NVIDIA GeForce RTX 5090

The NVIDIA GeForce RTX 5090 is the flagship consumer graphics card of the RTX 50 series, powered by the Blackwell architecture. Announced at CES 2025 on January 6 and launched at $1,999 MSRP, it succeeds the RTX 4090 as the most powerful consumer GPU ever made. Built on TSMC's custom 4N process, the GB202 die packs over 92 billion transistors, 21,760 CUDA cores, and 32 GB of GDDR7 memory on a 512-bit bus.

With 32 GB VRAM and fifth-gen Tensor Cores supporting FP4 and FP8 precision, the RTX 5090 became the new darling of the AI community, capable of running quantized 70B-parameter LLMs on a single card. It introduced DLSS 4 with Multi Frame Generation, DisplayPort 2.1b, and a 575 W TDP.

Built on the GB202 GPU, Nvidia's largest consumer chip ever, fabricated by TSMC on 4N (5 nm-class). The die contains over 92 billion transistors, compared to 76.3 billion on the RTX 4090's AD102.

Spec	RTX 5090	RTX 4090	RTX 3090
Architecture	Blackwell	Ada Lovelace	Ampere
Transistors	92.2B	76.3B	28.3B
CUDA Cores	21,760	16,384	10,496
VRAM	32 GB GDDR7	24 GB GDDR6X	24 GB GDDR6X
Memory Bandwidth	~1,792 GB/s	~1,008 GB/s	~936 GB/s
FP32 Throughput	~104.8 TFLOPS	~82.6 TFLOPS	~35.6 TFLOPS
TDP	575 W	450 W	350 W

The fifth-gen Tensor Cores add native support for FP4 (4-bit floating point) precision - a game-changer for AI inference on consumer hardware. Models quantized to FP4 use half the memory of FP8, theoretically allowing a 70B-parameter model to run on a single 32 GB card. The RTX 5090 delivers up to 3,352 AI TOPS (FP4 sparse).

Unveiled at CES 2025 on January 6, the RTX 5090 launched on January 30, 2025 at $1,999. Stock was immediately wiped out by scalpers, with cards appearing on eBay at $3,500-$8,000+. By mid-2026, street prices on PC part picker still showed the cheapest RTX 5090 at over $5,700 - nearly 3x MSRP. The shortage was exacerbated by GDDR7 supply constraints, massive AI demand, US tariff threats, and TSMC 4N capacity shared with H100/B200 data-center GPUs.

A special RTX 5090D was released for the Chinese market with reduced AI compute capabilities to comply with US export restrictions.

• Transformer-based upscaling - Replaced older CNN model with a vision transformer for sharper images and reduced ghosting.
• Multi Frame Generation (MFG) - Generates up to three interpolated frames per traditionally rendered frame, enabling 3x-6x frame generation multipliers. Exclusive to RTX 50 series.
• 30% less VRAM usage compared to DLSS 3 frame generation.
• 75 titles at launch including Cyberpunk 2077, Alan Wake 2, and Indiana Jones.

With 32 GB GDDR7 and 1.79 TB/s memory bandwidth, the RTX 5090 can run a 70B-parameter model quantized to 4-bit on a single card - pushing right up against the 32 GB limit. Inference benchmarks: Llama 3 8B (FP8) at ~120-150 t/s, Llama 3 70B (4-bit GPTQ) at ~20-30 t/s, Mistral 7B (FP8) at ~180-200 t/s. Fine-tuning up to ~13B with QLoRA (4-bit) and ~32B with aggressive quantization and gradient checkpointing.

The 5th-gen Tensor Cores with FP4/FP8 give it roughly 30-40% higher AI throughput than the RTX 4090 in most workloads.

In pure rasterization, the RTX 5090 delivers roughly 30-40% higher performance than the RTX 4090 at 4K. With DLSS 4 MFG, frame rates can appear 3-5x higher. At 4K with max settings and ray tracing, the 5090 consistently delivers 60+ FPS in even the most demanding titles.

• Availability and Pricing Fiasco - $1,999 MSRP proved meaningless with street prices above $5,000 for over a year. GDDR7 shortage and AI demand kept prices elevated.
• 12V-2x6 Connector Issues - Melting connector reports persisted despite the revised design. 575 W TDP meant any seating issue could cause catastrophic failure.
• Missing ROPs Controversy - Some cards shipped with fewer ROPs than advertised. Nvidia offered replacements.
• 32-bit CUDA/PhysX Removal - Broke compatibility with older games and CUDA applications.
• Black Screen and Driver Instability - Frequent crashes with DisplayPort 2.1b at high refresh rates.
• AI Demand vs. Gaming - Researchers buying 5090s in bulk frustrated gamers, reinforcing criticism that Nvidia had abandoned gaming for AI.

Card	VRAM	Bandwidth	Street Price (2026)	Relative AI Perf
RTX 5090	32 GB GDDR7	1,792 GB/s	$4,000-$6,000	1.0x (baseline)
RTX 4090	24 GB GDDR6X	1,008 GB/s	$1,400-$1,800	~0.7x
H100 SXM	80 GB HBM3	3,350 GB/s	$30,000+ (cloud)	~2-3x
RX 7900 XTX	24 GB GDDR6	960 GB/s	$700-$900	~0.3x

The RTX 5090's legacy is complicated. On paper, it is the most technologically impressive consumer GPU ever produced: Blackwell architecture, GDDR7, DLSS 4, FP4 Tensor Cores, and 32 GB VRAM. But in practice, the $1,999 MSRP proved meaningless, with street prices above $5,000 making it a paper launch. The card was effectively a prosumer/AI card disguised as a gaming GPU. It will likely be remembered as the card that crystallized Nvidia's transition from a gaming company to an AI company.