How much faster is the RTX 5090 than the 4090 for LLMs?

1792 vs 1008 GB/s — theoretically +78% decode: 8B Q4 ~154→~273 tok/s, Qwen3 32B Q4 ~38→~67 tok/s. The 32GB VRAM also moves the whole 27-32B tier from tight to comfortable.

Can the RTX 5090 run 70B?

Not on one card: 70B Q4 needs ~47.4GB, over 32GB — you need two. The extra 8GB over the 4090 buys quant headroom in the 27-32B tier (Qwen3 32B can attempt Q6_K at ~30.6GB, tight), not a new model class.

Buy a 4090 or a 5090 right now?

Neither is rational at street prices: used 4090s run $2,350-2,500 (discontinuation premium), 5090s actually cost $3,590-4,300 (~2× MSRP). At MSRP, take the 5090; otherwise two used 3090s ($2,200, 48GB) beat either card for pure LLMs.

GPU Comparisons / RTX 4090 vs RTX 5090

RTX 4090 vs RTX 5090 for Local LLMs: Which Should You Pick

Verdict

If you can buy at MSRP, take the 5090: the 32B tier moves from tight to comfortable and decode is 78% faster. At street prices neither card is rational — the 4090 carries a discontinuation premium, the 5090 trades at ~2× MSRP.

Spec Comparison

	RTX 4090	RTX 5090
Nominal VRAM	24 GB	32 GB
Usable VRAM (for models)	24.0 GB	32.0 GB
Memory bandwidth	1008 GB/s	1792 GB/s
Type	Discrete GPU	Discrete GPU
Release year	2022	2025
MSRP	$1,599	$1,999
Used reference price	$2,400	—
TDP	450 W	575 W

Model Verdict Matrix (Q4_K_M @8K)

Model	VRAM needed	RTX 4090	Theoretical speed	RTX 5090	Theoretical speed
Llama 3.2 3B	4.4 GB	Comfortable	≈385	Comfortable	≈684
Llama 3.1 8B	7.5 GB	Comfortable	≈154	Comfortable	≈273
Qwen3 8B	7.7 GB	Comfortable	≈151	Comfortable	≈268
Phi-4 14B	12.2 GB	Comfortable	≈84	Comfortable	≈149
Mistral Small 3.2 24B	17.5 GB	Comfortable	≈51	Comfortable	≈91
Gemma 3 27B	22.4 GB	Tight fit	≈45	Comfortable	≈80
Qwen3.8 27B	20.7 GB	Tight fit	≈44	Comfortable	≈79
Muse Glimmer 30B	20.1 GB	Tight fit	≈42	Comfortable	≈74
Qwen3 30B-A3B	21.0 GB	Tight fit	≈374	Comfortable	≈665
Qwen3 32B	23.7 GB	Tight fit	≈38	Comfortable	≈67
gpt-oss-20b	14.7 GB	Comfortable	≈343	Comfortable	≈610
Llama 3.3 70B	47.4 GB	Not feasible	—	Not feasible	—
gpt-oss-120b	73.6 GB	Not feasible	—	Not feasible	—
DeepSeek-R1 671B	413.1 GB	Not feasible	—	Not feasible	—

Highlighted rows = watershed models where the two cards' verdicts differ. Theoretical speed = bandwidth × 0.75 ÷ per-token weight bytes (active params for MoE); real-world results vary with framework/driver/CPU, ±30%. Speed is hidden for multi-GPU/infeasible verdicts.

Value Comparison

	RTX 4090	RTX 5090
Bandwidth per dollar	0.42 GB/s/$	0.90 GB/s/$
Usable VRAM per dollar	0.010 GB/$	0.016 GB/$

Pricing basis: RTX 4090 = used reference price; RTX 5090 = MSRP (no used price yet); used prices are market estimates, see analysis for volatility.

Analysis

The 5090's upgrade over the 4090 is twofold. Bandwidth jumps 1008→1792 GB/s (+78%), lifting decode proportionally across the board: 8B Q4 theoretically ~154→~273 tok/s, Qwen3 32B Q4 ~38→~67 tok/s — the 32B tier finally moves from 'readable' to 'fluid chat'. VRAM grows 24→32GB, moving the entire 27-32B class (Gemma 3 27B, Qwen3.8 27B, Muse Glimmer 30B, Qwen3 30B-A3B, Qwen3 32B) from tight to comfortable; Qwen3 32B can even attempt Q6_K (~30.6GB, tight), and Gemma 3 27B's heavy KV cache stops being a long-context worry.

But both prices are distorted: the 4090 is discontinued, with August 2026 used prices of $2,350-2,500 — over 50% above MSRP; the 5090's $1,999 MSRP is a fiction, with new cards actually at $4,288-4,381 and used at $3,590-4,304 (~2.1× MSRP). Hence an ironic outcome: at real transaction prices the 4090's 0.42 GB/s/$ is the worst of the pair, while the 5090's 0.90 looks great — if you can actually buy at MSRP.

Verdict in three cases: at or near MSRP, the 5090 is the only answer for single-card 32B and beats the 4090; at ~$4,300 street price for pure LLMs, two used 3090s ($2,200, 48GB) offer more VRAM at half the cost; existing 4090 owners should not upgrade — the VRAM tier didn't change, only the speed did. Wait for the next capacity jump.

FAQ

How much faster is the RTX 5090 than the 4090 for LLMs?: 1792 vs 1008 GB/s — theoretically +78% decode: 8B Q4 ~154→~273 tok/s, Qwen3 32B Q4 ~38→~67 tok/s. The 32GB VRAM also moves the whole 27-32B tier from tight to comfortable.
Can the RTX 5090 run 70B?: Not on one card: 70B Q4 needs ~47.4GB, over 32GB — you need two. The extra 8GB over the 4090 buys quant headroom in the 27-32B tier (Qwen3 32B can attempt Q6_K at ~30.6GB, tight), not a new model class.
Buy a 4090 or a 5090 right now?: Neither is rational at street prices: used 4090s run $2,350-2,500 (discontinuation premium), 5090s actually cost $3,590-4,300 (~2× MSRP). At MSRP, take the 5090; otherwise two used 3090s ($2,200, 48GB) beat either card for pure LLMs.

Related guides

Data verified 2026-09-01