Skip to content

GPT Neox 20B VRAM requirements

How much VRAM GPT Neox 20B needs
4-bit weights: 11.7 GBWith 2K context: 14.5 GB8-bit, 2K context: 23.3 GB
Parameters
20.7B
Longest context
2K
Released
Apr 2022
License
Apache 2.0
Strongest at
Not rated

How much VRAM GPT Neox 20B needs

Memory GPT Neox 20B needs at each precision and context length
PrecisionWeightsWith 2K
16-bit38.6 GB41.4 GB
8-bit20.5 GB23.3 GB
6-bit15.8 GB18.6 GB
5-bit13.7 GB16.5 GB
4-bit11.7 GB14.5 GB
3-bit9.4 GB12.2 GB
2-bit7.2 GB10.0 GB
1.6-bit4.0 GB6.8 GB

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

GPUs and Macs that run GPT Neox 20B

The best way GPT Neox 20B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: GPT Neox 20B on RTX 4060 report
RTX 5060Too bigPDF: GPT Neox 20B on RTX 5060 report
RTX 3070Too bigPDF: GPT Neox 20B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: GPT Neox 20B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: GPT Neox 20B on RTX 3060 12 GB report
Arc B580Too bigPDF: GPT Neox 20B on Arc B580 report
RTX 4070Too bigPDF: GPT Neox 20B on RTX 4070 report
RTX 4070 SuperToo bigPDF: GPT Neox 20B on RTX 4070 Super report
RTX 5070Too bigPDF: GPT Neox 20B on RTX 5070 report
RTX 4060 Ti 16 GB4-bit2K16 tok/s2.0 sPDF: GPT Neox 20B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB4-bit2K25 tok/s1.9 sPDF: GPT Neox 20B on RTX 5060 Ti 16 GB report
Arc A770 16 GB4-bit2K21 tok/s2.2 sPDF: GPT Neox 20B on Arc A770 16 GB report
Radeon RX 9070 XT4-bit2K29 tok/s1.2 sPDF: GPT Neox 20B on Radeon RX 9070 XT report
RTX 4070 Ti Super4-bit2K36 tok/s1.0 sPDF: GPT Neox 20B on RTX 4070 Ti Super report
RTX 4080 Super4-bit2K39 tok/s0.9 sPDF: GPT Neox 20B on RTX 4080 Super report
RTX 5070 Ti4-bit2K47 tok/s1.0 sPDF: GPT Neox 20B on RTX 5070 Ti report
RTX 50804-bit2K51 tok/s0.8 sPDF: GPT Neox 20B on RTX 5080 report
Radeon RX 7900 XTX8-bit2K25 tok/s2.0 sPDF: GPT Neox 20B on Radeon RX 7900 XTX report
RTX 30908-bit2K29 tok/s1.3 sPDF: GPT Neox 20B on RTX 3090 report
RTX 40908-bit2K31 tok/s0.5 sPDF: GPT Neox 20B on RTX 4090 report
Radeon AI PRO R97008-bit2K17 tok/s1.3 sPDF: GPT Neox 20B on Radeon AI PRO R9700 report
RTX 50908-bit2K54 tok/s0.4 sPDF: GPT Neox 20B on RTX 5090 report
2× RTX 30908-bit2K29 tok/s0.7 sPDF: GPT Neox 20B on 2× RTX 3090 report
2× RTX 40908-bit2K31 tok/s0.3 sPDF: GPT Neox 20B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit2K41 tok/s0.7 sPDF: GPT Neox 20B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit2K30 tok/s0.5 sPDF: GPT Neox 20B on RTX 6000 Ada report
L40S8-bit2K27 tok/s0.8 sPDF: GPT Neox 20B on L40S report
2× RTX 50908-bit2K54 tok/s0.2 sPDF: GPT Neox 20B on 2× RTX 5090 report
4× RTX 30908-bit2K29 tok/s0.4 sPDF: GPT Neox 20B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit2K54 tok/s0.3 sPDF: GPT Neox 20B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: GPT Neox 20B on Mac mini M4 16 GB report
Mac mini M4 24 GB4-bit2K7.7 tok/s28 sPDF: GPT Neox 20B on Mac mini M4 24 GB report
Mac mini M4 32 GB6-bit2K5.8 tok/s28 sPDF: GPT Neox 20B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB4-bit2K17 tok/s14 sPDF: GPT Neox 20B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit2K10.0 tok/s14 sPDF: GPT Neox 20B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit2K10.0 tok/s14 sPDF: GPT Neox 20B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit2K18 tok/s7.1 sPDF: GPT Neox 20B on M4 Max 128 GB report
M5 Max 128 GB8-bit2K20 tok/s2.0 sPDF: GPT Neox 20B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit2K23 tok/s4.0 sPDF: GPT Neox 20B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit2K23 tok/s4.0 sPDF: GPT Neox 20B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit2K23 tok/s4.0 sPDF: GPT Neox 20B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit2K32 tok/s1.2 sPDF: GPT Neox 20B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit2K7.0 tok/s4.1 sPDF: GPT Neox 20B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit2K7.0 tok/s4.1 sPDF: GPT Neox 20B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit2K9.6 tok/s1.4 sPDF: GPT Neox 20B on RTX Spark report
DGX Spark8-bit2K8.7 tok/s1.4 sPDF: GPT Neox 20B on DGX Spark report
Dual DGX Spark8-bit2K8.7 tok/s0.8 sPDF: GPT Neox 20B on Dual DGX Spark report
DGX Station GB3008-bit2K176 tok/s0.1 sPDF: GPT Neox 20B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit2K54 tok/s0.2 sPDF: GPT Neox 20B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit2K54 tok/s0.1 sPDF: GPT Neox 20B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit2K54 tok/s0.1 sPDF: GPT Neox 20B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit2K61 tok/s0.9 sPDF: GPT Neox 20B on A100 80 GB SXM report
H100 SXM8-bit2K95 tok/s0.3 sPDF: GPT Neox 20B on H100 SXM report
4× H200 NVL8-bit2K129 tok/s0.1 sPDF: GPT Neox 20B on 4× H200 NVL report
8× H200 NVL8-bit2K129 tok/s0.1 sPDF: GPT Neox 20B 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/gpt-neox-20b-i1-GGUF
vLLM
vllm serve EleutherAI/gpt-neox-20b

Questions

How much VRAM does GPT Neox 20B need?

GPT Neox 20B needs 11.7 GB for its weights at 4-bit, 14.5 GB with an 2K context and 14.5 GB with 2K. At 8-bit, 23.3 GB.

What is the minimum VRAM for GPT Neox 20B?

About 12.2 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 Ti 16 GB.

How big is GPT Neox 20B?

20.7B parameters: 38.6 GB of weights at 16-bit, 11.7 GB at 4-bit.

Can GPT Neox 20B run on an RTX 4090?

Yes. At 8-bit it holds 2K of context and writes about 31 tokens a second.

More models