Skip to content

Phi 2 VRAM requirements

How much VRAM Phi 2 needs
4-bit weights: 1.6 GBWith 2K context: 2.8 GB8-bit, 2K context: 4.0 GB
Parameters
2.8B
Longest context
2K
Released
Dec 2023
License
MIT
Strongest at
Not rated

How much VRAM Phi 2 needs

Memory Phi 2 needs at each precision and context length
PrecisionWeightsWith 2K
16-bit5.2 GB6.4 GB
8-bit2.8 GB4.0 GB
6-bit2.1 GB3.4 GB
5-bit1.8 GB3.1 GB
4-bit1.6 GB2.8 GB
3-bit1.3 GB2.5 GB
2-bit1.0 GB2.2 GB
1.6-bit547 MB1.8 GB

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

GPUs and Macs that run Phi 2

The best way Phi 2 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit2K64 tok/s0.4 sPDF: Phi 2 on RTX 4060 report
RTX 50608-bit2K101 tok/s0.3 sPDF: Phi 2 on RTX 5060 report
RTX 30708-bit2K101 tok/s0.3 sPDF: Phi 2 on RTX 3070 report
RTX 3080 10 GB8-bit2K160 tok/s0.2 sPDF: Phi 2 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit2K82 tok/s0.5 sPDF: Phi 2 on RTX 3060 12 GB report
Arc B5808-bit2K70 tok/s0.3 sPDF: Phi 2 on Arc B580 report
RTX 40708-bit2K112 tok/s0.2 sPDF: Phi 2 on RTX 4070 report
RTX 4070 Super8-bit2K112 tok/s0.2 sPDF: Phi 2 on RTX 4070 Super report
RTX 50708-bit2K144 tok/s0.2 sPDF: Phi 2 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit2K67 tok/s0.3 sPDF: Phi 2 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit2K101 tok/s0.2 sPDF: Phi 2 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit2K84 tok/s0.3 sPDF: Phi 2 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit2K112 tok/s0.2 sPDF: Phi 2 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit2K144 tok/s0.1 sPDF: Phi 2 on RTX 4070 Ti Super report
RTX 4080 Super8-bit2K156 tok/s0.1 sPDF: Phi 2 on RTX 4080 Super report
RTX 5070 Ti8-bit2K184 tok/s0.1 sPDF: Phi 2 on RTX 5070 Ti report
RTX 50808-bit2K194 tok/s0.1 sPDF: Phi 2 on RTX 5080 report
Radeon RX 7900 XTX8-bit2K154 tok/s0.3 sPDF: Phi 2 on Radeon RX 7900 XTX report
RTX 30908-bit2K190 tok/s0.2 sPDF: Phi 2 on RTX 3090 report
RTX 40908-bit2K202 tok/s0.1 sPDF: Phi 2 on RTX 4090 report
Radeon AI PRO R97008-bit2K112 tok/s0.2 sPDF: Phi 2 on Radeon AI PRO R9700 report
RTX 50908-bit2K312 tok/s0.1 sPDF: Phi 2 on RTX 5090 report
2× RTX 30908-bit2K190 tok/s0.1 sPDF: Phi 2 on 2× RTX 3090 report
2× RTX 40908-bit2K202 tok/s0.1 sPDF: Phi 2 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit2K253 tok/s0.1 sPDF: Phi 2 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit2K194 tok/s0.1 sPDF: Phi 2 on RTX 6000 Ada report
L40S8-bit2K178 tok/s0.1 sPDF: Phi 2 on L40S report
2× RTX 50908-bit2K312 tok/s0.1 sPDF: Phi 2 on 2× RTX 5090 report
4× RTX 30908-bit2K190 tok/s0.1 sPDF: Phi 2 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit2K312 tok/s0.1 sPDF: Phi 2 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit2K32 tok/s3.7 sPDF: Phi 2 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit2K32 tok/s3.7 sPDF: Phi 2 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit2K32 tok/s3.7 sPDF: Phi 2 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit2K66 tok/s1.8 sPDF: Phi 2 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit2K66 tok/s1.8 sPDF: Phi 2 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit2K66 tok/s1.8 sPDF: Phi 2 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit2K106 tok/s0.9 sPDF: Phi 2 on M4 Max 128 GB report
M5 Max 128 GB8-bit2K115 tok/s0.3 sPDF: Phi 2 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit2K127 tok/s0.5 sPDF: Phi 2 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit2K127 tok/s0.5 sPDF: Phi 2 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit2K127 tok/s0.5 sPDF: Phi 2 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit2K162 tok/s0.2 sPDF: Phi 2 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit2K50 tok/s0.5 sPDF: Phi 2 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit2K50 tok/s0.5 sPDF: Phi 2 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit2K70 tok/s0.2 sPDF: Phi 2 on RTX Spark report
DGX Spark8-bit2K64 tok/s0.2 sPDF: Phi 2 on DGX Spark report
Dual DGX Spark8-bit2K64 tok/s0.1 sPDF: Phi 2 on Dual DGX Spark report
DGX Station GB3008-bit2K655 tok/s0.1 sPDF: Phi 2 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit2K312 tok/s0.1 sPDF: Phi 2 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit2K312 tok/s0.1 sPDF: Phi 2 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit2K312 tok/s0.1 sPDF: Phi 2 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit2K341 tok/s0.1 sPDF: Phi 2 on A100 80 GB SXM report
H100 SXM8-bit2K463 tok/s0.1 sPDF: Phi 2 on H100 SXM report
4× H200 NVL8-bit2K556 tok/s0.1 sPDF: Phi 2 on 4× H200 NVL report
8× H200 NVL8-bit2K556 tok/s0.1 sPDF: Phi 2 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 TheBloke/phi-2-GGUF
vLLM
vllm serve microsoft/phi-2

Questions

How much VRAM does Phi 2 need?

Phi 2 needs 1.6 GB for its weights at 4-bit, 2.8 GB with an 2K context and 2.8 GB with 2K. At 8-bit, 4.0 GB.

What is the minimum VRAM for Phi 2?

About 2.5 GB at 3-bit with an 2K 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 Phi 2?

2.8B parameters: 5.2 GB of weights at 16-bit, 1.6 GB at 4-bit.

Can Phi 2 run on an RTX 4090?

Yes. At 8-bit it holds 2K of context and writes about 202 tokens a second.

More models