TextDense9B params8K max context

Gemma 2 9B VRAM Requirements: How Many GPUs & What It Costs

Gemma 2 9B needs 18.0 GB of VRAM for its weights at FP16 — 19.4 GB once a KV cache is loaded. Below is the full breakdown: which GPUs fit, how many you need, what it serves, and what it costs per million tokens. Every figure comes from tokcalc's open-source capacity model.

Weights (FP16)
18.0 GB
19.4 GB with KV cache
Fits on 1 GPU?
Yes
L40S 48GB or larger
Cheapest config
$0.0248
8× RTX 3080 10GB / 1M tokens
Fastest config
17916
tok/s on 8× RTX 3080 10GB
VRAM requirement by quantization
Weight footprint = params × bytes per param. The “with KV cache” column adds one request at 4K context.
QuantizationWeightsWith KV cacheSmallest single GPU
FP16recommended18.0 GB19.4 GBA10G 24GB (24 GB)
FP8 (E4M3)9.0 GB10.4 GBT4 16GB (16 GB)
AWQ 4-bit5.0 GB6.4 GBRTX 3080 10GB (10 GB)
Which GPU should you use?
Reference workload: 8K context, 8 concurrent users, continuous batching, FP16. Ranked by fitting first, then by tokens per dollar-hour.
GPUGPUsTotal VRAMDecode tok/sAggregate tok/sCost / 1M tokens
RTX 3080 10GB8×80 GB186.617916$0.0248
RTX 4080 16GB4×64 GB88.034226$0.0789
RTX 3090 24GB2×48 GB57.461379$0.1007
RTX 5090 32GB2×64 GB110.02640$0.1052
T4 16GB4×64 GB39.291886$0.1178
RTX A4000 16GB4×64 GB55.002640$0.1263
RTX 4090 24GB2×48 GB61.881485$0.1496
V100 SXM2 32GB2×64 GB55.251326$0.2095
A100 40GB SXM42×80 GB95.152284$0.2189
A10G 24GB2×48 GB36.83884.0$0.2262
AMD MI300X 192GB1×192 GB191.42297$0.3024
TPU v6e Trillium (32GB)2×64 GB50.281207$0.3683

RTX 3080 10GB has no NVLink, so 8-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.

How concurrency changes with context length
On 8× RTX 3080 10GB at FP16. KV cache scales linearly with context, so long contexts cut the number of users you can serve from the same GPUs.
ContextConcurrent usersKV per requestFits?
4K431.41 GB✓ fits
8K212.82 GB✓ fits
33K511.27 GB✓ fits
66K222.55 GB✓ fits
131K145.10 GB✓ fits
Frequently asked questions

Can Gemma 2 9B run on a single GPU?

Yes. At FP16 the model needs 19.4 GB of VRAM, so it fits on a single L40S 48GB or anything with more VRAM. At full FP16 it needs 19.4 GB, which fits on one A10G 24GB or larger.

How many GPUs does Gemma 2 9B need?

At FP16 with 8,192-token context and 8 concurrent users, the cheapest configuration that fits is 8× RTX 3080 10GB (80 GB total VRAM). $0.0248 per 1M output tokens.

How fast can Gemma 2 9B be served?

On 8× RTX 3080 10GB at FP16, the aggregate throughput is 17916 tokens/sec with continuous batching, at $0.0248 per 1M output tokens.

How much VRAM does Gemma 2 9B need?

Weights alone: 18.0 GB at FP16, 18.0 GB at FP16. Add KV cache for your context length and concurrency — every concurrent request at 8,192 tokens costs extra VRAM, so budget headroom above the weight figure.

Is it cheaper to self-host Gemma 2 9B than to call an API?

At 8× RTX 3080 10GB the marginal cost is $0.0248 per 1M output tokens, before you add idle capacity, ops staff, or egress. That is typically cheaper than frontier API pricing at sustained volume, but API calls win when your load is bursty or too small to keep a GPU busy. The Build vs Buy tab on tokcalc prices this against a specific provider for you.

Try Gemma 2 9B with your own context and concurrency

The calculator uses the exact same formulas as this page — change the workload and watch every number update.

Open in calculator

Gemma 2 9B on specific GPUs

Gemma 2 9B on RTX 5090 32GBGemma 2 9B on AMD MI300X 192GBGemma 2 9B on AMD MI325X 288GBGemma 2 9B on A100 80GB SXM4Gemma 2 9B on Intel Gaudi 3 128GBGemma 2 9B on H100 SXM5 80GBGemma 2 9B on B200 SXM 192GBGemma 2 9B on L40S 48GB
Model architecture
The numbers behind the KV-cache and throughput math.
Total parameters9.00B
Active parameters per token9.00B
ArchitectureDense
Attention layoutFull attention (GQA)
Transformer layers42
Hidden dimension3584
Attention heads16 query / 8 KV (GQA)
Head dimension256
Vocabulary size256,000
Max context (catalog)8K
KV cache per token @ 8K336.0 KB

Every layer caches keys and values for the whole context, so the KV cache grows linearly with conversation length.