Skip to content

Gemma 2 2B IT VRAM requirements

How much VRAM Gemma 2 2B IT needs
4-bit weights: 1.5 GBWith 8K context: 3.3 GB8-bit, 8K context: 4.4 GB
Parameters
2.6B
Longest context
8K
Released
Jul 2024
License
Gemma
Access
Gated on the Hub
Strongest at
Not rated

Hugging Face asks you to accept its terms before it serves the weights. Its licence (Gemma) is not a standard open-source one: check the terms before commercial use.

How much VRAM Gemma 2 2B IT needs

Memory Gemma 2 2B IT needs at each precision and context length
PrecisionWeightsWith 8K
16-bit4.9 GB6.7 GB
8-bit2.6 GB4.4 GB
6-bit2.0 GB3.8 GB
5-bit1.7 GB3.6 GB
4-bit1.5 GB3.3 GB
3-bit1.2 GB3.0 GB
2-bit935 MB2.7 GB
1.6-bit514 MB2.3 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Gemma 2 2B IT

The best way Gemma 2 2B IT runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit8K83 tok/s0.2 sPDF: Gemma 2 2B IT on RTX 4060 report
RTX 50608-bit8K131 tok/s0.2 sPDF: Gemma 2 2B IT on RTX 5060 report
RTX 30708-bit8K131 tok/s0.2 sPDF: Gemma 2 2B IT on RTX 3070 report
RTX 3080 10 GB8-bit8K207 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 3080 10 GB report
RTX 3060 12 GB8-bit8K107 tok/s0.3 sPDF: Gemma 2 2B IT on RTX 3060 12 GB report
Arc B5808-bit8K91 tok/s0.2 sPDF: Gemma 2 2B IT on Arc B580 report
RTX 40708-bit8K145 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 4070 report
RTX 4070 Super8-bit8K145 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 4070 Super report
RTX 50708-bit8K187 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 5070 report
RTX 4060 Ti 16 GB8-bit8K87 tok/s0.2 sPDF: Gemma 2 2B IT on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit8K131 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit8K108 tok/s0.2 sPDF: Gemma 2 2B IT on Arc A770 16 GB report
Radeon RX 9070 XT8-bit8K145 tok/s0.1 sPDF: Gemma 2 2B IT on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit8K187 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 4070 Ti Super report
RTX 4080 Super8-bit8K202 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 4080 Super report
RTX 5070 Ti8-bit8K238 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 5070 Ti report
RTX 50808-bit8K251 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 5080 report
Radeon RX 7900 XTX8-bit8K199 tok/s0.1 sPDF: Gemma 2 2B IT on Radeon RX 7900 XTX report
RTX 30908-bit8K246 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 3090 report
RTX 40908-bit8K261 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 4090 report
Radeon AI PRO R97008-bit8K145 tok/s0.1 sPDF: Gemma 2 2B IT on Radeon AI PRO R9700 report
RTX 50908-bit8K401 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 5090 report
2× RTX 30908-bit8K246 tok/s0.1 sPDF: Gemma 2 2B IT on 2× RTX 3090 report
2× RTX 40908-bit8K261 tok/s0.1 sPDF: Gemma 2 2B IT on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K326 tok/s0.1 sPDF: Gemma 2 2B IT on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K251 tok/s0.1 sPDF: Gemma 2 2B IT on RTX 6000 Ada report
L40S8-bit8K231 tok/s0.1 sPDF: Gemma 2 2B IT on L40S report
2× RTX 50908-bit8K401 tok/s0.1 sPDF: Gemma 2 2B IT on 2× RTX 5090 report
4× RTX 30908-bit8K246 tok/s0.1 sPDF: Gemma 2 2B IT on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K401 tok/s0.1 sPDF: Gemma 2 2B IT on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit8K42 tok/s2.1 sPDF: Gemma 2 2B IT on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit8K42 tok/s2.1 sPDF: Gemma 2 2B IT on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K42 tok/s2.1 sPDF: Gemma 2 2B IT on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit8K86 tok/s1.1 sPDF: Gemma 2 2B IT on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K86 tok/s1.1 sPDF: Gemma 2 2B IT on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K86 tok/s1.1 sPDF: Gemma 2 2B IT on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K136 tok/s0.5 sPDF: Gemma 2 2B IT on M4 Max 128 GB report
M5 Max 128 GB8-bit8K148 tok/s0.2 sPDF: Gemma 2 2B IT on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K163 tok/s0.3 sPDF: Gemma 2 2B IT on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K163 tok/s0.3 sPDF: Gemma 2 2B IT on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K163 tok/s0.3 sPDF: Gemma 2 2B IT on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K207 tok/s0.1 sPDF: Gemma 2 2B IT on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K65 tok/s0.3 sPDF: Gemma 2 2B IT on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K65 tok/s0.3 sPDF: Gemma 2 2B IT on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K91 tok/s0.1 sPDF: Gemma 2 2B IT on RTX Spark report
DGX Spark8-bit8K83 tok/s0.1 sPDF: Gemma 2 2B IT on DGX Spark report
Dual DGX Spark8-bit8K83 tok/s0.1 sPDF: Gemma 2 2B IT on Dual DGX Spark report
DGX Station GB3008-bit8K825 tok/s0.1 sPDF: Gemma 2 2B IT on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K401 tok/s0.1 sPDF: Gemma 2 2B IT on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K401 tok/s0.1 sPDF: Gemma 2 2B IT on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K401 tok/s0.1 sPDF: Gemma 2 2B IT on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K437 tok/s0.1 sPDF: Gemma 2 2B IT on A100 80 GB SXM report
H100 SXM8-bit8K589 tok/s0.1 sPDF: Gemma 2 2B IT on H100 SXM report
4× H200 NVL8-bit8K704 tok/s0.1 sPDF: Gemma 2 2B IT on 4× H200 NVL report
8× H200 NVL8-bit8K704 tok/s0.1 sPDF: Gemma 2 2B IT on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf bartowski/gemma-2-2b-it-GGUF
vLLM
vllm serve google/gemma-2-2b-it

Questions

How much VRAM does Gemma 2 2B IT need?

Gemma 2 2B IT needs 1.5 GB for its weights at 4-bit, 3.3 GB with an 8K context and 3.3 GB with 8K. At 8-bit, 4.4 GB.

What is the minimum VRAM for Gemma 2 2B IT?

About 3.0 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is Gemma 2 2B IT?

2.6B parameters: 4.9 GB of weights at 16-bit, 1.5 GB at 4-bit.

Can Gemma 2 2B IT run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 261 tokens a second.

More models