NVIDIA Blackwell architecture: what really changes with the RTX 50

FP4 Tensor cores, GDDR7, Multi Frame Generation and promises still pending, 2026 review

By the FlowUP team, gaming PC builder in Genappe (Belgium) · Updated 29 September 2026

In brief: Blackwell is NVIDIA’s graphics card architecture that powers the GeForce RTX 50, from the RTX 5050 to the RTX 5090, released between January and July 2025. Its main new features: 5th-generation Tensor cores capable of computing in FP4 (enabling the DLSS 4 Multi Frame Generation), 4th-generation RT cores designed for highly detailed geometry (Mega Geometry), GDDR7 memory, PCIe 5.0 and DisplayPort 2.1b. The same name also refers to the AI chips for data centres (B200, B300), which have nothing to do with your gaming PC. Some January 2025 promises, such as Reflex 2 and neural texture compression, were still not available in games by September 2026.

Blackwell: what exactly are we talking about?

At NVIDIA, each graphics card generation is named after a scientist: Ampere for the RTX 30, Ada Lovelace for the RTX 40, and Blackwell, in tribute to mathematician David Blackwell, for the RTX 50.

Small source of confusion: NVIDIA uses the same name for two very different families of chips. The GeForce RTX 50 for gamers and creators, and the B200, GB200 and B300 (Blackwell Ultra) for AI servers. They share ideas, but not the same chip. This article is about the GeForce version.

Graphics chip surrounded by its video memory modules on a graphics card

In the centre, the graphics chip; around it, the GDDR7 video memory modules.

What changes in the chip

ItemNew featureWhy it matters
Manufacturing processTSMC 4N (NVIDIA custom process)The same manufacturing process family as the RTX 40: the gains come from the architecture, not a finer process
5th-generation Tensor coresFP4 computationDoubles AI throughput and halves the required memory, according to NVIDIA: this is what enables Multi Frame Generation
4th-generation RT coresTriangle cluster intersection engineSpeeds up ray tracing in very detailed scenes (Mega Geometry)
Compute coresUnified INT32 and FP32 unitsMore flexibility for shaders, including neural shaders
MemoryGDDR7 (except RTX 5050 in GDDR6)More bandwidth than GDDR6X
ConnectionsPCIe 5.0, DisplayPort 2.1bUp to 4K at 480 Hz or 8K at 165 Hz
9th-generation NVENC video encoderH.264 and HEVC at 4:2:2Useful for video editing with files from recent cameras
AMP processor and Flip MeteringAI workload management and display timingEssential so that generated images are displayed at regular intervals

Source: NVIDIA, white paper on the RTX Blackwell architecture. The largest chip, the GB202 of the RTX 5090, has 92.2 billion transistors on 750 mm².

DLSS 4 and Multi Frame Generation: the real change for gamers

The most visible new feature of Blackwell is Multi Frame Generation. Where the RTX 40 generated an image between two computed frames, the RTX 50 can generate up to three images using DLSS 4, and then up to five with DLSS 4.5 and its 6X mode (March 2026). At CES 2025, NVIDIA announced performance multiplied by over 8 in a Cyberpunk 2077 scene on RTX 5090, combining upscaling and image generation.

That figure should be read with caution. Generated images make the display look smoother, but don’t make the game more responsive: latency still depends on the frames that are actually computed. Our view: Multi Frame Generation is excellent for enjoying a 240 Hz screen in single-player when you already have a baseline of at least 60 FPS; much less useful in competitive play. We break down the versions and settings in DLSS 4.5 versus FSR Redstone and FSR 4.1.

Les Numériques tested the RTX 5080 at launch, with a good preview of what DLSS 4 changes:

Les Numériques, “Test of the Nvidia RTX 5080: saved by DLSS 4!”, January 2025.

