La Nvidia GeForce RTX 3070 Ti poursuit la Déploiement de l’architecture Ampere, qui alimente les GPU derrière la plupart des meilleures cartes graphiques. La semaine dernière, Nvidia a lancé la GeForce RTX 3080 Ti , une carte qui, selon nous, augmentait trop le prix par rapport à la prochaine étape. Le RTX 3070 Ti devrait faire mieux, à la fois parce qu’il ne coûte que 599 $ (en théorie), et aussi parce qu’il y a jusqu’à 33 % de différence entre les GeForce RTX 3070 et GeForce RTX 3080. C’est une augmentation de prix de 100 $ par rapport au 3070 existant, mais les 3070 et 3080 continueront à être vendus, dans des versions à”taux de hachage limité”, pour le moment. Nous ajouterons le RTX 3070 Ti à nos tests GPU hiérarchie sous peu, si vous voulez voir comment tous les GPU se classent en termes de performances.
L’idée de base derrière le RTX 3070 Ti est assez simple. Nvidia prend le GPU GA104 qui alimente les RTX 3070 et RTX 3060 Ti, mais cette fois, il s’agit de la variante 48 SM complète de la puce, et l’associe à la GDDR6X. Alors que Nvidia aurait pu essayer de le faire l’année dernière, les RTX 3080 et RTX 3090 avaient déjà du mal à obtenir suffisamment de mémoire GDDR6X, et un retard de neuf mois a permis à Nvidia de constituer un inventaire suffisant du GPU et de la mémoire pour ce lancement. Nvidia a également implémenté son limiteur de hachage Ethereum, réduisant de moitié les performances de minage sur les pièces cryptographiques utilisant l’algorithme Ethash/Dagger-Hashimoto.
Est-ce que cela suffira pour éviter que les cartes se vendent immédiatement au lancement ? Laissez-moi réfléchir, non. Aucune chance. En fait, les mineurs essaient probablement toujours d’acheter les cartes RTX 3080 Ti, 3080, 3070, 3060 Ti et 3060 limitées. Peut-être qu’ils espèrent que le limiteur sera à nouveau fissuré ou déverrouillé accidentellement. Peut-être qu’ils ont gagné trop d’argent grâce à la hausse des prix de la cryptographie au cours des six derniers mois. Ou peut-être sont-ils simplement optimistes quant à l’avenir de la cryptographie. La bonne nouvelle, selon votre point de vue, est que la rentabilité de l’exploitation minière a considérablement diminué au cours du mois dernier, ce qui signifie que des cartes comme la RTX 3090 rapportent désormais moins de 7 $ par jour après les coûts d’électricité, et la RTX 3080 est tombée à un peu plus de 5 $. par jour.
GeForce RTX 3070 Ti : pas idéal pour l’exploitation minière mais toujours rentable
Même si le RTX 3070 Ti n’avait pas un taux de hachage limité, il ne rapporterait qu’environ 4,25 $ un jour. Avec le limiteur en place, Ravencoin (KAWPOW) et Conflux (Octopus) sont actuellement les crypto-monnaies les plus rentables, et ces deux algorithmes de hachage semblent toujours fonctionner à pleine vitesse. La rentabilité devrait être un peu plus élevée avec le réglage, mais pour le moment, nous estimons ne gagner que 3,50 $ environ par jour. C’est encore suffisant pour que les cartes atteignent le seuil de rentabilité dans environ six mois, mais encore une fois, la rentabilité a chuté et pourrait continuer à baisser.
Les joueurs parmi nous l’espèrent certainement, mais même sans extraction de pièces cryptographiques , la demande de GPU continue de dépasser largement l’offre. En lançant le RTX 3070 Ti, avec ses puces GA104 regroupées et sa mémoire GDDR6X, Nvidia continue d’augmenter régulièrement le nombre de GPU qu’il vend. Nvidia produit également plus de GPU Turing en ce moment, principalement pour la gamme de cartes de mineurs CMP, et à un moment donné, l’offre devrait rattraper son retard. Cela arrivera-t-il avant l’arrivée des GPU de nouvelle génération ? Probablement, mais seulement parce que les GPU de nouvelle génération sont susceptibles d’être repoussés en raison des mêmes pénuries auxquelles sont confrontées les puces de génération actuelle.
