TL;DR: Yes, the rumored NVIDIA RTX 5090 “Blackwell” flagship is slated for a late 2024 reveal, featuring a 192-bit GDDR7 memory bus and a 25% performance uplift over the RTX 4090. Expect a 2-slot design with a single 16-pin power connector, marking a shift toward efficiency over brute-force power draw.
Architecture and Core Specs
NVIDIA’s next-gen Blackwell architecture (GB202 die) moves to TSMC’s 3nm process, enabling 192 streaming multiprocessors (SMs) — up from 128 on the current Ada Lovelace flagship. That yields roughly 24,576 CUDA cores, a 15% increase, but clock speeds are expected to plateau near 2.9 GHz due to thermal limits. The real leap is memory: GDDR7 at 32 Gbps across a 384-bit bus delivers 1.5 TB/s bandwidth (vs. 1.0 TB/s on the 4090). VRAM remains at 24 GB, but with error-correction improvements for AI workloads.
If you want to dig deeper, check out our guide on 10 Easy Lifestyle Changes for a Happier, Healthier Life.
Power and Cooling Redesign
Contrary to earlier rumors of a 600W TDP, NVIDIA has targeted 450W — matching the 4090 — thanks to a dual-die design that splits the GPU into two compute chiplets. This allows for better binning and lower leakage currents. The cooler is a radical departure: a vapor chamber with a rear-exhaust fin stack, similar to the RTX 3090 FE, but thinner (2.2 slots) and quieter at load. The 16-pin 12V-2×6 connector is mandatory, but the board includes an onboard power monitor to prevent melting incidents seen on early 4090 adapters.
Industry Impact and Software Stack
Blackwell introduces “Neural Texture Compression,” which offloads VRAM usage by 30% for texture-heavy games, using dedicated tensor cores for real-time decompression. This is a direct answer to AMD’s upcoming RDNA 4, which lacks a dedicated AI upscaler. On the software side, DLSS 4.0 will feature “Frame Warp” — inserting AI-generated frames between rendered ones at a 3x ratio, reducing input latency by 40% versus DLSS 3.5. For creators, NVENC now supports AV1 4:4:4 12-bit, a first for consumer cards, targeting prosumer video editors.
Industry analysts predict the RTX 5090 will launch at $1,899 (up from $1,599), reflecting inflation and advanced packaging costs. However, NVIDIA is bundling a 3-year warranty and free GeForce Now Priority membership to soften the blow. AMD’s counter, the RX 8900 XTX, is rumored but lacks a concrete release date, giving NVIDIA a 6-month uncontested window in the high-end segment.
FAQ
Q: Will the RTX 5090 fit in my existing ATX case?
A: Yes, the card is 310mm long (same as 4090 FE) but only 2.2 slots thick, so most mid-tower cases with 350mm+ GPU clearance will work, but check that your PSU has a native 16-pin cable (or use the included adapter).
Q: Is GDDR7 memory compatible with current motherboards?
A: Memory is on the GPU itself, so no motherboard change is needed. However, you’ll require PCIe 5.0 for full bandwidth; PCIe 4.0 will work but loses ~3% performance in read-intensive tasks.
Q: Does the 5090 support multi-GPU (NVLink) for AI training?
A: No, NVIDIA has removed NVLink entirely from the GeForce lineup. For multi-GPU LLM work, you must buy the professional RTX 6000 Blackwell (which retains NVLink) or use PCIe peer-to-peer via
Leave a Reply