Skip to content

Gpt2 XL VRAM requirements

How much VRAM Gpt2 XL needs
4-bit weights: 930 MBWith 2K context: 2.1 GB8-bit, 2K context: 2.8 GB
Parameters
1.6B
Longest context
2K
Released
Mar 2022
License
MIT
Strongest at
Not rated

How much VRAM Gpt2 XL needs

Memory Gpt2 XL needs at each precision and context length
PrecisionWeightsWith 2K
16-bit3.0 GB4.2 GB
8-bit1.6 GB2.8 GB
6-bit1.2 GB2.4 GB
5-bit1.1 GB2.2 GB
4-bit930 MB2.1 GB
3-bit749 MB1.9 GB
2-bit575 MB1.7 GB
1.6-bit316 MB1.5 GB

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

GPUs and Macs that run Gpt2 XL

The best way Gpt2 XL runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit2K100 tok/s0.2 sPDF: Gpt2 XL on RTX 4060 report
RTX 50608-bit2K151 tok/s0.2 sPDF: Gpt2 XL on RTX 5060 report
RTX 30708-bit2K151 tok/s0.2 sPDF: Gpt2 XL on RTX 3070 report
RTX 3080 10 GB8-bit2K223 tok/s0.1 sPDF: Gpt2 XL on RTX 3080 10 GB report
RTX 3060 12 GB8-bit2K127 tok/s0.3 sPDF: Gpt2 XL on RTX 3060 12 GB report
Arc B5808-bit2K100 tok/s0.2 sPDF: Gpt2 XL on Arc B580 report
RTX 40708-bit2K166 tok/s0.1 sPDF: Gpt2 XL on RTX 4070 report
RTX 4070 Super8-bit2K166 tok/s0.1 sPDF: Gpt2 XL on RTX 4070 Super report
RTX 50708-bit2K204 tok/s0.1 sPDF: Gpt2 XL on RTX 5070 report
RTX 4060 Ti 16 GB8-bit2K105 tok/s0.2 sPDF: Gpt2 XL on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit2K151 tok/s0.1 sPDF: Gpt2 XL on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit2K115 tok/s0.2 sPDF: Gpt2 XL on Arc A770 16 GB report
Radeon RX 9070 XT8-bit2K151 tok/s0.1 sPDF: Gpt2 XL on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit2K204 tok/s0.1 sPDF: Gpt2 XL on RTX 4070 Ti Super report
RTX 4080 Super8-bit2K218 tok/s0.1 sPDF: Gpt2 XL on RTX 4080 Super report
RTX 5070 Ti8-bit2K248 tok/s0.1 sPDF: Gpt2 XL on RTX 5070 Ti report
RTX 50808-bit2K259 tok/s0.1 sPDF: Gpt2 XL on RTX 5080 report
Radeon RX 7900 XTX8-bit2K192 tok/s0.2 sPDF: Gpt2 XL on Radeon RX 7900 XTX report
RTX 30908-bit2K255 tok/s0.1 sPDF: Gpt2 XL on RTX 3090 report
RTX 40908-bit2K267 tok/s0.1 sPDF: Gpt2 XL on RTX 4090 report
Radeon AI PRO R97008-bit2K151 tok/s0.1 sPDF: Gpt2 XL on Radeon AI PRO R9700 report
RTX 50908-bit2K366 tok/s0.1 sPDF: Gpt2 XL on RTX 5090 report
2× RTX 30908-bit2K255 tok/s0.1 sPDF: Gpt2 XL on 2× RTX 3090 report
2× RTX 40908-bit2K267 tok/s0.1 sPDF: Gpt2 XL on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit2K316 tok/s0.1 sPDF: Gpt2 XL on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit2K259 tok/s0.1 sPDF: Gpt2 XL on RTX 6000 Ada report
L40S8-bit2K243 tok/s0.1 sPDF: Gpt2 XL on L40S report
2× RTX 50908-bit2K366 tok/s0.1 sPDF: Gpt2 XL on 2× RTX 5090 report
4× RTX 30908-bit2K255 tok/s0.1 sPDF: Gpt2 XL on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit2K366 tok/s0.1 sPDF: Gpt2 XL on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit2K49 tok/s2.2 sPDF: Gpt2 XL on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit2K49 tok/s2.2 sPDF: Gpt2 XL on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit2K49 tok/s2.2 sPDF: Gpt2 XL on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit2K90 tok/s1.1 sPDF: Gpt2 XL on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit2K90 tok/s1.1 sPDF: Gpt2 XL on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit2K90 tok/s1.1 sPDF: Gpt2 XL on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit2K128 tok/s0.5 sPDF: Gpt2 XL on M4 Max 128 GB report
M5 Max 128 GB8-bit2K136 tok/s0.2 sPDF: Gpt2 XL on M5 Max 128 GB report
M3 Ultra 96 GB8-bit2K145 tok/s0.3 sPDF: Gpt2 XL on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit2K145 tok/s0.3 sPDF: Gpt2 XL on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit2K145 tok/s0.3 sPDF: Gpt2 XL on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit2K169 tok/s0.1 sPDF: Gpt2 XL on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit2K77 tok/s0.3 sPDF: Gpt2 XL on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit2K77 tok/s0.3 sPDF: Gpt2 XL on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit2K109 tok/s0.1 sPDF: Gpt2 XL on RTX Spark report
DGX Spark8-bit2K101 tok/s0.1 sPDF: Gpt2 XL on DGX Spark report
Dual DGX Spark8-bit2K101 tok/s0.1 sPDF: Gpt2 XL on Dual DGX Spark report
DGX Station GB3008-bit2K566 tok/s0.1 sPDF: Gpt2 XL on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit2K366 tok/s0.1 sPDF: Gpt2 XL on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit2K366 tok/s0.1 sPDF: Gpt2 XL on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit2K366 tok/s0.1 sPDF: Gpt2 XL on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit2K388 tok/s0.1 sPDF: Gpt2 XL on A100 80 GB SXM report
H100 SXM8-bit2K469 tok/s0.1 sPDF: Gpt2 XL on H100 SXM report
4× H200 NVL8-bit2K520 tok/s0.1 sPDF: Gpt2 XL on 4× H200 NVL report
8× H200 NVL8-bit2K520 tok/s0.1 sPDF: Gpt2 XL 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/gpt2-xl-GGUF
vLLM
vllm serve openai-community/gpt2-xl

Questions

How much VRAM does Gpt2 XL need?

Gpt2 XL needs 930 MB for its weights at 4-bit, 2.1 GB with an 2K context and 2.1 GB with 2K. At 8-bit, 2.8 GB.

What is the minimum VRAM for Gpt2 XL?

About 1.9 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 Gpt2 XL?

1.6B parameters: 3.0 GB of weights at 16-bit, 930 MB at 4-bit.

Can Gpt2 XL run on an RTX 4090?

Yes. At 8-bit it holds 2K of context and writes about 267 tokens a second.

More models