D’accord, assez d’informations de base. Jetons un coup d’œil aux spécifications du RTX 3070 Ti, ainsi qu’aux GPU Nvidia associés tels que le 3080, le 3070 et le RTX 2070 Super de génération précédente :
La GeForce RTX 3070 Ti offre des performances de calcul un peu plus théoriques que la 3070, grâce à l’ajout de deux SM supplémentaires. Il a également des horloges légèrement plus élevées, ce qui lui donne 7 % de TFLOPS en plus — et il a toujours 27 % de TFLOPS en moins que le 3080. Le plus important est que le 3070 Ti passe de 14 Gbit/s de GDDR6 et 448 Go/s de bande passante à 19 Gbit/s GDDR6X et 608 Go/s de bande passante, une amélioration de 36 %. En général, nous nous attendons à ce que les performances se situent entre le 3080 et le 3070, mais plus près du 3070.
Outre les spécifications de performance, il est également important d’examiner la puissance. C’est un peu choquant de voir que le 3070 Ti a un TDP supérieur de 70 W à celui du 3070, et nous supposons que presque tout cela va dans la mémoire GDDR6X. Certains d’entre eux permettent également des horloges légèrement plus élevées, mais en général, il s’agit d’une augmentation significative du TDP juste pour un changement de VRAM.
Il reste la question de savoir si 8 Go de mémoire suffisent. De nos jours, nous dirions que c’est suffisant pour tout jeu auquel vous voulez jouer, mais il y a certainement des cas où vous rencontrerez des problèmes de capacité de mémoire. Sans surprise, beaucoup d’entre eux viennent dans les jeux promus par AMD, c’est presque comme si AMD avait convaincu les développeurs de cibler 12 Go ou 16 Go de VRAM avec des paramètres de qualité maximale. Mais quelques ajustements judicieux aux paramètres (comme baisser la qualité de la texture d’un cran) suffiront généralement.
La difficulté est qu’il n’y a pas de bon moyen d’obtenir plus de mémoire autre que de simplement le faire. L’interface 256 bits signifie que Nvidia peut faire 8 Go ou 16 Go-rien entre les deux. Et avec les 3080 et 3080 Ti offrant respectivement 10 Go et 12 Go, il n’y avait pratiquement aucune chance que Nvidia équipe un GPU moindre avec plus de mémoire GDDR6X. (Oui, je sais, mais le RTX 3060 12 Go reste un peu une anomalie dans ce département.)
Conception GeForce RTX 3070 Ti : Un mélange des 3070 et 3080
Unlike the RTX 3080 Ti, Nvidia actually made some changes to the RTX 3070 Ti’s design. Basically, the 3070 Ti has a flow-through cooling fan at the’back’of the card, similar to the 3080 and 3090 Founders Edition cards. In comparison, the 3070 just used two fans on the same side of the card. This also required some tweaks to the PCB layout, so the 3070 Ti doesn’t use the exact same boards as the 3070 and 3060 Ti. It’s not clear exactly how much the design tweak helps with cooling, but considering the 290W vs. 220W TDP, presumably Nvidia did plenty of testing before settling on the final product.
Overall, whether the change significantly improves the cooling or not, we think it does improve the look of the card. The RTX 3070 and 3060 Ti Founders Editions looked a bit bland, as they lacked even a large logo indicating the product name. The 3080 and above (FE models) include RGB lighting, though, which the 3070 Ti and below lack. Third party cards can, of course, do whatever they want with the GPU, and we assume many of them will provide beefier cooling and RGB lighting, along with factory overclocks.
One question we had going into this review was how well the card would cool the GDDR6X memory. The various Founders Edition cards with GDDR6X memory can all hit 110 degrees Celsius on the memory with various crypto mining algorithms, at which point the fans kick into high gear and the GPU throttles. Gaming tends to be less demanding, but we still saw 102C-104C on the 3080 Ti. The 3070 Ti doesn’t have that problem. Even with mining algorithms, the memory peaked at 100C, and temperatures in games were generally 8C–12C cooler. That’s the benefit of only having to cool 8GB of GDDR6X instead of 10GB, 12GB, or 24GB.
GeForce RTX 3070 Ti: Standard Gaming Performance
TOM’S HARDWARE GPU TEST PC
Our test setup remains unchanged from previous reviews, and like the 3080 Ti, we’ll be doing additional testing with ray tracing and DLSS — using the same tests as our AMD vs. Nvidia: Ray Tracing Showdown. We’re using the test equipment shown above, which consists of a Core i9-9900K, 32GB DDR4-3600 memory, 2TB M.2 SSD, and the various GPUs being tested — all of which are reference models here, except for the RTX 3060 (an EVGA model running reference clocks).
