NBCUni 9.5.23

Review: Nvidia GeForce RTX 4070 Ti GPU

By Brady Betzel

A few weeks ago, Nvidia announced a behemoth of a GPU in the RTX 4090 Founders Edition. It retails for $1,599, which might be beyond some people’s budgets. But do not fear. This week Nvidia introduced the RTX 4070 Ti, with a retail price of $799. The Nvidia GeForce RTX 4070 Ti packs a lot of power similar to the 4090 except for the CUDA cores and memory.

For comparison:
– RTX 4090 CUDA cores: 16,384; memory: 24GB ($1,599)
– RTX 4070 Ti CUDA cores: 7,680; memory: 12GB ($799)

As you can see, the 4070 Ti is more affordable, but that price comes with a CUDA speed and memory deficit. However, if you’re a modest multimedia creator, the RTX 4070 Ti might be all you need. The technical specs of the Nvidia GeForce RTX 4070 Ti are as follows:

Nvidia CUDA cores: 7680; boost clock (GHz): 2.61; base clock (GHz): 2.31; memory size: 12GB; memory type: GDDR6X; memory interface width: 192-bit; power connectors: 2x PCIe 8-pin cables (adapter in box) or 300W or greater PCIe Gen 5 cable.

The 4070 Ti embraces Nvidia’s latest tech, including the Nvidia Ada Lovelace architecture, dedicated raytracing cores, Nvidia DLSS 3, reflex low latency, Nvidia Encoder, Nvidia Broadcast, Nvidia Studio, Game Ready and Studio Drivers.

In this review, I am using the Studio Drivers and won’t be covering gaming. That being said, if you play any modern games, the Nvidia RTX 40-series GPUs should be at the top of your list. I was sent the ASUS TUF GeForce RTX 4070 Ti OC Edition, which has a few unique features:

  • OC Mode: 2760 MHz (OC mode) / 2730 MHz (default mode)
  • Axial-tech fans (21% larger cooling fans)
  • 25-slot design
  • Diecast shroud, frame, and backplate
  • Three DisplayPort /two HDMI

Testing
Multimedia creators, like me, want a seamless playback experience and ultra-fast export speed. The first task I performed with the RTX 4070 Ti was inside of Blackmagic’s Resolve 18.1.2. Nvidia sends a few sample files to test, which I layout below with some frames-per-second details:

  • Wedding Styles (:50): Vimeo 2160p – :37
  • Bride Face Refinement (:24): Vimeo 2160p – :33 / ~18fps
  • Optical Flow – Enhanced Better 50% Retime (:08): Vimeo 2160p – :36 / ~6fps
  • SuperScale2x 4K source (:17): Vimeo 2160p – :18 / 26.5 fps
  • SuperScale4x HD source (1:16): Vimeo 2160p – 1:35 / 19.5 fps
  • SpeedWarp example (:30) – .mov/H.264 – :19 / ~112 – 125fps
  • Scene Edit Detect (:40) ~ :05

Overall, the AI performance inside of the Resolve was excellent. The newly updated Magic Mask is an amazing feature that is very fast (but not instant) and mostly accurate at selecting foreground images. Even snappier is the Scene Edit Detect. For sample exports, Nvidia sent a few test files:

H.264 Exports:
ProRes422HQ 4K/30 – :16 /271fps

ProRes422HQ 8K/30 – :55 /77fps

AV1 Exports:
ProRes422HQ 4K/30 – :14 / 284fps

ProRes422HQ 8K/30 – :55 / 77.5fps

You can compare these numbers with the RTX 4090 Founders Edition results from a few weeks ago:

RTX 4090 – H.264 Exports:
ProRes422HQ 4k/30 – :14

ProRes422HQ 8k/30 – :47

RTX 4090 – AV1 Exports:
ProRes422HQ 4k/30 – :14

ProRes422HQ 8k/30 – :50

There really isn’t a gigantic difference, but it’s something to keep in mind when deciding on a GPU.

Benchmarks
I tried to keep the benchmarks as GPU-focused as possible, but to shed a little light on the system I built for this: It houses an AMD 5950x CPU (16 cores/32 threads), 64GB DDR4 3200 memory, 1.9TB NVME SSD and an Asus ProArt motherboard. The system is nice, but it’s not top-of-the line or bottom-of-the-line. It’s slightly above an average multimedia workstation.

