NVIDIA32 GB VRAM1,792 GB/s$0.5/hr

Phi-4 14B on RTX 5090 32GB: Does It Fit, How Fast, What It Costs

Phi-4 14B at FP16 needs 28.0 GB for its weights and 38.7 GB in total at 8K context with 8 concurrent users. That means it needs 2× RTX 5090 32GB. Every figure is computed by tokcalc's open-source capacity model.

Fits?
Yes, 2×
38.7 GB needed / 64 GB available
Decode speed
70.72
tok/s per stream
Aggregate
1697
tok/s @ 8 concurrent
Cost / 1M tokens
$0.1637
2× RTX 5090 32GB

RTX 5090 32GB has no NVLink, so 2-way tensor parallelism runs over PCIe. tokcalc applies a flat 0.85× multi-GPU efficiency factor for this, which is realistic for NVLink-equipped datacenter cards but optimistic for PCIe scale-out — treat the throughput and cost figures for this configuration as an upper bound, and measure before committing.

What changes as you add RTX 5090 32GBs
Tensor-parallel degree vs VRAM, throughput, and cost. Cost scales linearly with GPU count; decode throughput scales near-linearly too, so cost per token stays roughly flat.
GPUsTotal VRAMVRAM neededDecode tok/sAggregate tok/sTTFTCost / 1M tokFits?
1× RTX32 GB38.7 GB0.0000.000——✗
2× RTX64 GB38.7 GB70.7216972.23 s$0.1637✓
4× RTX128 GB38.7 GB141.467891.13 s$0.0818✓
8× RTX256 GB38.7 GB282.927156579.7 ms$0.0409✓
Concurrency vs context length
On 2× RTX 5090 32GB at FP16. KV cache scales linearly with context, so this is the number of users your GPUs can actually serve.
ContextConcurrent usersKV per requestFits?
4K53687.2 MB✓ fits
8K261.34 GB✓ fits
33K65.37 GB✓ fits
66K310.74 GB✓ fits
131K121.47 GB✓ fits
Other GPUs that run Phi-4 14B
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
AMD MI300X 192GB1×1476$0.4704
A100 40GB SXM41×431.8$0.5790
AMD MI325X 288GB1×1671$0.5817
A100 80GB SXM41×557.1$0.5983
Intel Gaudi 3 128GB1×919.3$0.6043
H100 SXM5 80GB1×933.2$0.7441
Frequently asked questions

Can Phi-4 14B run on RTX 5090 32GB?

Yes, but it needs 2× RTX 5090 32GB. At FP16 the total footprint is 38.7 GB (28.0 GB of weights plus 10.7 GB of KV cache at 8K × 8 concurrent users), against 64 GB of available VRAM across 2 GPUs.

How fast is Phi-4 14B on RTX 5090 32GB?

70.72 tokens/sec per stream and 1697 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 2.23 s, and inter-token latency is 14.1 ms.

What does Phi-4 14B on RTX 5090 32GB cost per million tokens?

$0.1637 per 1M output tokens at 2× RTX 5090 32GB ($0.5000/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.

How much VRAM does Phi-4 14B need on RTX 5090 32GB?

28.0 GB for the weights at FP16, plus 10.7 GB of KV cache for 8 concurrent requests at 8K context — 38.7 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is RTX 5090 32GB the right choice for Phi-4 14B?

It works. Compare $0.1637 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 4× RTX 3080 10GB at $0.0772, and the fastest is 1× Cerebras CS-3 (WSE-3) at 5571 tok/s.

Adjust the workload yourself

Same formulas, your context length, concurrency, and quantisation.

Open in calculator