Skip to content

Phi 1 5 VRAM requirements

How much VRAM Phi 1 5 needs
4-bit weights: 820 MBWith 2K context: 1.8 GB8-bit, 2K context: 2.4 GB
Parameters
1.4B
Longest context
2K
Released
Sep 2023
License
MIT
Strongest at
Not rated

How much VRAM Phi 1 5 needs

Memory Phi 1 5 needs at each precision and context length
PrecisionWeightsWith 2K
16-bit2.6 GB3.6 GB
8-bit1.4 GB2.4 GB
6-bit1.1 GB2.1 GB
5-bit962 MB1.9 GB
4-bit820 MB1.8 GB
3-bit661 MB1.6 GB
2-bit507 MB1.5 GB
1.6-bit279 MB1.3 GB

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

GPUs and Macs that run Phi 1 5

The best way Phi 1 5 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit2K124 tok/s0.2 sPDF: Phi 1 5 on RTX 4060 report
RTX 50608-bit2K193 tok/s0.1 sPDF: Phi 1 5 on RTX 5060 report
RTX 30708-bit2K193 tok/s0.1 sPDF: Phi 1 5 on RTX 3070 report
RTX 3080 10 GB8-bit2K299 tok/s0.1 sPDF: Phi 1 5 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit2K160 tok/s0.2 sPDF: Phi 1 5 on RTX 3060 12 GB report
Arc B5808-bit2K132 tok/s0.2 sPDF: Phi 1 5 on Arc B580 report
RTX 40708-bit2K214 tok/s0.1 sPDF: Phi 1 5 on RTX 4070 report
RTX 4070 Super8-bit2K214 tok/s0.1 sPDF: Phi 1 5 on RTX 4070 Super report
RTX 50708-bit2K271 tok/s0.1 sPDF: Phi 1 5 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit2K131 tok/s0.1 sPDF: Phi 1 5 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit2K193 tok/s0.1 sPDF: Phi 1 5 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit2K156 tok/s0.1 sPDF: Phi 1 5 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit2K207 tok/s0.1 sPDF: Phi 1 5 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit2K271 tok/s0.1 sPDF: Phi 1 5 on RTX 4070 Ti Super report
RTX 4080 Super8-bit2K292 tok/s0.1 sPDF: Phi 1 5 on RTX 4080 Super report
RTX 5070 Ti8-bit2K340 tok/s0.1 sPDF: Phi 1 5 on RTX 5070 Ti report
RTX 50808-bit2K358 tok/s0.1 sPDF: Phi 1 5 on RTX 5080 report
Radeon RX 7900 XTX8-bit2K276 tok/s0.1 sPDF: Phi 1 5 on Radeon RX 7900 XTX report
RTX 30908-bit2K351 tok/s0.1 sPDF: Phi 1 5 on RTX 3090 report
RTX 40908-bit2K371 tok/s0.1 sPDF: Phi 1 5 on RTX 4090 report
Radeon AI PRO R97008-bit2K207 tok/s0.1 sPDF: Phi 1 5 on Radeon AI PRO R9700 report
RTX 50908-bit2K546 tok/s0.1 sPDF: Phi 1 5 on RTX 5090 report
2× RTX 30908-bit2K351 tok/s0.1 sPDF: Phi 1 5 on 2× RTX 3090 report
2× RTX 40908-bit2K371 tok/s0.1 sPDF: Phi 1 5 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit2K454 tok/s0.1 sPDF: Phi 1 5 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit2K358 tok/s0.1 sPDF: Phi 1 5 on RTX 6000 Ada report
L40S8-bit2K330 tok/s0.1 sPDF: Phi 1 5 on L40S report
2× RTX 50908-bit2K546 tok/s0.1 sPDF: Phi 1 5 on 2× RTX 5090 report
4× RTX 30908-bit2K351 tok/s0.1 sPDF: Phi 1 5 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit2K546 tok/s0.1 sPDF: Phi 1 5 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit2K62 tok/s1.8 sPDF: Phi 1 5 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit2K62 tok/s1.8 sPDF: Phi 1 5 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit2K62 tok/s1.8 sPDF: Phi 1 5 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit2K123 tok/s0.9 sPDF: Phi 1 5 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit2K123 tok/s0.9 sPDF: Phi 1 5 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit2K123 tok/s0.9 sPDF: Phi 1 5 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit2K188 tok/s0.4 sPDF: Phi 1 5 on M4 Max 128 GB report
M5 Max 128 GB8-bit2K202 tok/s0.1 sPDF: Phi 1 5 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit2K220 tok/s0.3 sPDF: Phi 1 5 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit2K220 tok/s0.3 sPDF: Phi 1 5 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit2K220 tok/s0.3 sPDF: Phi 1 5 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit2K270 tok/s0.1 sPDF: Phi 1 5 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit2K97 tok/s0.3 sPDF: Phi 1 5 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit2K97 tok/s0.3 sPDF: Phi 1 5 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit2K136 tok/s0.1 sPDF: Phi 1 5 on RTX Spark report
DGX Spark8-bit2K125 tok/s0.1 sPDF: Phi 1 5 on DGX Spark report
Dual DGX Spark8-bit2K125 tok/s0.1 sPDF: Phi 1 5 on Dual DGX Spark report
DGX Station GB3008-bit2K999 tok/s0.1 sPDF: Phi 1 5 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit2K546 tok/s0.1 sPDF: Phi 1 5 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit2K546 tok/s0.1 sPDF: Phi 1 5 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit2K546 tok/s0.1 sPDF: Phi 1 5 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit2K589 tok/s0.1 sPDF: Phi 1 5 on A100 80 GB SXM report
H100 SXM8-bit2K760 tok/s0.1 sPDF: Phi 1 5 on H100 SXM report
4× H200 NVL8-bit2K881 tok/s0.1 sPDF: Phi 1 5 on 4× H200 NVL report
8× H200 NVL8-bit2K881 tok/s0.1 sPDF: Phi 1 5 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 mradermacher/phi-1_5-GGUF
vLLM
vllm serve microsoft/phi-1_5

Questions

How much VRAM does Phi 1 5 need?

Phi 1 5 needs 820 MB for its weights at 4-bit, 1.8 GB with an 2K context and 1.8 GB with 2K. At 8-bit, 2.4 GB.

What is the minimum VRAM for Phi 1 5?

About 1.6 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 1 5?

1.4B parameters: 2.6 GB of weights at 16-bit, 820 MB at 4-bit.

Can Phi 1 5 run on an RTX 4090?

Yes. At 8-bit it holds 2K of context and writes about 371 tokens a second.

More models