RTX 5090

The RTX 5090 arrived January 2025 on NVIDIA’s Blackwell architecture, the successor to Ada Lovelace. The VRAM jump from 24 GB to 32 GB is the headline for inference plebs: 70B-class models finally fit at Q8 on a single card with usable context. GDDR7 on a 512-bit bus delivers ~1.8 TB/s — nearly 2× the 4090 — so tok/s scales proportionally on memory-bound workloads.

Who it’s for: professionals and well-funded enthusiasts who need single-card 70B performance without moving to H100/H200 workstation territory.

Models it runs comfortably: Llama 3 70B at Q8 with 8K context, Qwen 2.5 72B at Q4 with 32K context, Mixtral 8x22B at Q3. Also the first consumer card where FP8 training of ~7B LoRAs is genuinely practical.

Hashcenter notes: dual-slot (!) Founders Edition despite 575 W TDP — NVIDIA shifted to a vapor-chamber flow-through design. 575 W is aggressive for a single 120V/15A branch; 120V/20A or 240V strongly recommended. The 12V-2×6 connector evolved from 12VHPWR with better detection — still use a quality cable. Pricing reality check, July 2026: the GDDR7/DRAM shortage has kept street prices around $2,900–$5,000 against the $1,999 MSRP — memory is now the dominant share of the card’s bill of materials, and NVIDIA has passed wholesale increases on to board partners. Blackwell credits go all the way back to the Tesla architecture that started NVIDIA’s compute journey in 2006.

Further reading: This card is a core component of a pleb-grade AI Hashcenter. Need more than 32 GB in one slot? The RTX PRO 6000 Blackwell is the same bandwidth with 96 GB. Pair it with the sovereignty argument in the Sovereign AI for Bitcoiners Manifesto, or look at how the same 120V envelope powers a Bitcoin space heater in our mining catalog. Running both workloads on one rig? See Heating Your Home With Inference.