Phi-4 14B on Quad RTX 5090 workstation (4x 32 GB)
14.66B on 128 GB at 7,168 GB/s. Hardware details · Model details
Yes. Phi-4 14B fits on Quad RTX 5090 workstation (4x 32 GB) at the recommended
Q8_0 configuration, requiring approximately 19.0 GB at 8K context. Practical context capacity is 16K. Expected decode for the recommended configuration is ~128–185 t/s. EstimatedCompatibilityComfortable
Calculated fit
Yes
Calculated Uses under 80% of usable memory. Room for a long context and other work.
Recommended quantization
Q8_0
Calculated gguf
Memory required
19.0 GB
Calculated of 112.6 GB usable — 17%
Max practical context
16K
Calculated model supports 16K
Memory budget at 8K context
Model weights ⓘ
14.8 GB Calculated
KV cache ⓘ
1.6 GB Calculated
Runtime overhead ⓘ
2.6 GB Estimated
Total required
19.0 GB Calculated
Headroom ⓘ
93.6 GB Calculated
Highest-precision quantization that leaves headroom: uses 17% of usable memory at 8K context.
4 GPUs providing 128 GB aggregate VRAM. Assumes tensor- or layer-parallel sharding; each GPU carries its own context and communication buffers.
Multi-GPU throughput depends heavily on interconnect. Without NVLink or NVSwitch, tensor parallelism over PCIe adds latency that partially offsets the extra bandwidth.
Performance
Estimated performance · recommended Q8_0
Decode, prefill and TTFT below are estimates for Q8_0. We do not have a comparable Q8_0 measurement on this machine.
Estimated decode
~128–185 t/s
Estimated Q8_0; calculated range
Estimated prefill
~6830–14190 t/s
Estimated Q8_0; calculated range
Estimated TTFT at 8K
~0.6–1.3 s
Estimated Q8_0; calculated range
Hardware load reference
2250 W
Measured machine-level load; not this model run
How the estimate is calculated
- Multi-GPU: 1792 GB/s per card, with each additional card contributing 40% of its bandwidth — 3942 GB/s effective, not the 7168 GB/s aggregate. Cross-GPU collectives use PCIe rather than a dedicated GPU fabric.
- Decode: reading 14.7B active parameters at 8.50 bits/weight takes 4.99 ms at 3942 GB/s x 79% achieved efficiency.
- Per-token overhead of 1.4 ms (kernel launches, attention bookkeeping, sampling) is significant here — this model is not purely bandwidth-bound on this hardware.
- Prefill: 838 TFLOPS (FP16) x 2 for native FP8 tensor cores x 0.71 calibrated against measured prefill on this platform x 26% assumed model-FLOPs utilisation, divided by 2 x 14.7B parameters per token.
- With few active parameters on fast memory, fixed per-token overhead rather than bandwidth sets the ceiling. Real engines vary widely in how well they hide it.
- Prefill throughput is highly engine-dependent. Flash attention, batch size and quantized KV all move this number substantially, and for sparse mixture-of-experts models it is the least reliable figure we produce.
- This is a calculated estimate, not a measurement. It assumes a single request, a short prompt, no speculative decoding and a warm model already resident in memory.
Estimator version estimator_v1. Stored with every estimated row so old estimates can be regenerated when the model improves.
Submit a benchmarkSubmissions stay pending until reviewed.
This model at other quantizations on this machine
| Quantization | Resident | Disk | Total needed | Utilisation | Fit | Max context |
|---|---|---|---|---|---|---|
Q4_K_M | 8.5 GB | 8.5 GB | 12.5 GB | 11% | Comfortable | 16K |
Q5_K_M | 10.0 GB | 10.0 GB | 14.1 GB | 12% | Comfortable | 16K |
Q8_0recommended | 14.8 GB | 14.8 GB | 19.0 GB | 17% | Comfortable | 16K |
BF16 | 27.3 GB | 27.3 GB | 31.9 GB | 28% | Comfortable | 16K |
Cost of running Phi-4 14B on this machineUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$170
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $70 more
Net cash at purchase
$12,611
Calculated VAT not reclaimable
Total over 5 years
$10,224
Calculated after tax, after resale
Cost per USD/1M tokens
$8.59
Calculated 19.8M tokens/month
The $100 comparison uses the Codex Pro 5x / Claude Max 5x plans. A fixed planning conversion is used for USD. This compares monthly spend only: subscriptions have usage limits, local hardware has different capabilities and constraints, and taxes or regional pricing may change the charged amount. Prices checked 7 September 2026.
Purchase-price input: $12,611 in United States (federal), tax/VAT excluded · estimated · NL retail (system build) (converted catalogue estimate). This is the localized purchase input; the after-tax economic cost is calculated separately below.
