Nvidia GeForce RTX 4060 Ti 8GB review: the disappointment is real
Meagre performance boosts and big questions remain over memory.
There's seven pages here of RTX 4060 Ti analysis and data in this review, but the more I used the card, the more I realised that this is massive overkill for a GPU where the salient points can be summed up in one paragraph. The RTX 4060 Ti is essentially a hybrid RTX 3060 Ti/RTX 3070 with the added bonus of DLSS 3, a good media block and extra efficiency, but beyond that it's a missed opportunity and a disappointment. There are some positive points to highlight, but they're crushed by the sense that Nvidia isn't delivering good value to the sector of the market that most demands it.
And this is actually a shame, because there are positive points to discuss. While gen-on-gen gains up against the RTX 3060 Ti are variable and VRAM allocation is an issue, the advantages of the Ada Lovelace architecture can still shine through. DLSS 3 frame generation's ability to overcome CPU limitations has palpable worth in the mainstream, as you'll see later on.
Perhaps more impressively, Cyberpunk 2077 RT Overdrive - a massive statement about the future of games - doesn't just run well on the RTX 4060 Ti at 1080p, I actually found that via the RT optimisation mod we discussed in the past, it actually runs slightly faster than RTX 4070 in an unmodded state at 1440p in DLSS balanced mode. A path-traced Cyberpunk 2077 running with better image quality and faster frame-rates than the Series X and PS5 versions in their performance modes is quite the thing to see and it's all coming from a $400 GPU. [UPDATE: Having put some time into this now, certain areas of the game see performance come crashing down at 1440p with frame-gen active and the solution is to use 1080p at DLSS quality mode instead to stop this happening - my feeling is that we hit memory limits with frame-gen and higher resolution working in concert.]
Cyberpunk 2077 RT Overdrive, 1440p, Balanced Upscaling
But elsewhere, it's very difficult to get excited about what Nvidia has delivered here - and the sense of a lack of momentum the lower down the stack we go is difficult to shake. Consider this: at the top of the range, RTX 4090 delivered features and performance we'd never seen before at around the same price as its predecessor. RTX 4080? Another superb GPU, but saddled with a price-point at least $300 higher than it should have been. RTX 4070 and and RTX 4070 Ti? These made sense from a strategic standpoint by delivering performance equivalent or better than RTX 3080, but in each case, the products were backed with DLSS 3 and more memory.
The 70-class could have rational arguments made for them - they were 'solid but unspectacular' - but it's much harder to do the same for the RTX 4060 Ti. There's no extra memory this time unless you want to wait a couple of months and pay a gigantic 25 percent premium for it. While RT performance is OK (usually in RTX 3070 territory), rasterisation performance varies significantly - but doesn't represent any kind of game-changing improvement up against the existing RTX 3060 Ti even in the best case. Push the card out of its comfort zone and it can actually be beaten by its predecessor. Yes, the use-case scenarios where this happens aren't a great match for the card, but any kind of gen-on-gen reduction in performance is a red flag.
Below you'll see how the RTX 40-series stack looks so far in terms of specs - with the 4060 Ti continuing the trend of Nvidia splitting its desktop graphics cards among a larger amount of GPUs. For example, RTX 3090 Ti to RTX 3060 Ti were spread across just its GA102 and G104 silicon, while RTX 4090 to 4060 Ti have been equipped with AD102, AD104 and AD106 GPUs - with the 4060 Ti getting the latter, with a modest CUDA core count and a 128-bit memory bus. On the memory front, Nvidia spent a great deal of effort explaining how the much larger L2 cache offsets the more narrow bus and as you'll see, this works to a certain extent - although bandwidth-heavy games suffer and it's why we see scenarios why the older 3060 Ti can occasionally best its successor.
RTX 4090 24GB | RTX 4080 16GB | RTX 4070 Ti 12GB | RTX 4070 12GB | RTX 4060 Ti 8GB | |
---|---|---|---|---|---|
Processor | AD102 | AD103 | AD104 | AD106 | AD106 |
Transistors | 76.3B | 45.9B | 35.8B | 35.6B | 22.9B |
Die Size | 608mm² | 379mm² | 295mm² | 295mm² | 190mm² |
CUDA Cores | 16384 | 9728 | 7680 | 5888 | 4352 |
Boost Clock | 2.52GHz | 2.51GHz | 2.61GHz | 2.475GHz | 2.535GHz |
Memory Interface | 384-bit | 256-bit | 192-bit | 192-bit | 128-bit |
Memory Bandwidth | 1018GB/s | 742GB/s | 557GB/s | 504GB/s | 288GB/s |
TGP | 450W | 320W | 285W | 200W | 160W |
PSU Recommendation | 850W | 750W | 700W | 650W | 550W |
PSU Cables | 4x 8-pin | 3x 8-pin | 2x 8-pin | 2x 8-pin | 2x 8-pin |
Base Price | $1499/£1649 | $1199/£1199 | $799/£799 | $599/£589 | $399/£389 |
Release date | October 12th, 2022 | November 16th, 2022 | January 5th, 2023 | April 13th, 2023 | May 23rd, 2023 |
In terms of spec comparisons against the outgoing RTX 3060 Ti, there's a reduction in CUDA core count from 4864 to 4352, a 128-bit memory interface vs a wider 256-bit equivalent on the old product (offset by the massive L2 cache increase), while boost clocks rise from 1665MHz to 2535MHz. The 8nm Samsung-fabbed die reduces from 293mm2 to just 190mm2 on the new TSMC 4nm node, while transistor count rises from 17.4 billion to 22.9 billion. That's a fairly modest rise overall and while the new process used by Nvidia is expensive, we do have to wonder whether this chip is more expensive to produce than its predecessor when it's so much smaller.
