A few days ago, NVIDIA presented during its annual presentation at CES your new RTX 4070 Ti, which came as no surprise since there had been rumors for some time that the “cancelled” RTX 4080 12GB had been rebranded under its new name. Our impressions? It completely put us off, that is to say, it is the first time that a launch of the brand does not excite us.
One of the consequences of the end of the mining bubble is that the world has suddenly become filled with older generation graphics cards, both on store shelves and on the second-hand market. This collided with the launch of a new generation of graphics cards with technical capabilities superior to the previous one. The way to solve such a mess? Do not associate the price of graphics cards with their actual cost, but with their performance compared to what exists in the market. The consequences of this? A generation of graphics cards with a price well above the average.
What problem do we see with the RTX 4070 Ti?
We must start from the fact that for us it is not a technical problem, but rather the cost of the graphics card for the end user. It looks like these were designed when the GPU mining bubble was active and priced for a market that no longer exists. We’re not saying it just to say it, although we’ll review its official specs first so you can see the issue we’re referring to.
| Graphic card | RTX4070Ti |
|---|---|
| manufacturing node | TSMC N4 |
| MHz base | 2310 |
| MHz increase | 2610 |
| FP32 Units (CUDA Cores) | 7680 |
| Core (SM) | 60 |
| Ray Tracing Units | 60 |
| AI units (Tensor Cores) | 240 |
| Hidden L1 | 128 KB per core |
| Hidden L2 | 48 MB |
| VRAM type | GDDR6X |
| VRAM bus | 192 bit |
| VRAM quantity | 12 GB |
| VRAM Bandwidth | 504 GB/s |
| PDT | 285W |
We found the key to everything on your 192 bit busThis is the first time we’ve seen a 70-terminal graphics card where the memory bus is so tight. Which also shows us that we are dealing with a chip that is much smaller than the average and since the cost of chips is linked to their size, we can immediately deduce that if we are not in the situation where the market is currently said price would not be justified at all.
And we say it to say it? No, because the previous generation graphics card with a bus of this size was the RTX 3060. However, aside from all the bad, one must start from the fact that the RTX 4070 Ti sweeps the floor with said graphics card. on all performance measures. It is much better in performance per watt and performance per mm2 of the GPU, relative to said graphics card. However, Nvidia decided to take as a benchmark not the size of the components or their configuration, but the relative performance of other graphics cards on the market.
At the performance level of an RTX 3090
Obviously, those at NVIDIA are marketing geniuses and they have no interest in comparing their new RTX 4070 Ti with the RTX 3060, even if they come out on top in all the comparisons and that’s normal from the moment the official sale price of 910 $ on the European market, i.e. double. So they will look to compare it with something in that price range. And they did it with an RTX 3090 Ti.
However, looking at its technical specifications, it is clear that it is more at the level of an RTX 3090. And no, we don’t just say it. Although it is better to leave you a table comparing these graphics cards so that you can see that our statement is not free.
| Graphic card | RTX4070Ti | RTX-3090 |
|---|---|---|
| manufacturing node | TSMC N4 | Samsung 8nm |
| MHz base | 2310 | 1395 |
| MHz increase | 2610 | 1695 |
| FP32 Units (CUDA Cores) | 7680 | 10496 |
| TFLOPS (FP32) (minimum) | 35.4 TFLOPS | 29.3 TFLOPS |
| TFLOPS (FP32) (maximum) | 40 TFLOPS | 35.5 TFLOPS |
| Core (SM) | 60 | 82 |
| Ray Tracing Units | 60 | 82 |
| AI units (Tensor Cores) | 240 | 328 |
| Texture units | 240 | 328 |
| Textured Tasa (Min) | 554 Gtexels/s | 457.6 Gtexels/s |
| Textured Tasa (Max) | 626 Gtexels/s | 555 Gtexels/s |
| Hidden L1 | 128 KB per core | 128 KB per core |
| Hidden L2 | 48 MB | 6 MB |
| VRAM type | GDDR6X | GDDR6X |
| ROPS | 80 | 112 |
| Fill rate (min) | 184.6 GPixels | 156 GPixels/s |
| Maximum filling rate | 208.8 GPixels | 190 GPixels/s |
| VRAM bus | 192 bit | 384 bit |
| VRAM quantity | 12 GB | 24 GB |
| VRAM Bandwidth | 504 GB/s | 912 GB/s |
| PDT | 285W | 350W |
By using TSMC’s N4 node, Nvidia managed to create a graphics card with costs, for them, lower than an RTX 3070, but with performance on par with an RTX 3090 in Boost, but without having to force the machinery. We are therefore dealing with a card that is totally valid for 4K due to its technical specifications. Even with its shortened bus, NVIDIA has therefore done an excellent job with this graphics card. The only negative point we see is its price, which is out of the market, and the fact that it does not have a better memory configuration, both in quantity and in bandwidth.