  • Blender Benchmark:

CPU – Monster: 190.482643 – Junkshop; 126.605179 – Classroom: 91.325746

GPU – Monster: 3769.565198 – Junkshop; 1689.343314 – Classroom: 1869.861870

  • Octane Bench:

RTX Off: 565.72

RTX On: 739.08

  • VRay:

CPU: 20161 vsamples

GPU CUDA: 2326 vpaths

GPU RTX: 3062 vrays

  • Blackmagic RAW Speed Test:

8K CPU: 70fps

8K CUDA: 102fps

Testing
First up was testing inside of Blackmagic Resolve 18.1.2, where I took clips from different cameras and did a basic color correction in a 3840×2160 timeline. I use these same sequences and effects in a lot of reviews. The clips include:

  • ARRI RAW: 3840×2160 24fps – 7 seconds, 12 frames
  • ARRI RAW: 4448×1856 24fps – 7 seconds, 12 frames
  • BMD RAW: 6144×3456 24fps – 15 seconds
  • Red RAW: 6144×3072 23.976fps – 7 seconds, 12 frames
  • Red RAW: 6144×3160 23.976fps – 7 seconds, 12 frames
  • Sony a7siii: 3840×2160 23.976fps – 15 seconds

I then added Blackmagic noise reduction, sharpening and grain. Finally, I replaced the noise reduction with Neat Video’s noise reduction. From there I exported multiple versions: DNxHR 444 10-bit OP1a .MXF file, DNxHR 444 10-bit .mov, H.264 .mp4, H.265 .mp4, and then an IMF package using the default settings.

Resolve 18 Exports DNxHR 444 10-bit .mxf DNxHR 444 10-bit .mov H.264 .mp4 H.265 .mp4 IMF
Color Correction Only 00:40 00:36 00:22 00:22 01:21
CC + Resolve Noise Reduction 01:37 01:37 01:33 01:33 01:38
CC, Resolve NR, Sharpening, Grain 02:16 02:16 02:12 02:24 02:16
CC + Neat Video Noise Reduction 03:20 03:21 03:11 03:12 03:18

When compared to a previous review of the Lenovo P620 Gen 2 workstation with an Nvidia A6000 GPU, the RTX 4070 Ti performs faster, overall. There are some hit-or-miss sections possibly due to new Resolve improvements or Nvidia Studio Driver improvements, but you can really get a sense with the Neat Video noise reduction testing, which for the RTX 4070 Ti is a pretty good improvement.

I also ran similar tests inside of Adobe Premiere Pro 2023, exporting using Adobe Media Encoder. I find it hard to run comparisons in Premiere Pro, as they always come out a little quirky, but for those interested, I ran the same tests. The footage assets were the same as the ones I used in the Resolve test, but I used Adobe’s noise reduction, sharpening and grain.

Adobe Premiere Pro 2023 (Individual Exports in Media Encoder) DNxHR 444 10-bit .mxf H.264 .mp4 H.265 .mp4
Color Correction Only 02:02 01:07 01:08
CC +  NR, Sharpening, Grain 17:48 46:30 46:35
Adobe Premiere Pro 2023 (Simultaneous Exports in Media Encoder)
Color Correction Only 02:37 02:50 02:50
CC +  NR, Sharpening, Grain 17:33 17:33 17:33

You can export Premiere Pro sequences in Media Encoder individually or simultaneously where they export concurrently. I’m not sure why the export speeds are so different, but I ran them multiple times and had similar results.

For my final test, I wanted to use Blackmagic’s new Resolve Proxy Generator, which is a dead-simple way of creating proxy-sized media. I chose H.265 10-bit/1080p for the proxy file type, pointed it to some Red RAW and Sony files and let it rip. I noticed the GPU monitor in windows hitting around 90% usage, which was nice to see. Here are the results:

Sony a7iii UHD (3840×2160) .mp4 Master Files

  • 46 files
  • 166fps
  • Total master file size: 15GB

Proxies: H.265/10-bit 1080p

  • Total proxy file size: 1.04GB

Red RAW – various resolutions

  • Two files
  • 28fps
  • Total master file size: 3.81GB

Proxies: H.265/10-bit 1080p

  • Total proxy file size: 18MB

The Sony MP4 files flew through Blackmagic’s Proxy Generator, averaging around 166fps. The Red RAW files, which were the same I used in the Resolve test from earlier, were a bit slower. They averaged 28fps. But for chewing through Red RAW files, this was a nice speed.

Summing Up
If you are in the market for a new GPU, you don’t want to hit that $1,599 price tag of the RTX 4090, but you want some of the same performance improvements, the Nvidia RTX 4070 Ti is the sweet spot, retailing for $799. It’s still not “cheap” by any means, but the speed improvements you will achieve will likely allow you to make more money and pay for the GPU in no time.

You can find out more about the Asus TUF RTX 4070 Ti OC version that I tested here.


Leave a Reply

Your email address will not be published. Required fields are marked *

I accept the Privacy Policy

This site uses Akismet to reduce spam. Learn how your comment data is processed.