All of which brings us to the big question of VRAM allocation and the topic of heated discussion between us and Nvidia not just in the product briefing but in a series of follow-up emails. Let's make it perfectly clear where we stand on this. Developers should do their best to make their console ports work on 8GB GPUs and we shouldn't be seeing situations like the launch of Forspoken, Hogwarts Legacy and The Last of Us Part 1, where the 8GB experience was awful. But at the same time, if you're buying a GPU in 2023 with the aim of matching or exceeding console visuals, or allowing triple-A games to run with the best textures, it's pretty clear that a 12GB minimum is needed as we finally leave the cross-gen period behind.
Put simply, the RTX 3060 Ti offered rasterisation and RT performance that were faster than the consoles and naturally, the even-faster RTX 4060 Ti should follow suit - but it'll need the memory to get the job done fully, not just to house the larger footprint of higher resolution textures, but also to host the extra features we should expect from PC gaming. Ray tracing, for example, requires a good degree of extra memory for the BVH structures required to make the feature work. Nvidia talks about the RTX 4060 Ti as a 1080p card - and it's true that render targets reduce in size with lower resolutions, but textures don't and neither do BVH structures. As things stand, I'd contend that if the RTX 4060 Ti can run Cyberpunk 2077 RT Overdrive as well as it does at 1440p, I'm not sure that the label of 'a 1080p card' best reflects the GPU's capabilities - but it does help to justify the 8GB memory allocation.
GeForce RTX 4060 Ti | GeForce RTX 3060 Ti | GeForce RTX 2060 Super | |
---|---|---|---|
Control, 1440p, High RT | 39.8fps/147W - 3.69 Joules per Frame | 38.7fps/198W - 5.11 Joules per Frame | 25.8fps/171W - 6.63 Joules per Frame |
Dying Light 2, 1440p, Ultra RT | 44.4fps/145W - 3.27 Joules per Frame | 41.6fps/197W - 4.74 Joules per Frame | 25.6fps/169W - 6.60 Joules per Frame |
Forza Horizon 5, 1440p, Extreme, RT Off | 85.4fps/109W - 1.28 Joules per Frame | 82.9fps/189W - 2.28 Joules per Frame | 54.6fps/155W - 2.83 Joules per Frame |
Hitman 3, 1440p, Max, RT Off | 155.7fps/146W - 0.94 Joules per Frame | 138.4fps/198W - 1.43 Joules per Frame | 96.6fps/173W - 1.79 Joules per Frame |
The Ada Lovelace architecture has delivered some remarkable power efficiency metrics in previous reviews and it continues to do so here as we look at efficiency of the 4060 Ti compared to similar products from the last two generations. We calculate efficiency by benchmarking all games with frame-rate fully unlocked, then divide power by performance, giving us a value defined by joules per frame. The lower that is, the better.
With Control, there's a 28 percent reduction in power per frame up against the RTX 3060 Ti, rising to 44 percent up against the RTX 2060 Super - a strong start for a game where the generational gains in performance are not spectacular. There's a similar showing with Dying Light 2, with a 31 percent reduction in power per frame compared to the last-gen RTX 3060 Ti, rising to a 50.5 percent reduction up against RTX 2060 Super. These two titles cover off our RT testing.
Moving onto rasterisation, Forza Horizon 5 has been an exceptionally efficient performer on all Ada cards and in all cases, it's never troubled the power limit of the GPU - which can't be said of prior cards. The same thing happens here with similarly spectacular results. Power reduction per frame is in the region of 44 percent up against the RTX 3060 Ti, rising to 55 percent when compared against RTX 2060 Super. With Hitman running at max settings without RT, there's a 34 percent reduction in power per frame against the 4060 Ti's immediate predecessor, which boosts up to a 47 percent reduction when we compare it against the RT 2060 Super.
That's one of the biggest areas of good news with this GPU, but now it's time to move onto raw performance numbers, where bench all cards using a Core i9 13900K at stock settings, paired with 6000MHz DDR5 from GSkill.
Nvidia GeForce RTX 4060 Ti analysis
- Introduction, hardware and power analysis [This Page]
- RT benchmarks: Dying Light 2, Cyberpunk 2077, Control
- RT benchmarks: Hitman 3, Metro Exodus Enhanced Edition, F1 22
- RT/DLSS/FSR2/DLSS3 benchmarks: Cyberpunk 2077, Dying Light 2, Forza Horizon 5
- Game benchmarks: Control, Cyberpunk 2077, F1 22, Forza Horizon 5
- Game benchmarks: Hitman 3, A Plague Tale: Requiem, Returnal
- Nvidia GeForce RTX 4060 Ti: the 8GB problem and the Digital Foundry verdict