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

Llama 4 Scout 17B-16E (MoE, VLM) on RTX 5090 32GB: Does It Fit, How Fast, What It Costs

Llama 4 Scout 17B-16E (MoE, VLM) at AWQ 4-bit needs 60.0 GB for its weights and 72.8 GB in total at 8K context with 8 concurrent users. That means it needs 4× RTX 5090 32GB. Every figure is computed by tokcalc's open-source capacity model.

Fits?
Yes, 4×
72.8 GB needed / 128 GB available
Decode speed
372.7
tok/s per stream
Aggregate
17891
tok/s @ 8 concurrent
Cost / 1M tokens
$0.0311
4× RTX 5090 32GB

RTX 5090 32GB has no NVLink, so 4-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 GB72.8 GB0.0000.000——✗
2× RTX64 GB72.8 GB0.0000.000——✗
4× RTX128 GB72.8 GB372.7178911.36 s$0.0311✓
8× RTX256 GB72.8 GB745.571565694.5 ms$0.0155✓
Concurrency vs context length
On 4× RTX 5090 32GB at AWQ 4-bit. KV cache scales linearly with context, so this is the number of users your GPUs can actually serve.
ContextConcurrent usersKV per requestFits?
4K84824.6 MB✓ fits
8K421.61 GB✓ fits
33K106.44 GB✓ fits
66K512.88 GB✓ fits
131K225.77 GB✓ fits
Other GPUs that run Llama 4 Scout 17B-16E (MoE, VLM)
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
AMD MI300X 192GB1×3891$0.1785
L40S 48GB2×2157$0.1803
AMD MI325X 288GB1×4405$0.2207
A100 80GB SXM41×1468$0.2270
Intel Gaudi 3 128GB1×2423$0.2293
H100 SXM5 80GB1×2459$0.2824
Frequently asked questions

Can Llama 4 Scout 17B-16E (MoE, VLM) run on RTX 5090 32GB?

Yes, but it needs 4× RTX 5090 32GB. At AWQ 4-bit the total footprint is 72.8 GB (60.0 GB of weights plus 12.9 GB of KV cache at 8K × 8 concurrent users), against 128 GB of available VRAM across 4 GPUs.

How fast is Llama 4 Scout 17B-16E (MoE, VLM) on RTX 5090 32GB?

372.7 tokens/sec per stream and 17891 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 1.36 s, and inter-token latency is 2.7 ms.

What does Llama 4 Scout 17B-16E (MoE, VLM) on RTX 5090 32GB cost per million tokens?

$0.0311 per 1M output tokens at 4× 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 Llama 4 Scout 17B-16E (MoE, VLM) need on RTX 5090 32GB?

60.0 GB for the weights at AWQ 4-bit, plus 12.9 GB of KV cache for 8 concurrent requests at 8K context — 72.8 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is RTX 5090 32GB the right choice for Llama 4 Scout 17B-16E (MoE, VLM)?

It works. Compare $0.0311 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 3080 10GB at $0.0146, and the fastest is 2× Cerebras CS-3 (WSE-3) at 49920 tok/s.

Adjust the workload yourself

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

Open in calculator