Monthly breakdown
| Depreciation 10,719 over 5 years, straight-line to a 1,892 residual | $178.65 |
| Electricity 106.0 kWh/month at 0.140/kWh | $14.83 |
| Cost of capital 4.0%/yr on 7,251 average capital employed | $24.17 |
| Monthly cost before tax | $217.65 |
| Electricity tax shield Running costs are deductible business expenses | −$3.11 |
| First-year expensing §179 (100.0%) | −$44.14 |
| Monthly economic cost after tax | $170.40 |
Three different numbers, deliberately
Cash cost
$12,611
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$178.65/month
$10,719 written down over 5 years to a $1,892 residual.
After-tax economic cost
$170.40/month
Depreciation plus running costs plus cost of capital, less the tax those deductions save. This is the figure to compare between machines.
Assumptions and sources (verified 2026-09-07)
VAT rate
0% Official spec
VAT recoverable
0% Official spec
Federal C-corporation estimate at the flat 21% rate. Pass-through entities should use the sole-proprietor estimate as a rougher proxy.
Effective deduction rate
21.00% Calculated
Headline marginal rate 21.00%.
Depreciation
5 years, straight-line Assumption
Residual value
$1,892 (15%) Assumption
Two GPU generations later, the card is worth a fraction of its list price.
Electricity
$0.140/kWh Assumption
Average power draw
602 W Calculated
Load 2250 W for 20% of powered hours, idle 190 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $2,648 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
$24.17/month Assumption
- VAT rate: No US federal VAT · verified 2026-09-07
- Marginal tax rate: IRS Publication 542 — Corporations · verified 2026-09-07
- VAT recoverable fraction: site assumption · verified 2026-09-07
- Useful life: IRS Publication 946 — How To Depreciate Property · verified 2026-09-07
- Electricity price: Site assumption · verified 2026-09-07
Assumes the machine generates tokens 20% of its 176 powered hours per month. Cost per token scales inversely with this number — halve the utilisation and the cost per token doubles.
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.
Efficiency metrics
Decode per $1,000 spent
12.4 t/s
Calculated purchase price only
Decode per $100/month
91.8 t/s
Calculated after-tax ownership cost
Tokens per joule
0.26
Calculated same as tokens/s per watt
USD per 1M output tokens
$8.59
Calculated at 20% utilisation
Local versus hosted APIs
| Hosted model | USD/1M output | Break-even | API at your volume | Verdict | Comparison type |
|---|---|---|---|---|---|
| Qwen3.8 Flash Alibaba Cloud | $0.42 | 110.7M/mo | $31 | API cheaper | Different model |
| DeepSeek-V4.1 Flash DeepSeek | $0.60 | 94.7M/mo | $36 | API cheaper | Different model |
| DeepSeek-V4 Pro DeepSeek | $1.98 | 23.5M/mo | $144 | API cheaper | Different model |
| DeepSeek V4 Pro (Together) Together AI | $4.40 | 8.0M/mo | $420 | Local cheaper | Different model |
| GLM-5.3 Z.ai | $4.40 | 10.9M/mo | $309 | Local cheaper | Different model |
| Claude Haiku 4.5 Anthropic | $5.00 | 13.1M/mo | $258 | Local cheaper | Different model |
Different-model comparison: the hosted model is not the model you would run locally. Treat this as a workload-quality trade-off, not a direct economic equivalence — the frontier model may complete a task in fewer tokens, or complete tasks the local model cannot. Break-even is the monthly output volume at which API spend equals the $170/month economic cost of owning this machine, assuming 8 input tokens per output token. Hosted prices are published in USD. This table is the canonical US default.
Phi-4 14B on other hardware
| Hardware | Decode | Price | Fit |
|---|---|---|---|
| NVIDIA DGX H200 (8x H200, 1,128 GB) | ~416–599 t/s | $359,430 * | Comfortable |
| Lenovo ThinkStation PX (4x RTX PRO 6000, 384 GB) | ~114–164 t/s | $58,553 * | Comfortable |
| Dual RTX 5090 workstation (2x 32 GB) | ~88.7–128 t/s | $5,854 * | Comfortable |
| Dell Precision 7960 Rack (2x RTX PRO 6000, 192 GB) | ~81.8–118 t/s | $36,663 * | Comfortable |
| RTX 5090 workstation (1x 32 GB) | ~69.6–100 t/s | $3,332 * | Comfortable |
| RTX PRO 6000 Blackwell workstation (96 GB) | ~65.6–94.4 t/s | $10,359 * | Comfortable |
Nearest alternatives to the Quad RTX 5090 workstation (4x 32 GB)
Phi-4 14B on Dual RTX 5090 workstation (2x 32 GB)64 GBPhi-4 14B on RTX 5090 workstation (1x 32 GB)32 GBPhi-4 14B on RTX PRO 6000 Max-Q workstation (96 GB, 300 W)96 GBPhi-4 14B on RTX PRO 6000 Blackwell workstation (96 GB)96 GBPhi-4 14B on Dual used RTX 4090 workstation (48 GB)48 GB
Compare Quad RTX 5090 workstation (4x 32 GB) against Dual RTX 5090 workstation (2x 32 GB) →
Compare Quad RTX 5090 workstation (4x 32 GB) against Dual RTX 5090 workstation (2x 32 GB) →