How much VRAM does Gemma 4 31B need?

62.5 GB for its weights at 16-bit, 31.3 GB at 8-bit and 17.6 GB at 4-bit, plus the KV cache: 81.9 KB for each token, 3.52 GB for a conversation of 32K tokens and 22.3 GB at its longest context of 262,144 tokens. The server also keeps about 8% of each GPU free.

Can Gemma 4 31B run on 1 GPU?

Yes, at 4-bit: the smallest single GPU here that holds it with a 32K conversation is the L4.

How many H100s does Gemma 4 31B need?

With 1 conversation of 32K tokens, 1 H100 at 16-bit and 1 H100 at 4-bit. Each H100 has 80 GB, of which vLLM uses 92% by default.

Can I run Gemma 4 31B on an RTX 5090?

The RTX 5090 has 32 GB. With a 32K conversation, Gemma 4 31B needs 1 GeForce RTX 5090 at 4-bit and 4 GeForce RTX 5090s at 16-bit.

How fast does Gemma 4 31B run?

At 4-bit on 1 H100, memory lets 1 conversation get at most 159 tokens a second. Real servers reach part of that ceiling, and each conversation gets less when many run at once.

How much GPU memory does Gemma 4 31B need?

Gemma 4 31B by Google has 31.3B parameters. Its weights take 62.5 GB at 16-bit, 31.3 GB at 8-bit and 17.6 GB at 4-bit; released in BF16, its files are 62.5 GB. Each token of a conversation adds 81.9 KB of KV cache, so a 32K conversation holds 3.52 GB.

Size Gemma 4 31B for your setup Questions

Google · released in BF16

Gemma 4 31B at a glance

Parameters: 31.3B
As released: 62.5 GB; Its BF16 files
At 16-bit: 62.5 GB
At 4-bit: 17.6 GB; NVFP4
KV cache a token: 81.9 KB; At 16-bit
Longest context: 262,144; Tokens

GPUs needed at each number format

For 1 conversation of 32,768 tokens with a 16-bit KV cache, counting 92% of each GPU's memory as vLLM does. Under each count, the most tokens a second memory allows.

GPU	Memory	16-bit	8-bit	4-bit
L4	24 GB	4 GPUs18 tokens/s	2 GPUs17 tokens/s	1 GPU14 tokens/s
GeForce RTX 5090	32 GB	4 GPUs108 tokens/s	2 GPUs103 tokens/s	1 GPU85 tokens/s
L40S	48 GB	2 GPUs26 tokens/s	1 GPU25 tokens/s	1 GPU41 tokens/s
A100 80GB	80 GB	1 GPU31 tokens/s	1 GPU59 tokens/s	1 GPU97 tokens/s
H100	80 GB	1 GPU51 tokens/s	1 GPU96 tokens/s	1 GPU159 tokens/s
RTX PRO 6000 Blackwell	96 GB	1 GPU27 tokens/s	1 GPU51 tokens/s	1 GPU85 tokens/s
DGX Spark	128 GB	1 GPU4.1 tokens/s	1 GPU7.8 tokens/s	1 GPU13 tokens/s
H200	141 GB	1 GPU73 tokens/s	1 GPU138 tokens/s	1 GPU227 tokens/s
B200	180 GB	1 GPU121 tokens/s	1 GPU230 tokens/s	1 GPU379 tokens/s
Instinct MI300X	192 GB	1 GPU80 tokens/s	1 GPU152 tokens/s	1 GPU251 tokens/s
Instinct MI325X	256 GB	1 GPU91 tokens/s	1 GPU172 tokens/s	1 GPU284 tokens/s
Instinct MI355X	288 GB	1 GPU121 tokens/s	1 GPU230 tokens/s	1 GPU379 tokens/s

Models of a similar size

Every model

Questions

How much VRAM does Gemma 4 31B need?: 62.5 GB for its weights at 16-bit, 31.3 GB at 8-bit and 17.6 GB at 4-bit, plus the KV cache: 81.9 KB for each token, 3.52 GB for a conversation of 32K tokens and 22.3 GB at its longest context of 262,144 tokens. The server also keeps about 8% of each GPU free.
Can Gemma 4 31B run on 1 GPU?: Yes, at 4-bit: the smallest single GPU here that holds it with a 32K conversation is the L4.
How many H100s does Gemma 4 31B need?: With 1 conversation of 32K tokens, 1 H100 at 16-bit and 1 H100 at 4-bit. Each H100 has 80 GB, of which vLLM uses 92% by default.
Can I run Gemma 4 31B on an RTX 5090?: The RTX 5090 has 32 GB. With a 32K conversation, Gemma 4 31B needs 1 GeForce RTX 5090 at 4-bit and 4 GeForce RTX 5090s at 16-bit.
How fast does Gemma 4 31B run?: At 4-bit on 1 H100, memory lets 1 conversation get at most 159 tokens a second. Real servers reach part of that ceiling, and each conversation gets less when many run at once.

Sources

Gemma 4 31B at a glance

GPUs needed at each number format

Models of a similar size

Questions

Where each fact comes from