Announced back at CES 2025, Nvidia's RTX 50-series is here, built on the latest Blackwell architecture. First up for review is the new flagship, the RTX 5090, which hits the market at an eye-watering £1939/$1999. Packing in a total of 92.2 billion transistors, 21760 CUDA Cores and 32GB of GDDR7 memory, Nvidia claims that the RTX 5090 is ‘up to 2x' faster than the out-going RTX 4090, but is that really the case? And what's power consumption like given the rated 575W TGP? All that and more is covered in today's review…

Timestamps

00:00 Intro
00:49 RTX 5090 recap, pricing + specs
02:50 Test setup
04:31 Alan Wake 2
05:11 Black Myth: Wukong
05:47 Cyberpunk 2077
06:17 Final Fantasy XVI
06:42 Forza Horizon 5
07:11 Ghost of Tsushima
07:38 Horizon Forbidden West
08:05 The Last of Us Part 1
08:30 Plague Tale: Requiem
08:53 Senua’s Saga: Hellblade 2
09:22 Starfield
09:50 Total War: Warhammer III
10:37 12-game average results
11:41 Cost per frame analysis
12:47 RT Alan Wake II
13:19 RT Black Myth: Wukong
13:47 RT Cyberpunk 2077
14:14 RT F1 24
14:45 RT Ratchet & Clank
15:07 RT Returnal
15:34 RT Shadow of the Tomb Raider
16:07 RT Star Wars Outlaws
16:33 RT 8-game average
17:31 DLSS 4 – new Transformer model
19:59 DLSS 4 – Multi Frame Generation
27:09 AI & productivity benchmarks
28:33 Closer look at the Founders Edition
29:44 Thermals & acoustics (with soundest)
31:31 Power draw and efficiency
33:33 Closing thoughts

If you missed the announcement earlier in the month, the Nvidia RTX 5090 is arriving as the first of four 50-series GPUs. The RTX 5080 is the next cab off the rank, launching on January 30th, while the RTX 5070 and RTX 5070 Ti are set to arrive some time next month. This review is firmly focused on the flagship SKU, though, and there is plenty to cover. We look at raster performance, ray tracing, new upgrades made to DLSS 4, alongside AI benchmarks, power, efficiency and more.

First, a quick spec recap. The RTX 5090 is built on the new GB202 die, measuring 750mm², though it's not quite a full implementation of the silicon. Instead we find a total of 11 Graphics Processing Clusters (GPCs), each holding up to eight Texture Processing Clusters (TPC), for a total of 85. Each TPC is home to two Streaming Multiprocessors (SMs), giving us 170, and each SM still holds 256 CUDA Cores, meaning the RTX 5090 has an eye-watering total of 21760 shaders. We also find 170 RT cores, 680 Tensor cores, 680 Texture Units, and 176 ROPs.

This time around, however, there's no node-shrink, and GB202 remains fabricated on TSMC's N4 node, as per the RTX 40-series. As such, rated clock speed is not increased this generation and is actually touted slightly below that of the RTX 4090, with the RTX 5090 delivering a rated 2407MHz boost clock, compared to its predecessor's 2520MHz boost.

The memory configuration has seen significant upgrades, though.The RTX 5090 now comes equipped with a super-wide 512-bit memory interface, paired with 32GB GDDR7 memory running at 28Gbps, and that puts total memory bandwidth at a staggering 1792 GB/s. L2 cache is also increased to 98MB, up from the 74MB of the RTX 4090.

Considering the large increases to die size and core count, but with no node shrink, it's perhaps unsurprising to see power draw has increased, this time boasting a 575W TGP. This is something we focus on closely in this review, using our enhanced GPU power testing methodology, so read on for our most detailed power and efficiency testing yet.