NVIDIA24 GB VRAM1,008 GB/s$0.4/hr

GPT-OSS 20B (MoE) on RTX 4090 24GB: Does It Fit, How Fast, What It Costs

GPT-OSS 20B (MoE) at FP8 (E4M3) needs 20.9 GB for its weights and 22.6 GB in total at 8K context with 8 concurrent users. That means it fits on a single RTX 4090 24GB. Every figure is computed by tokcalc's open-source capacity model.

Fits?
Yes, 1 GPU
22.6 GB needed / 24 GB available
Decode speed
273.0
tok/s per stream
Aggregate
3276
tok/s @ 8 concurrent
Cost / 1M tokens
$0.0339
1× RTX 4090 24GB
What changes as you add RTX 4090 24GBs
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× RTX24 GB22.6 GB273.032761.46 s$0.0339✓
2× RTX48 GB22.6 GB464.111138746.2 ms$0.0200✓
4× RTX96 GB22.6 GB928.244554388.1 ms$0.009976✓
8× RTX192 GB22.6 GB1856178214209.1 ms$0.004988✓
Concurrency vs context length
On 1× RTX 4090 24GB at FP8 (E4M3). KV cache scales linearly with context, so this is the number of users your GPUs can actually serve.
ContextConcurrent usersKV per requestFits?
4K29206.2 MB✓ fits
8K15412.3 MB✓ fits
33K31.61 GB✓ fits
66K13.22 GB✓ fits
131K06.44 GB✗ does not fit
Other GPUs that run GPT-OSS 20B (MoE)
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
RTX 5090 32GB1×5824$0.0238
RTX 3090 24GB1×2472$0.0281
AMD MI300X 192GB1×17225$0.0403
V100 SXM2 32GB1×2925$0.0475
A100 40GB SXM41×5038$0.0496
AMD MI325X 288GB1×19500$0.0499
Frequently asked questions

Can GPT-OSS 20B (MoE) run on RTX 4090 24GB?

Yes. GPT-OSS 20B (MoE) at FP8 (E4M3) needs 22.6 GB including the KV cache, and RTX 4090 24GB has 24 GB — so it fits on a single GPU.

How fast is GPT-OSS 20B (MoE) on RTX 4090 24GB?

273.0 tokens/sec per stream and 3276 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 1.46 s, and inter-token latency is 3.7 ms.

What does GPT-OSS 20B (MoE) on RTX 4090 24GB cost per million tokens?

$0.0339 per 1M output tokens at 1× RTX 4090 24GB ($0.4000/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.

How much VRAM does GPT-OSS 20B (MoE) need on RTX 4090 24GB?

20.9 GB for the weights at FP8 (E4M3), plus 1.6 GB of KV cache for 8 concurrent requests at 8K context — 22.6 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is RTX 4090 24GB the right choice for GPT-OSS 20B (MoE)?

It works. Compare $0.0339 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 4× RTX 3080 10GB at $0.006615, and the fastest is 1× Cerebras CS-3 (WSE-3) at 65000 tok/s.

Adjust the workload yourself

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

Open in calculator