where it can make a difference, so spending a bit more — or in this case, potentially saving money — to get a superior feature set is something plenty of gamers seem willing to do.
Besides all the usual graphics hardware, Nvidia's latest Ada Lovelace GPUs have a few extras. AMD still doesn't have a Tensor core equivalent — its AI Accelerators share the same execution units as the GPU shaders — which means in FP16 workloads the 4070 Ti has potentially triple the throughput of the RX 7900 XT. And for AI inference work that can leverage the FP8 hardware, Nvidia doubles performance yet again.
There's also the enhanced Optical Flow Accelerator (OFA), which powers the DLSS 3 algorithm and may be used for other video related tasks. Nvidia says that over 50 games with DLSS 3 are in the works, with about 15 currently available. Even better news is that every one of those games with DLSS 3 also supports DLSS 2 and Reflex, which means anyone with an RTX card can benefit — just not with Frame Generation.
If you're interested in AI research, especially without forking over the money for a professional workstation, Nvidia's GeForce cards certainly warrant a look. There are quite a few AI tools already available, including things like Stable Diffusion, ChatGPT, and more. It's possible to get those tools running on non-Nvidia GPUs, but after months of looking for a good AI benchmark, I can say that many of the open source tools focus on Nvidia while AMD (or Intel) GPUs are at best an afterthought.