Introducing the Nvidia GeForce RTX 5080
The Nvidia GeForce RTX 5080 Founders Edition takes the honor guard position for the RTX 5090. In times past, the penultimate Nvidia GPU of each generation has often been the best overall pick. But the gap between first and second place has widened significantly in the past two generations, at least for 4K gaming and other demanding workloads. The 5080 also takes over from the RTX 4080 and RTX 4080 Super, often with only modest gains. It may still be one of the best graphics cards when the dust clears, but it doesn’t have the wow factor of its big brother.
Both the RTX 5080 and RTX 5090 will go on sale tomorrow, January 30, 2025. While we anticipate a lot of demand for the halo card, the 5080 will hopefully be more readily available — but probably only after the initial wave of eager buyers clears. And there’s still the risk that businesses looking for affordable AI hardware might drive inventory shortages because while the 5080 can’t match a 5090 in raw performance, two of them would certainly provide plenty of computing for nominally the same price.
RTX 5080 will have the same core feature set, meaning stuff like native FP4 support that could entice AI researchers and developers. But it still ‘only’ has 16GB of VRAM, and many AI models tend to be voracious when it comes to memory requirements — though DeepSeek has certainly shaken many of the foundational thoughts about AI training and inference, as well as Nvidia’s stock price.
We were extremely crunched for time on the RTX 5090 review, and things have only been slightly better on the RTX 5080. There’s still a lot to dissect, and unfortunately, we can’t shake the feeling that the initial Blackwell drivers are holding the cards back. The 1080p results are particularly bad at times, and Nvidia’s heavy reliance on Multi Frame Generation (MFG) for the initial performance preview suggests that was probably at the forefront of the driver team’s work, rather than general performance.
You can check the boxout with additional links and information on the Nvidia Blackwell and RTX 50-series GPUs. The succinct story for the RTX 5080 is that, outside of certain AI workloads and MFG, it’s currently a pretty minor upgrade over the prior generation 4080 cards. (The 4080 Super was only a few percent faster, with its primary attraction being a $200 price cut compared to the vanilla model.) The specs basically say most of what you need to know.
| Graphics Card | RTX 5080 | RTX 4080 Super | RTX 4080 | RTX 3080 Ti | RTX 3080 12GB | RTX 3080 | RTX 2080 Super | RTX 2080 |
|---|---|---|---|---|---|---|---|---|
| Architecture | GB203 | AD103 | AD103 | GA102 | GA102 | GA102 | TU104 | TU104 |
| Process Technology | TSMC 4N | TSMC 4N | TSMC 4N | Samsung 8N | Samsung 8N | Samsung 8N | TSMC 12FFN | TSMC 12FFN |
| Transistors (Billion) | 45.6 | 45.9 | 45.9 | 28.3 | 28.3 | 28.3 | 13.6 | 13.6 |
| Die size (mm^2) | 378 | 378.6 | 378.6 | 628.4 | 628.4 | 628.4 | 545 | 545 |
| SMs / CUs / Xe-Cores | 84 | 80 | 76 | 80 | 70 | 68 | 48 | 46 |
| GPU Shaders (ALUs) | 10752 | 10240 | 9728 | 10240 | 8960 | 8704 | 3072 | 2944 |
| Tensor / AI Cores | 336 | 320 | 304 | 320 | 280 | 272 | 384 | 368 |
| Ray Tracing Cores | 84 | 80 | 76 | 80 | 70 | 68 | 48 | 46 |
| Boost Clock (MHz) | 2617 | 2550 | 2505 | 1665 | 1845 | 1710 | 1815 | 1800 |
| VRAM Speed (Gbps) | 30 | 23 | 22.4 | 19 | 19 | 19 | 15.5 | 14 |
| VRAM (GB) | 16 | 16 | 16 | 12 | 12 | 10 | 8 | 8 |
| VRAM Bus Width | 256 | 256 | 256 | 384 | 384 | 320 | 256 | 256 |
| L2 / Infinity Cache | 64 | 64 | 64 | 6 | 6 | 5 | 4 | 4 |
| Render Output Units | 112 | 112 | 112 | 112 | 96 | 96 | 64 | 64 |
| Texture Mapping Units | 336 | 320 | 304 | 320 | 280 | 272 | 192 | 184 |
