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

Phi-3 Mini 3.8B on RTX 5090 32GB: Does It Fit, How Fast, What It Costs

Phi-3 Mini 3.8B at FP16 needs 7.6 GB for its weights and 33.4 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×
33.4 GB needed / 64 GB available
Decode speed
260.5
tok/s per stream
Aggregate
6253
tok/s @ 8 concurrent
Cost / 1M tokens
$0.0444
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 GB33.4 GB0.0000.000——✗
2× RTX64 GB33.4 GB260.56253626.8 ms$0.0444✓
4× RTX128 GB33.4 GB521.125013328.4 ms$0.0222✓
8× RTX256 GB33.4 GB1042100050179.2 ms$0.0111✓
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?
4K351.61 GB✓ fits
8K173.22 GB✓ fits
33K412.88 GB✓ fits
66K225.77 GB✓ fits
131K151.54 GB✓ fits
Other GPUs that run Phi-3 Mini 3.8B
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
AMD MI300X 192GB1×5439$0.1277
A100 40GB SXM41×1591$0.1572
AMD MI325X 288GB1×6158$0.1579
A100 80GB SXM41×2053$0.1624
Intel Gaudi 3 128GB1×3387$0.1640
H100 SXM5 80GB1×3438$0.2020
Frequently asked questions

Can Phi-3 Mini 3.8B run on RTX 5090 32GB?

Yes, but it needs 2× RTX 5090 32GB. At FP16 the total footprint is 33.4 GB (7.6 GB of weights plus 25.8 GB of KV cache at 8K × 8 concurrent users), against 64 GB of available VRAM across 2 GPUs.

How fast is Phi-3 Mini 3.8B on RTX 5090 32GB?

260.5 tokens/sec per stream and 6253 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 626.8 ms, and inter-token latency is 3.8 ms.

What does Phi-3 Mini 3.8B on RTX 5090 32GB cost per million tokens?

$0.0444 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-3 Mini 3.8B need on RTX 5090 32GB?

7.6 GB for the weights at FP16, plus 25.8 GB of KV cache for 8 concurrent requests at 8K context — 33.4 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is RTX 5090 32GB the right choice for Phi-3 Mini 3.8B?

It works. Compare $0.0444 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 4× RTX 3080 10GB at $0.0209, and the fastest is 1× Cerebras CS-3 (WSE-3) at 20526 tok/s.

Adjust the workload yourself

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

Open in calculator