Kept promises, pending promises

In January 2025, NVIDIA unveiled several technologies related to Blackwell. Twenty months later, the verdict is mixed:

TechnologyStatus as of 29 September 2026Detail
Multi Frame Generation (DLSS 4)AvailableSince the launch of the RTX 50, expanded to 6X mode in March 2026
RTX Mega GeometryAvailable in a few gamesFirst game: Alan Wake 2 (update dated 31 January 2025), path tracing 15 to 20% faster
Reflex 2 with Frame WarpStill not availableAnnounced for The Finals and Valorant, but no game was using it by the end of July 2026
Neural texture compression (NTC)Development kit in betaBeta version 0.10 in September 2026; no commercial game uses it yet
4:2:2 encodingAvailableSupported in compatible video editing software

Sources: NVIDIA, TweakTown, Hardware Busters (July 2026), RTXNTC GitHub repository and Wccftech (September 2026).

The GeForce RTX 50 range

All RTX 50 desktop cards use the Blackwell architecture. The prices below are US launch prices, in dollars and before tax: they’re meant to place each card within the range, not to estimate what you’ll pay today.

The choice depends mostly on your screen and video memory: 8 GB is getting tight in some recent games, which is why we recommend the RTX 5060 Ti with 16 GB rather than the 8 GB model. Everything is explained in RTX 5060 or RTX 5060 Ti: 8 or 16 GB.

Close-up of a graphics chip and its GDDR7 memory modules

GDDR7 is one of the main hardware upgrades of the RTX 50.

CardCUDA coresMemoryLaunch price (United States)Release
RTX 509021,76032 GB GDDR7$1,999January 2025
RTX 508010,75216 GB GDDR7$999January 2025
RTX 5070 Ti8,96016 GB GDDR7$749February 2025
RTX 50706,14412 GB GDDR7$549March 2025
RTX 5060 Ti4,60816 or 8 GB GDDR7$429 or $379April 2025
RTX 50603,8408 GB GDDR7$299May 2025
RTX 50502,5608 GB GDDR6$249July 2025

Sources: NVIDIA (press releases and product sheets), Tom's Hardware, TweakTown.

And what next? The RTX 50 Super models never materialised: according to The Information (February 2026), their launch was pushed back with no date set due to a memory shortage, with NVIDIA only saying that memory supply was constrained. On the data-centre side, the next generation, Vera Rubin, entered production in 2026. We’re covering the story in RTX 50 Super: where are we at?.

Our Blackwell gaming PCs

FlowUP Swift gaming PC with GeForce RTX 5060 Ti 16 GB

Swift PC: RTX 5060 Ti 16 GB

The 16 GB version of the RTX 5060 Ti, to make the most of DLSS 4 and Multi Frame Generation at 1080p and 1440p.

See the PC

FlowUP Novara gaming PC with GeForce RTX 5070 and Ryzen 5 7500X3D

Novara PC: RTX 5070 and X3D

RTX 5070 and Ryzen 5 7500X3D: 1440p with DLSS 4, plus a 3D-cache processor for steady FPS.

See the PC

FlowUP Orion gaming PC with GeForce RTX 5080 in a HYTE Y60 case

Orion PC: RTX 5080 for 4K

RTX 5080 and Ryzen 7 X3D: 4K with ray tracing and Multi Frame Generation.

See the PC

Before installing an RTX 50 in an existing PC, check the power supply and the case: our guide RTX 5000: what power supply, which case, and which motherboard? takes a look.

FAQ: Blackwell architecture

What is NVIDIA Blackwell architecture?

It’s the generation of NVIDIA graphics chips that succeeds Ada Lovelace (RTX 40). It powers the GeForce RTX 50 and, in another form, the B200 and B300 AI accelerators.

Which cards use Blackwell?

All desktop GeForce RTX 50 cards: RTX 5050, 5060, 5060 Ti, 5070, 5070 Ti, 5080 and 5090, as well as the RTX 50 for laptops.

What’s the difference between Blackwell and Ada Lovelace?

Tensor cores capable of computing in FP4, enabling Multi Frame Generation; RT cores optimised for complex geometry; GDDR7; PCIe 5.0; and DisplayPort 2.1b. The fabrication process remains within the same family.

Does Multi Frame Generation work on the RTX 40?

No. The RTX 40 generate a single intermediate frame; multi frame generation is reserved for the RTX 50.

Is Reflex 2 available?

Not yet as of 29 September 2026: no game used Frame Warp in late July 2026, and NVIDIA still lists it as “coming soon”.

Sources

Intel Arc Battlemage 24GB: Arc Pro B60, B70 and the entire lineup in 2026
Read Next