That gives us two sets of results. First is the traditional rendering performance, using thirteen games, at 1080p, 1440p, and 4K with ultra/maximum quality settings. Then we have ten more games with RT (and sometimes DLSS, where applicable). We’ll start with 4K, since this is a top-tier GPU more likely to be used at that resolution, plus it’s where the card does best relative to the other GPUs — CPU bottlenecks are almost completely eliminated at 4K, but more prevalent at 1080p. If you want to check 1080p/1440p/4K medium performance, we’ll have those results in our best graphics cards and GPU benchmarks articles — though only for nine of the games.
The RTX 3070 Ti does best as a 1440p gaming solution, which remains the sweet spot in terms of image quality and performance requirements. Overall performance ended up 9% faster than the RTX 3070 and 13% slower than the RTX 3080, so the added memory bandwidth only goes so far toward removing bottlenecks. However, a few games benefit more, like Assassin’s Creed Valhalla, Dirt 5, Horizon Zero Dawn, Shadow of the Tomb Raider, and Strange Brigade — all of which show double-digit percentage improvements relative to the 3070.
Some of the games are also clearly hitting other bottlenecks, like the GPU cores. Borderlands 3, The Division 2, Far Cry 5, FFXIV, Metro Exodus, and Red Dead Redemption 2 all show performance gains closer to the theoretical 7% difference in compute that we get from core counts and clock speeds. Meanwhile, Watch Dogs Legions ends up showing the smallest change in performance, improving just 3% compared to the RTX 3070.
The RTX 3070 Ti makes for a decent showing here, but we’re still looking at an MSRP increase of 20% for a slightly less than 10% increase in performance. Compared to AMD’s RX 6000 cards, the 3070 Ti easily beats the RX 6700 XT, but it comes in 6% behind the RX 6800 — which, of course, means it trails the RX 6800 XT as well.
On the one hand, AMD’s GPUs tend to sell at higher prices, even when you see them in places like the Newegg Shuffle. At the same time, RTX 30-series hardware on eBay remains extremely expensive, with the 3070 selling for around $1,300, compared to around $1,400 for the RX 6800. Considering the RTX 3070 Ti is faster than the RTX 3070, it remains to be seen where street pricing lands. Of course, the reduced hashrates for Ethereum mining on the 3070 Ti may also play a role.
Next up is 1080p testing. Lowering the resolution tends to make games more CPU limited, and that’s exactly what we see. The 3070 Ti was 7% faster than the 3070 this time and 11% slower than the 3080. It was also 7% faster than the 6700 XT and 6% slower than the 6800. While you can still easily play games at 1080p on the RTX 3070 Ti, the same is true of most of the other GPUs on our charts.
We won’t belabor the point, other than to note that our current test suite is slightly more tilted in favor of AMD GPUs (six AMD-promoted games compared to four Nvidia-promoted games, with three’agnostic’games). We’ll make up for that when we hit the ray tracing benchmarks in a moment.
Not surprisingly, while 4K ultra gaming gave the RTX 3070 Ti its biggest lead over the RTX 3070 (11%), it also got its biggest loss (17%) against the 3080. 4K also narrowed the gap between the 3070 Ti and the RX 6800, as AMD’s Infinity Cache starts to hit its limits at 4K.
Technically, the RTX 3070 Ti can still play all of the test games at 4K, just not always at more than 60 fps. Nearly half of the games we tested came in below that mark, with Valhalla and Watch Dogs Legion being the two lowest scores — and they’re still in the mid-40s. The RTX 3070 was already basically tied with the previous generation RTX 2080 Ti, which means the RTX 3070 Ti is now clearly faster than the previous-gen halo card, at half the price.
GeForce RTX 3070 Ti: Ray Tracing and DLSS Gaming Performance
So far, we’ve focused on gaming performance using traditional rasterization graphics. We’ve also excluded using Nvidia’s DLSS technology in order to provide an apples-to-apples comparison. Now we’ll focus on ray tracing performance, with DLSS 2.0 enabled where applicable. We’re only using DLSS in Quality mode (2x upscaling) in the six games where it is supported. We’ll have to wait for AMD’s FSR to see if it can provide a reasonable alternative to DLSS 2.0 in the coming months, though Nvidia clearly has a lengthy head start. Note that these are the same tests we used in our recent AMD vs. Nvidia Ray Tracing Battle.
Nvidia’s RTX 3070 Ti does far better — at least against the AMD competition — in ray tracing games. It’s not a complete sweep, as the RX 6800 still leads in Godfall, but the 3070 Ti ties or wins in every other game. In fact, the 3070 Ti basically ties the RX 6800 XT in our ray tracing test suite, and that’s before we enable DLSS 2.0.
