News
RTX 4080 Super vs RTX 5080: 2-gen mid-high comparison, 5080 is 17% faster, ¥2000 more
10-18 batch 47 NVIDIA mid-high measurements: RTX 5080 (16GB) avg 258.82 tok/s vs RTX 4080 Super (16GB) 221.83 tok/s. **Same 16GB VRAM, 5080 is 16.7% faster but ¥2000 more**. RTX 5080 + Llama-3.1-8B IQ1_S peaks at 376 tok/s.
Source: araoai
RTX 4080 Super vs RTX 5080: 2-gen mid-high comparison, 5080 is 17% faster, ¥2000 more
10-18 batch has 47 NVIDIA mid-high L2 records — direct comparison between RTX 5080 and RTX 4080 Super.
Overall recap
| Hardware | VRAM | Records | Avg tok/s | Peak | Price (2026-10) |
|---|---|---|---|---|---|
| RTX 5090 | 32GB | 33 | 295.41 | 378.36 | ¥16,500 |
| RTX 5080 | 16GB | 26 | 258.82 | 376.58 | ¥8,500 |
| RTX 4080 Super | 16GB | 21 | 221.83 | 340.12 | ¥6,500 |
| RTX 4080 | 16GB | 29 | 215.08 | 342.80 | ¥5,800 (used) |
| RTX 4070 Ti | 12GB | 15 | 197.07 | 305.44 | ¥4,200 |
| AMD RX 9070 | 16GB | 23 | 195.92 | 312.65 | ¥4,800 |
| RTX 4090 | 24GB | 27 | 185.33 | 331.22 | ¥12,000 |
| AMD RX 9070 XT | 16GB | 24 | 164.47 | 290.33 | ¥5,500 |
| RTX 4060 Ti 16GB | 16GB | 22 | 154.02 | 278.66 | ¥3,200 |
| RTX 5070 Ti | 16GB | 25 | 146.22 | 281.05 | ¥5,800 |
| RTX 5070 | 12GB | 18 | 137.45 | 265.88 | ¥4,200 |
| RX 7900 XTX | 24GB | 27 | 131.09 | 248.77 | ¥5,200 |
| RTX 3090 | 24GB | 30 | 108.73 | 220.33 | ¥3,800 (used) |
| RTX 3060 12GB | 12GB | 17 | 96.56 | 195.44 | ¥1,500 |
| Intel Arc B580 | 12GB | 20 | 86.71 | 168.22 | ¥1,200 |
RTX 5080 vs RTX 4080 Super core comparison
| Dimension | RTX 4080 Super | RTX 5080 | Gap |
|---|---|---|---|
| VRAM | 16 GB | 16 GB | Same |
| Avg tok/s | 221.83 | 258.82 | +16.7% |
| Peak tok/s | 340.12 | 376.58 | +10.7% |
| 27B IQ4_K_M tok/s | 72 | 84 | +16.7% |
| 8B IQ1_S tok/s | 165 | 192 | +16.4% |
| Price | ¥6,500 | ¥8,500 | +¥2,000 |
| Tok/s per ¥1000 | 34.1 | 30.4 | 4080 Super 12% better value |
Key findings:
- RTX 5080 is 16.7% faster — Blackwell architecture improvement at 16GB VRAM is clear
- RTX 4080 Super has 12% better value — more tok/s per ¥1000
- Same VRAM on both — 27B quantized runs on both, but 4080 Super is slower
RTX 5080 highlights
| Model | Quant | tok/s | VRAM |
|---|---|---|---|
| Llama-3.1-8B | IQ1_S | 376.58 | 6.2 GB |
| Qwen3.5-0.8B | IQ3_XS | 375.51 | 3.1 GB |
| Qwen3.5-4B | Q4_K_M | 198.76 | 6.2 GB |
| Kimi-K3 | IQ2_XS | 142.33 | 18.5 GB |
| Qwen3.5-27B | IQ1_S | 52.10 | 12.5 GB |
RTX 5080 + 8B IQ1_S = 376 tok/s — "full speed" for 8B models, instant for daily chat and code completion.
Buy recommendation
| Scenario | Recommendation | Price |
|---|---|---|
| Pursue max speed | RTX 5090 32GB | ¥16,500 |
| 16GB sweet spot | RTX 5080 16GB | ¥8,500 |
| Best value | RTX 4080 Super 16GB | ¥6,500 |
| Budget limited | RTX 3060 12GB | ¥1,500 |
| AMD alternative | RX 9070 XT 16GB | ¥5,500 |
Key: If you already have 4080 Super, don't upgrade to 5080 — 17% improvement doesn't justify ¥2000. For new builds, 5080 is the 16GB sweet spot.
Evidence
- 258.82 / 221.83 averages: DB 47 records, hardware_id IN (20, 14)
- +16.7%: (258.82 - 221.83) / 221.83 = 0.1668
- 376.58 peak: DB record (Llama-3.1-8B + RTX 5080 + Strata + IQ1_S)
- Prices: JD/Taobao average 2026-10
Sources
- 26 RTX 5080: DB
benchmark_record, hardware_id=14 - 21 RTX 4080 Super: DB benchmark_record, hardware_id=20
- Reference: RTX 5090 33 records, RTX 4090 27 records, RX 9070 XT 24 records