Nvidia's RTX 50 series will debut with the RTX 5090 and RTX 5080 starting January 30. While official reviews, including ours, are under embargo, a potential leak has surfaced that suggests a modest 30% improvement over the last generation. The RTX 5090 has reportedly been benchmarked in Geekbench 5 (via Benchleaks) using the CUDA API, likely by a reviewer who inadvertently made the test results public. As always, spread some salt over this leak, even though the performance claims are similar to what Nvidia has depicted in its slides.
The test bench features AMD's Zen 4-based Ryzen 9 7900X with 12 cores / 24 threads and an Asus ProArt X870E Creator WiFi motherboard. The system has 32GB of DDR5-6000 memory and uses Windows 11 Pro as the Operating System for this test. The benchmark was carried out using the CUDA API, the results of which are hard to come by as Geekbench does not publicly maintain a database for CUDA benchmarks.
We've gathered a few RTX 40 numbers to give you a general performance overview; however, your results might differ slightly. Remember that Geekbench is just a synthetic benchmark and may not accurately reflect how this GPU performs in real-world scenarios.
The RTX 5090 amasses 542,157 points in the CUDA API, landing a solid 27% lead over its predecessor, the RTX 4090. This isn't a massive leap generation-on-generation as we're used to seeing, plus the RTX 5090 has 32% more CUDA cores than the RTX 4090. Then you have the apparent elephant in the room: the pricing. Nvidia's $1,999 price tag for the RTX 5090 makes it roughly 25% more expensive than the RTX 4090, assuming you got one at MSRP. Again, this is leaked information, so almost everything you see here is subject to change.
| GPU | CUDA Cores | VRAM | Memory Type | Memory Speed | Bus Width | Bandwidth | CUDA Score | % vs RTX 5090 |
|---|---|---|---|---|---|---|---|---|
| RTX 5090 | 21,760 | 32GB | GDDR7 | 28 Gbps | 512-bit | 1.79 TB/s | 542,157 | 100.00% |
| RTX 4090 | 16,384 | 24GB | GDDR6X | 21 Gbps | 384-bit | 1.01 TB/s | 424,332 | 78.27% |
| RTX 4080 | 9,728 | 16GB | GDDR6X | 22.4 Gbps | 256-bit | 0.71 TB/s | 300,728 | 55.47% |
The GB202, which powers the RTX 5090, is 744mm2 and has roughly 92 billion transistors. This equates to around 123 million transistors per mm2 on the updated 4NP process, similar to the 4N used on Ada Lovelace. The L1 and L2 cache sizes per SM also show slight improvement, compensated by the faster GDDR7 memory. Nvidia claims a 2x increase in performance over the RTX 40 series, suggesting that an RTX 5070 equals an RTX 4090, but this requires enabling Multi Frame Generation.
Conversely, the RX 9070 XT from AMD has been rumored to match the RTX 4080 Super in raster performance. All that power has been crammed into a near-390mm2 Navi 48 chip, expected to be priced around $500. AMD has been tight-lipped about RDNA 4, though we might hear more from them in February when the mid-ranged RTX 5070 hits shelves and AMD finalizes its pricing strategy.