Even 1080p DXR generally ends up being GPU limited, so the rankings don’t change much from above. DLSS doesn’t help quite as much at 1080p, but otherwise, the 3070 Ti ends up right around 25% faster than the RX 6800 — the same as at 1440p. We’ve mentioned before that Fortnite is probably the best’neutral’look at advanced ray tracing techniques, and the 3070 Ti is about 5–7% faster there. Turn on DLSS Quality and it’s basically double the framerate of the RX 6800.
GeForce RTX 3070 Ti: Power, Clocks, and Temperatures
We’ve got our Powenetics equipment working again, so we’ve added the 3080 Ti to these charts. Unfortunately, there was another slight snafu: We couldn’t get proper fan speeds this round. It’s always one thing or another, I guess. Anyway, we use Metro Exodus running at 1440p ultra (without RT or DLSS) and FurMark running at 1600×900 in stress test mode for our power testing. Each test runs for about 10 minutes, and we log the result to generate the charts. For the bar charts, we only average data where the GPU load is above 90% (to avoid skewing things in Metro when the benchmark restarts).
Nvidia gives the RTX 3070 Ti a 290W TDP, and it mostly makes use of that power. It averaged about 282W for our Metro testing, but that’s partly due to the lull in GPU activity between benchmark iterations. FurMark showed 291W of power use, right in line with expectations.
Core clocks were interesting, as the GeForce RTX 3070 Ti actually ended up with slightly lower clocks than the RTX 3070 in FurMark and Metro. On the other hand, both cards easily exceeded the official boost clocks by about 100 MHz. Custom third-party cards will likely hit higher clocks and performance, though also higher power consumption.
While we don’t have fan data (or noise data — sorry, I’m still trying to get unpacked from the move), the RTX 3070 Ti did end up hitting the highest temperatures of any of the GPUs in both Metro and FurMark. As we’ve noted before, however, none of the cards are running”too hot,”and we’re more concerned with memory temperatures. The 3070 Ti thankfully didn’t get above 100C on GDDR6X junction temperatures when testing, and even that value occured while testing crypto coin mining.
GeForce RTX 3070 Ti: Good but With Diminishing Returns
We have to wonder what things would have been like for the RTX 3070 Ti without the double whammy of the Covid pandemic and the cryptocurr ency boom. If you look at the RTX 20-series, Nvidia started at higher prices ($599 for the RTX 2070 FE) and then dropped things $100 with the’Super’updates a year later. Ampere has gone the opposite route: Initial prices were excellent, at least on paper, and every one of the cards sold out immediately. That’s still happening today, and the result is a price increase — along with improved performance — for the 3070 Ti and 3080 Ti.
Thankfully, the jump in pricing on the 3070 Ti relative to the 3070 isn’t too arduous. $100 more for the switch to GDDR6X is almost palatable. Except, while the 3070 offers about 90% of the 3070 Ti performance for 80% of the price and represents an arguably better buy, the real problem is the RTX 3080. It’s about 12–20% faster across our 13 game test suite and only costs $100 more (a 17% price increase).
Well, in theory anyway. Nobody is really selling RTX 3080 for $700, and they haven’t done so since it launched. The 3080 often costs over $1,000 even in the lottery-style Newegg Shuffle, and the typical price is still above $2,000 on eBay. It’s one of the worst cards to buy on eBay, based on how big the markup is. In comparison, the RTX 3070 Ti might only end up costing twice its MSRP on eBay, but that’s still $1,200. And it could very well end up costing more than that.
We’ll have to see what happens in the coming months. Hopefully, the arrival of two more desktop graphics cards in the form of the RTX 3080 Ti and RTX 3070 Ti will alleviate the shortages a bit. The hashrate limiter can’t hurt either, at least if you’re only interested in gaming performance, and the drop in mining profitability might help. But we’re far from being out of the shortage woods.
If you can actually find the RTX 3070 Ti for close to its $600 MSRP, and you’re in the market for a new graphics card, it’s a good option. Finding it will be the difficult part. This is bound to be a repeat of every AMD and Nvidia GPU launch of the past year. If you haven’t managed to procure a new card yet, you can try again (and again, and again…). But for those who already have a reasonable graphics card, there’s nothing really new to see here: slightly better performance and higher power consumption at a higher price. Let’s hope supply and prices improve by the time fall blows in.
MORE: Best Graphics Cards
MORE: GPU Benchmarks and Hierarchy
MORE: All Graphics Content