Skip to content

HarmBench Llama 2 13B Cls VRAM requirements

How much VRAM HarmBench Llama 2 13B Cls needs
4-bit weights: 7.3 GBWith 2K context: 9.6 GB8-bit, 2K context: 15.1 GB
Parameters
13B
Longest context
2K
Released
Feb 2024
License
MIT
Strongest at
Not rated

How much VRAM HarmBench Llama 2 13B Cls needs

Memory HarmBench Llama 2 13B Cls needs at each precision and context length
PrecisionWeightsWith 2K
16-bit24.2 GB26.5 GB
8-bit12.9 GB15.1 GB
6-bit9.9 GB12.2 GB
5-bit8.6 GB10.9 GB
4-bit7.3 GB9.6 GB
3-bit5.9 GB8.2 GB
2-bit4.5 GB6.8 GB
1.6-bit2.5 GB4.7 GB

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

GPUs and Macs that run HarmBench Llama 2 13B Cls

The best way HarmBench Llama 2 13B Cls runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: HarmBench Llama 2 13B Cls on RTX 4060 report
RTX 5060Too bigPDF: HarmBench Llama 2 13B Cls on RTX 5060 report
RTX 3070Too bigPDF: HarmBench Llama 2 13B Cls on RTX 3070 report
RTX 3080 10 GB4-bit2K63 tok/s0.9 sPDF: HarmBench Llama 2 13B Cls on RTX 3080 10 GB report
RTX 3060 12 GB5-bit2K27 tok/s2.2 sPDF: HarmBench Llama 2 13B Cls on RTX 3060 12 GB report
Arc B5805-bit2K24 tok/s1.5 sPDF: HarmBench Llama 2 13B Cls on Arc B580 report
RTX 40705-bit2K37 tok/s1.0 sPDF: HarmBench Llama 2 13B Cls on RTX 4070 report
RTX 4070 Super5-bit2K37 tok/s0.8 sPDF: HarmBench Llama 2 13B Cls on RTX 4070 Super report
RTX 50705-bit2K48 tok/s0.9 sPDF: HarmBench Llama 2 13B Cls on RTX 5070 report
RTX 4060 Ti 16 GB8-bit2K15 tok/s1.3 sPDF: HarmBench Llama 2 13B Cls on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit2K22 tok/s1.2 sPDF: HarmBench Llama 2 13B Cls on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit2K20 tok/s1.4 sPDF: HarmBench Llama 2 13B Cls on Arc A770 16 GB report
Radeon RX 9070 XT8-bit2K27 tok/s0.8 sPDF: HarmBench Llama 2 13B Cls on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit2K33 tok/s0.6 sPDF: HarmBench Llama 2 13B Cls on RTX 4070 Ti Super report
RTX 4080 Super8-bit2K36 tok/s0.5 sPDF: HarmBench Llama 2 13B Cls on RTX 4080 Super report
RTX 5070 Ti8-bit2K44 tok/s0.6 sPDF: HarmBench Llama 2 13B Cls on RTX 5070 Ti report
RTX 50808-bit2K46 tok/s0.5 sPDF: HarmBench Llama 2 13B Cls on RTX 5080 report
Radeon RX 7900 XTX8-bit2K39 tok/s1.2 sPDF: HarmBench Llama 2 13B Cls on Radeon RX 7900 XTX report
RTX 30908-bit2K45 tok/s0.8 sPDF: HarmBench Llama 2 13B Cls on RTX 3090 report
RTX 40908-bit2K49 tok/s0.3 sPDF: HarmBench Llama 2 13B Cls on RTX 4090 report
Radeon AI PRO R97008-bit2K27 tok/s0.8 sPDF: HarmBench Llama 2 13B Cls on Radeon AI PRO R9700 report
RTX 50908-bit2K83 tok/s0.3 sPDF: HarmBench Llama 2 13B Cls on RTX 5090 report
2× RTX 30908-bit2K45 tok/s0.5 sPDF: HarmBench Llama 2 13B Cls on 2× RTX 3090 report
2× RTX 40908-bit2K49 tok/s0.2 sPDF: HarmBench Llama 2 13B Cls on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit2K64 tok/s0.4 sPDF: HarmBench Llama 2 13B Cls on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit2K46 tok/s0.3 sPDF: HarmBench Llama 2 13B Cls on RTX 6000 Ada report
L40S8-bit2K42 tok/s0.5 sPDF: HarmBench Llama 2 13B Cls on L40S report
2× RTX 50908-bit2K83 tok/s0.2 sPDF: HarmBench Llama 2 13B Cls on 2× RTX 5090 report
