RTX 5090 workstation (1x 32 GB) vs Used RTX 4090 workstation (24 GB)
RTX 5090 has 32 GB memory, 1,792 GB/s bandwidth, 700 W nominal power; RTX 4090 has 24 GB memory, 1,008 GB/s bandwidth, 600 W nominal power. Compare model compatibility, like-for-like tokens/sec and ownership cost below.
RTX 5090 workstation (1x 32 GB) · Used RTX 4090 workstation (24 GB)
Verdict
- throughputThe RTX 5090 workstation (1x 32 GB) delivers approximately 57% higher aggregate decode throughput across 5 like-for-like model configurations (same quantization, runtime, context and inference settings).
- capacityBoth machines run all 8 compared models in their best practical configurations. Compare their quantizations and evidence labels before treating any throughput difference as a hardware result.
- costAt 8 hours a day, the Used RTX 4090 workstation (24 GB) has approximately 41% lower total monthly cost ($27 against $46), for a C corporation in United States (federal).
- evidence1 of 5 like-for-like rows compare measured evidence on both machines; the remaining comparable rows use matched estimates.
- recommendationThere is no single answer: the RTX 5090 workstation (1x 32 GB) is faster, the Used RTX 4090 workstation (24 GB) is cheaper to own, and the Used RTX 4090 workstation (24 GB) gives more throughput per unit of purchase currency. Pick on whichever constraint actually binds you — interactive latency, monthly budget, or capacity headroom for larger models later.
Every statement above is generated from the calculations on this page by fixed rules — no prose is written for individual comparisons. If the numbers change, the verdict changes with them.
Hardware
| RTX 5090 workstation (1x 32 GB) | Used RTX 4090 workstation (24 GB) | |
|---|---|---|
| Purchase price estimate | $3,332 * | $2,071 * |
| Memory | 32 GB | 24 GB |
| Memory bandwidth | 1,792 GB/s | 1,008 GB/s |
| Usable bandwidth indicator | 1,419 GB/s (79%) | 806 GB/s (80%) |
| Nominal power | 700 W | 600 W |
| Measured load power | 640 W | 520 W |
| Idle power | 70 W | 60 W |
| Architecture | discrete gpu | discrete gpu |
Like-for-like performance5 comparable models
Ternary Bonsai 2 27BNemotron 3.5 Lightning 30B-A3BQwen3-Coder 30B-A3BQwen3 30B-A3BGemma 4 26B-A4Bgpt-oss-20bDeepSeek Coder V2 Lite 16B-A2.4BGranite 4.2 30B
| Model | RTX 5090 workstation (1x 32 GB) | Used RTX 4090 workstation (24 GB) | Winner |
|---|---|---|---|
| Ternary Bonsai 2 27B 27.36B | 130 t/s PQ2_0 | 81.2 t/s PQ2_0 | RTX 5090 workstation (1x 32 GB) +60% measured · Prism ML llama.cpp fork · 0K context |
| Nemotron 3.5 Lightning 30B-A3B 31.578B (3B active) | ~355–511 t/s Q4_K_M | ~237–342 t/s Q4_K_M | RTX 5090 workstation (1x 32 GB) +50% estimated · estimator_v1 · 8K context |
| Qwen3-Coder 30B-A3B 30.532B (3.3B active) | ~333–480 t/s Q4_K_M | ~221–317 t/s Q4_K_M | RTX 5090 workstation (1x 32 GB) +51% estimated · estimator_v1 · 8K context |
| Qwen3 30B-A3B 30.532B (3.3B active) | — no matched configuration | — no matched configuration | Not comparable |
| Gemma 4 26B-A4B 25.806B (3.8B active) | ~270–388 t/s Q5_K_M | ~173–249 t/s Q5_K_M | RTX 5090 workstation (1x 32 GB) +56% estimated · estimator_v1 · 8K context |
| gpt-oss-20b 20.915B (3.6B active) | — no matched configuration | — no matched configuration | Not comparable |
| DeepSeek Coder V2 Lite 16B-A2.4B 15.706B (2.4B active) | — no matched configuration | — no matched configuration | Not comparable |
| Granite 4.2 30B 29.277B | ~61.6–88.7 t/s Q4_K_M | ~35.9–51.7 t/s Q4_K_M | RTX 5090 workstation (1x 32 GB) +71% estimated · estimator_v1 · 8K context |
Hardware deltas are calculated only when quantization, runtime, exact context, execution/decode mode and relevant inference settings match. Measured and estimated results are never compared to each other. A difference under 10% is reported as a tie.
Best practical configuration
| Model | RTX 5090 workstation (1x 32 GB) | Used RTX 4090 workstation (24 GB) | Practical result |
|---|---|---|---|
| Ternary Bonsai 2 27B | 121 t/s PTQ1_0 · measured | 91.1 t/s PTQ1_0 · measured | Both fit; configurations may differ |
| Nemotron 3.5 Lightning 30B-A3B | ~319–459 t/s Q5_K_M · estimated | ~237–342 t/s Q4_K_M · estimated | Both fit; configurations may differ |
| Qwen3-Coder 30B-A3B | ~299–430 t/s Q5_K_M · estimated | ~221–317 t/s Q4_K_M · estimated | Both fit; configurations may differ |
| Qwen3 30B-A3B | ~299–430 t/s Q5_K_M · estimated | ~221–317 t/s Q4_K_M · estimated | Both fit; configurations may differ |
| Gemma 4 26B-A4B | ~270–388 t/s Q5_K_M · estimated | ~197–284 t/s Q4_K_M · estimated | Both fit; configurations may differ |
| gpt-oss-20b | 419 t/s MXFP4 · measured | ~228–329 t/s MXFP4 · estimated | Both fit; configurations may differ |
| DeepSeek Coder V2 Lite 16B-A2.4B | — Q8_0 · estimated | — Q8_0 · estimated | Both fit; configurations may differ |
| Granite 4.2 30B | ~61.6–88.7 t/s Q4_K_M · estimated | ~35.9–51.7 t/s Q4_K_M · estimated | Both fit; configurations may differ |
This is a deployment comparison: each machine uses its own recommended configuration. Different quantizations can change speed and quality, so these figures are not raw hardware benchmarks.
Economics — United States (federal), C corporation, 8h/day
| RTX 5090 workstation (1x 32 GB) | Used RTX 4090 workstation (24 GB) | |
|---|---|---|
| Purchase price estimate | $3,332 | $2,071 |
| Reclaimable VAT | $0 | $0 |
| Net cash at purchase | $3,332 | $2,071 |
| Monthly depreciation | $47.21 | $26.92 |
| Monthly electricity | $4.53 (32 kWh) | $3.75 (27 kWh) |
| Tax programme | §179 100.0% | §179 100.0% |
| Tax-programme benefit | $700 | $435 |
| Monthly economic cost | $45.51 | $26.84 |
| Total over 5 years | $2,731 | $1,611 |
Purchase prices are localized inputs in USD: tax/VAT excluded for RTX 5090 workstation (1x 32 GB) and tax/VAT excluded for Used RTX 4090 workstation (24 GB). Net cash and monthly economic cost then apply the selected United States (federal) / C corporation tax treatment. Tax assumptions last verified 2026-09-07. Federal planning estimate only, not tax advice. State and local income tax, sales/use tax and incentives are excluded. It assumes 100% business use, a Section 179 election, enough business income to use it, and that the full annual limit remains available.