NVIDIA RTX 5090 at 575 Watts, RTX 5080, 5070 Ti, & 5070 Specs

RTX Blackwell

NVIDIA spent a lot of time talking about AI, and we’ll get into some of that briefly at the end.

We’ll start with the hard information that we have.

NVIDIA announced its Blackwell GPU as it’ll arrive to the consumer market in January. The GPU, presumably the full Blackwell die, was noted as a 92 billion transistor solution (as compared to the RTX 4090’s approximate 76 billion).

Notably, the card CEO Jensen Huang showed off at the event appeared to be a 2-slot card, which he noted has 2 fans. This is a stark contrast to the huge prototype NVIDIA cooler we just tore down. This is a big swing from the 3-slot and 4-slot cards we’ve become used to, which were necessary for thermal and power management.

Hard Specs

NVIDIA also posted some of the hard specs to its website.

We can start with the power, which is listed as 575W total graphics power for the 5090. 

That’s a huge amount of load to put on a single 16-pin connector and we’re concerned about the strain on it. We also wonder if this will be coupled with a redesign of coolers to try and actively cool the area of the power connector.

The RTX 5090 Blackwell GPU is listed as having 21,760 CUDA cores, memory at 32 GB of GDDR7, clocks at 2.41 GHz boost and 2.01 GHz base, and a large 512-bit memory interface width. NVIDIA has iterated its Tensor core generation to 5 and its RT core generation to 4 for Blackwell, though we don’t yet have architectural details on what that actually means on the consumer side. We expect those details soon.

The FE card is listed as 304mm by 137mm for dimensions, and 2 slots thick.

Jumping over to the prior RTX 4090 (watch our review) for reference: The 4090 has 16,384 CUDA cores, down notably from the 21,760 of the 5090 -- but cores can’t be linearly compared, especially cross generation, so the real-world impact likely won’t track linearly. The 4090 also ran 24 GB of GDDR6X rather than the 32GB GDDR7 on the 5090. The 4090’s clocks are higher as advertised, though, at 2.52 GHz boost and 2.23 GHz base -- but clocks, like core count, aren’t everything.

Zooming out to look at memory capacity, we see the 5090 at 32 GB, the 5080 at 16 GB, the 5070 Ti at 16 GB, and the 5070 at 12 GB. For perspective, the RTX 4070 (watch our review) is also 12 GB, the 4070 Ti is 12GB, the 4070 Ti Super (read our review) is 16GB, and the 4080 is 16GB.

Going to a fuller look at the specs, the 5080 is listed at 10,752 CUDA Cores, which is slightly more than the RTX 4080’s 9,728 CUDA cores and not the same huge change we see with the 4090 to 5090. The 5070 Ti lists 8960, against 7680 on the 4070 Ti (watch our review), and the 5070 lists 6144, up from 5888 on the 4070 configuration. 

Again, these aren’t directly comparable as they’re different generations, but are useful for establishing how NVIDIA is positioning the cards.

The full specs page shows a 2.62 GHz max boost on the 5080, 2.45 GHz on the 5070 Ti, and 2.51 GHz on the 5070. We already said the 5090 has a larger memory bus. The 5080 and 5070 Ti both run a 256-bit bus, with the 5070 at 192 bits.

Other than the 575W of the 5090, NVIDIA lists the other cards at 360W, 300W, and 250W for total board power. Broadly speaking, NVIDIA’s power consumption appears to be increasing. 

This is good timing with our efficiency testing: It’s possible efficiency is up despite power draw also going up, but that’s what we’ll find out.

First-Party Marketing Claims

We’ll have our own benchmarks soon enough and you should rely on those or the independent benchmarks of other trusted reviewers. We can still reference NVIDIA’s first-party claims to get an idea for where they stand.

NVIDIA’s webpage has a relative performance chart that’s pretty hard to actually read, but we can get an idea. NVIDIA claims the 5090 outperforms the 4090 by over 2x in some situations, such as with Cyberpunk and Wukong, among others. In these tests, they list “DLSS+Full RT” as the settings. The footnote that’s nearly the same color as the page background says that the 40-series used frame generation in this testing, but the 50 series used MFG 4x mode. This makes the comparison not like-for-like.

We consider this approach flawed, but we’ll look at their other claims for full perspective. Switching to the RTX 5080, NVIDIA shows it as outperforming the RTX 4080 (watch our review) by, again, sometimes 2x -- this is tough to filter through differing settings tested, unfortunately.

The 5070 Ti shows beyond 2x gains against the 4070 Ti in the same way and with the same settings difference, with the 5070 and 4070 showing the same in NVIDIA’s first-party claims.

Of all of these, the Plague Tale comparison is maybe the most fair since NVIDIA notes that it only has DLSS 3.

As for what DLSS 4 actually is, NVIDIA ran this new blog post to introduce it. Of MFG, or Multi-Frame Generation, the post reads, “DLSS Multi Frame Generation generates up to three additional frames per traditionally rendered frame, working in unison with the complete suite of DLSS technologies to multiply frame rates by up to 8X over traditional brute-force rendering.”

The page continues to say, “Our new frame generation AI model is 40% faster, uses 30% less VRAM, and only needs to run once per rendered frame to generate multiple frames. For example, in Warhammer 40,000: Darktide, this model provided a 10% faster frame rate, while using 400MB less memory at 4K, max settings, using DLSS Frame Generation.”

This section of the article also indicates that you’ll be able to override the DLSS model used in games that don’t get updates from devs, which is already possible to be done manually by some users who replace .dll files. This new approach looks like it will be more user-friendly.

NVIDIA stated that its Blackwell GPUs will have 2x the memory bandwidth of Ada, at 1.8TB/s for the cited spec, it claimed 2x the the RT TFLOPS, and 1.5x Ada shader performance. 

The PCB showcased was a relatively small square -- like a further cut-down version of the 4090 FE PCB if you were to chop the wings off -- that appears to be sandwiched between two full flow-through fans. 

The design is shown in this explosion diagram where NVIDIA illustrates the PCB centrally with densely populated components on both sides of the board. The cooler also utilizes a vapor chamber cooling solution with what appears to be 5 heatpipes, assuming the render is accurate. The GPU directly contacts the vapor chamber as you would expect, with the flow through area handling the flanking heatsinks.

Uniquely, that appears to take some learnings NVIDIA had from this cooler, which is a full flow-through design. 

AI on GeForce

NVIDIA also spent limited time talking about other AI features for gaming. Although it is publishing materials to its site, we’ll wait for the full architectural briefing to get into more depth. For now, the company highlighted these:

• RTX Neural Material, wherein it showed an example material cut from 47 MB to 16 MB with this solution.
• DLSS improvements by moving to a Transformer model rather than the previous CNN model.

NVIDIA’s keynote was lengthy and covered topics outside of our typical coverage scope. For now, we’re just focusing on getting these basics out to you and will revisit the other announcements in more depth as we have time to adequately read through all the released materials.

As for release dates: NVIDIA cited January in its keynote, but its website specifically lists January 30th for the RTX 5090 and 5080. It says the 5070 Ti and 5070 will arrive in February.