4× RTX 30908-bit2K45 tok/s0.2 sPDF: HarmBench Llama 2 13B Cls on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit2K83 tok/s0.2 sPDF: HarmBench Llama 2 13B Cls on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB5-bit2K10 tok/s18 sPDF: HarmBench Llama 2 13B Cls on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit2K7.0 tok/s18 sPDF: HarmBench Llama 2 13B Cls on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit2K7.0 tok/s18 sPDF: HarmBench Llama 2 13B Cls on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit2K16 tok/s9.0 sPDF: HarmBench Llama 2 13B Cls on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit2K16 tok/s9.0 sPDF: HarmBench Llama 2 13B Cls on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit2K16 tok/s9.0 sPDF: HarmBench Llama 2 13B Cls on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit2K27 tok/s4.5 sPDF: HarmBench Llama 2 13B Cls on M4 Max 128 GB report
M5 Max 128 GB8-bit2K30 tok/s1.3 sPDF: HarmBench Llama 2 13B Cls on M5 Max 128 GB report
M3 Ultra 96 GB8-bit2K35 tok/s2.5 sPDF: HarmBench Llama 2 13B Cls on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit2K35 tok/s2.5 sPDF: HarmBench Llama 2 13B Cls on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit2K35 tok/s2.5 sPDF: HarmBench Llama 2 13B Cls on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit2K48 tok/s0.8 sPDF: HarmBench Llama 2 13B Cls on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit2K11 tok/s2.6 sPDF: HarmBench Llama 2 13B Cls on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit2K11 tok/s2.6 sPDF: HarmBench Llama 2 13B Cls on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit2K15 tok/s0.9 sPDF: HarmBench Llama 2 13B Cls on RTX Spark report
DGX Spark8-bit2K14 tok/s0.9 sPDF: HarmBench Llama 2 13B Cls on DGX Spark report
Dual DGX Spark8-bit2K14 tok/s0.5 sPDF: HarmBench Llama 2 13B Cls on Dual DGX Spark report
DGX Station GB3008-bit2K253 tok/s0.1 sPDF: HarmBench Llama 2 13B Cls on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit2K83 tok/s0.1 sPDF: HarmBench Llama 2 13B Cls on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit2K83 tok/s0.1 sPDF: HarmBench Llama 2 13B Cls on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit2K83 tok/s0.1 sPDF: HarmBench Llama 2 13B Cls on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit2K93 tok/s0.6 sPDF: HarmBench Llama 2 13B Cls on A100 80 GB SXM report
H100 SXM8-bit2K142 tok/s0.2 sPDF: HarmBench Llama 2 13B Cls on H100 SXM report
4× H200 NVL8-bit2K190 tok/s0.1 sPDF: HarmBench Llama 2 13B Cls on 4× H200 NVL report
8× H200 NVL8-bit2K190 tok/s0.1 sPDF: HarmBench Llama 2 13B Cls 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/HarmBench-Llama-2-13b-cls-i1-GGUF
vLLM
vllm serve cais/HarmBench-Llama-2-13b-cls

Questions

How much VRAM does HarmBench Llama 2 13B Cls need?

HarmBench Llama 2 13B Cls needs 7.3 GB for its weights at 4-bit, 9.6 GB with an 2K context and 9.6 GB with 2K. At 8-bit, 15.1 GB.

What is the minimum VRAM for HarmBench Llama 2 13B Cls?

About 8.2 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 3080 10 GB.

How big is HarmBench Llama 2 13B Cls?

13B parameters: 24.2 GB of weights at 16-bit, 7.3 GB at 4-bit.

Can HarmBench Llama 2 13B Cls run on an RTX 4090?

Yes. At 8-bit it holds 2K of context and writes about 49 tokens a second.

More models