VRAM
16 GB
TDP
360 W
MSRP
$999
Perf/W
0.46 emb/s/W
Cost/1K tok
$0.06/M
Tested
2026-02-08
NVIDIA GeForce RTX 5080 16GB produces approximately 7.0 tok/s on Llama 3 70B at Q4_K_M quantization (batch 1, 4096-token context, 16GB VRAM, 360W TDP). That figure comes from 1 measured run on llama.cpp. The 70B model needs roughly 40GB of VRAM at Q4, so headroom and KV cache budget matter as much as raw throughput.
NVIDIA GeForce RTX 5080 is the Blackwell-architecture enthusiast card released in 2025. 16GB GDDR7 at ~960 GB/s bandwidth, 360W TDP. Blackwell compute with FP4 support but capped at 16GB VRAM — the entry point for Blackwell consumer AI.
16GB VRAM limits full-GPU models to ~13B-class at Q4. 32B models require aggressive quantization or partial offload. Real-world: Treydol-LLM-Asure-12B Q4 at ~70 tok/s (verified-lab on this device). Best for 7-13B daily use, coding agents on smaller models, and as a secondary GPU in multi-card rigs. Overkill for sub-13B if you're budget-conscious.
Buy for mixed gaming+AI or as a secondary GPU in multi-card rigs. Skip if your primary use case is local AI — used 3090 gives you 24GB at lower cost, and 5070 Ti 16GB gives similar VRAM for $250 less.
13B-class Q4 full-GPU at 70+ tok/s — excellent for coding agents on Qwen 2.5 Coder 14B or similar
VRAM ceiling — 16GB fills fast with 32B models at useful context. KV cache eats the headroom quickly.
Ubuntu 24.04 (excellent), Windows 11 (excellent), WSL2 (excellent), macOS (unsupported)
360W TGP, ~250-300W sustained decode. 750W PSU minimum. Reasonable thermals in standard cases. The most power-efficient Blackwell card for AI.
NVIDIA GeForce RTX 5080 16GB with 16GB VRAM at 360W TDP, scored across 11 workloads with 15 benchmark records.
Best workload
Embedding throughput
8200 emb/s
Quantization
FP16
1K context · batch 1
| Workload | Score | Quant | Context | σ | Status | Tested |
|---|---|---|---|---|---|---|
| Llama 3 8B Q4 llm | 165.0tok/s | Q4_K_M | 4K | ±4.8 | Curated Aggregate | 2026-02-08 |
| Llama 3 8B FP16 llm | 88.0tok/s | FP16 | 4K | , | Curated Aggregate | 2026-02-12 |
| Llama 3 70B Q4 llm | 7.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2026-02-15 |
| DeepSeek-R1 7B llm | 138.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2026-02-20 |
| Qwen 2.5 14B llm | 86.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2025-03-16 |
| SDXL image gen image | 18.0img/min | FP16 | , | , | Curated Aggregate | 2026-02-25 |
| Whisper transcription audio | 95.0x RT | FP16 | , | , | Curated Aggregate | 2026-03-01 |
| Embedding throughput embedding | 8200.0emb/s | FP16 | 1K | , | Curated Aggregate | 2025-03-04 |
| Mistral 7B llm | 122.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2025-02-15 |
| Gemma 2 9B llm | 105.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2025-02-25 |
| Asure 12B llm | 69.9tok/s | Q4_K_M | 8K | ±1.5 | Lab Verified | 2025-05-01 |
NVIDIA GeForce RTX 5080 16GB clears a 9.1/10 based on workload-normalized throughput, memory headroom, efficiency, value, trust, and coverage.
What models fit this 16GB card at different quantization levels.
Trust score
6/10
MyAI rating
9.1
Runs
1
Freshness
Stale
Source-linked row with explicit verification status.
Tested on 2025-03-04; 541 days old.
Measured on your own machine via ollama-local-api on 2026-05-28.
Batch=1, 2048 context, 512 generated tokens, consecutive warm-model runs. Generation TPS uses Ollama eval_count/eval_duration.
kumru-2b
2.4B / Q4_K_M
median 443.66 / min 397.42 / max 454.33 / n=5
brooqs-mistral-turkish-v2-latest
7.2B / Q4_0
median 161.11 / min 159.88 / max 161.79 / n=5
Best place to start
Retailer we'd check first
Amazon search
For mainstream AI hardware, Amazon usually updates street pricing, seller availability, and shipping speed faster than most comparison sites.
Search Amazon listingsThis is one of the highest-scoring parts in our database, so it is worth tracking when inventory lands near fair-market pricing.
Affiliate note: we do not have a device-level ASIN yet, so this opens tagged Amazon search results for the exact product name.