| TFLOPS FP32 (Boost) | 56.3 | 52.2 | 48.7 | 34.1 | 33.1 | 29.8 | 11.2 | 10.6 |
| TFLOPS FP16 (FP4/FP8 TFLOPS) | 450 (1801) | 418 (836) | 390 (780) | 273 | 264 | 238 | 89 | 85 |
| Bandwidth (GB/s) | 960 | 736 | 717 | 912 | 912 | 760 | 496 | 448 |
| TBP (watts) | 360 | 320 | 320 | 350 | 350 | 320 | 250 | 215 |
| Launch Date | Jan 2025 | Jan 2024 | Nov 2022 | Jun 2021 | Jan 2022 | Sep 2020 | Jul 2019 | Sep 2018 |
| Launch Price | $999 | $999 | $1,199 | $1,199 | N/A | $699 | $699 | $699-$799 |
The biggest change, outside of AI and MFG, is support for faster GDDR7 memory. The RTX 5080 has 960 GB/s of bandwidth, compared to 736 GB/s on the 4080 Super and 717 GB/s on the original 4080. So, depending on your point of reference, that’s 30–34 percent more bandwidth, a pretty sizeable upgrade.
But in core processing power, ignoring the new native FP4 number format support, the upgrades are far less impressive. RTX 5080 has 84 Streaming Multiprocessors (SMs) and 10752 CUDA cores, compared to the 4080 Super’s 80 SMs and the 4080’s 76 SMs. Clock speeds are slightly higher in theory, but in practice, it’s mostly a wash. Raw compute ends up being 8% more than the 4080 Super and 16% more than the 4080.
Most of the other specs scale with the number of SMs, so there’s a similar potential 8% and 16% uplift in tensor compute for the existing FP8, FP16, and other formats. However, Blackwell adds native FP4 support (Ada relied on FP4 running as an FP8 calculation), which doubles the potential throughput if you don’t need the higher precision of FP8. That’s where the 1.8 petaFLOPS of compute comes from, compared to just 836 teraFLOPS on the 4080 Super.
ROPS is the same 112 count on the 5080 and 4080-class GPUs, so pixel shading throughput hasn’t changed. Ray tracing, on the other hand, sees another doubling in ray/triangle intersection calculations, and Nvidia says the 5080 offers 170.6 teraFLOPS of RT compute, compared to 121 and 113 teraFLOPS of RT on the 4080 Super and 4080, respectively.
There’s also a new PCIe 5.0 interface, though that shouldn’t matter much for most tasks. The biggest benefit will be for multi-GPU configurations running AI and GPGPU tasks — not for gaming, which no longer has NVLink or multi-GPU support. Power consumption also sees a modest bump from 320W with the previous generation to 360W with the 5080.
The good news is that the RTX 5080 won’t cost more than the outgoing RTX 4080 Super. Or that’s the theory. It’s really going to depend on supply and demand, and as we’ve seen with the dwindling inventories of RTX 4080 and 4090 parts over the past few months, there’s still enough demand to push prices up if Nvidia doesn’t provide an adequate supply. And, much to no one’s surprise, Nvidia says the 5090 and 5080 may experience stock shortages in the coming days.
Why isn’t that a surprise? Because there’s a limited number of TSMC wafers to go around right now. Every GB202 or GB203 wafer that Nvidia orders mean one less GB200 wafer and Nvidia previously said its Blackwell B200 supply is already allocated for 2025. That means there’s limited incentive to produce a bunch of consumer GPUs that sell for an order of magnitude less money than the most powerful data center parts.
That means we’ll likely see third-party AIB (add-in board) partner cards selling for far more than the base $999 MSRP of the RTX 5080. There are already hints that some card models could cost $1,399 or more, and if there’s a supply deficit, then we aren’t likely to see many base-price cards after the initial stock lands. Hopefully, the shortages won’t be as severe as we saw with the 3080 cards in 2020–2021 (those were driven by cryptomining), but only time will tell.
For now, let’s take a closer look at the RTX 5080 Founders Edition, and then we’ll hit the benchmarks.
