Skip to content

Falcon 40B VRAM requirements

How much VRAM Falcon 40B needs
4-bit weights: 23.6 GBWith 4K context: 24.9 GB8-bit, 4K context: 42.7 GB
Parameters
41.8B
Longest context
4K
Released
May 2023
License
Apache 2.0
Strongest at
Not rated

How much VRAM Falcon 40B needs

Memory Falcon 40B needs at each precision and context length
PrecisionWeightsWith 4K
16-bit77.9 GB79.2 GB
8-bit41.4 GB42.7 GB
6-bit31.9 GB33.2 GB
5-bit27.7 GB29.0 GB
4-bit23.6 GB24.9 GB
3-bit19.0 GB20.3 GB
2-bit14.6 GB15.9 GB
1.6-bit8.0 GB9.3 GB

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

GPUs and Macs that run Falcon 40B

The best way Falcon 40B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Falcon 40B on RTX 4060 report
RTX 5060Too bigPDF: Falcon 40B on RTX 5060 report
RTX 3070Too bigPDF: Falcon 40B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Falcon 40B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Falcon 40B on RTX 3060 12 GB report
Arc B580Too bigPDF: Falcon 40B on Arc B580 report
RTX 4070Too bigPDF: Falcon 40B on RTX 4070 report
RTX 4070 SuperToo bigPDF: Falcon 40B on RTX 4070 Super report
RTX 5070Too bigPDF: Falcon 40B on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Falcon 40B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Falcon 40B on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Falcon 40B on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Falcon 40B on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Falcon 40B on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Falcon 40B on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Falcon 40B on RTX 5070 Ti report
RTX 5080Too bigPDF: Falcon 40B on RTX 5080 report
Radeon RX 7900 XTX3-bit4K26 tok/s4.0 sPDF: Falcon 40B on Radeon RX 7900 XTX report
RTX 30903-bit4K31 tok/s2.5 sPDF: Falcon 40B on RTX 3090 report
RTX 40903-bit4K33 tok/s1.1 sPDF: Falcon 40B on RTX 4090 report
Radeon AI PRO R97005-bit4K13 tok/s2.6 sPDF: Falcon 40B on Radeon AI PRO R9700 report
RTX 50905-bit4K40 tok/s0.9 sPDF: Falcon 40B on RTX 5090 report
2× RTX 30908-bit4K15 tok/s1.4 sPDF: Falcon 40B on 2× RTX 3090 report
2× RTX 40908-bit4K16 tok/s0.6 sPDF: Falcon 40B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit4K21 tok/s1.3 sPDF: Falcon 40B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit4K15 tok/s0.9 sPDF: Falcon 40B on RTX 6000 Ada report
L40S8-bit4K13 tok/s1.6 sPDF: Falcon 40B on L40S report
2× RTX 50908-bit4K27 tok/s0.5 sPDF: Falcon 40B on 2× RTX 5090 report
4× RTX 30908-bit4K15 tok/s0.8 sPDF: Falcon 40B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit4K27 tok/s0.7 sPDF: Falcon 40B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Falcon 40B on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Falcon 40B on Mac mini M4 24 GB report
Mac mini M4 32 GB3-bit4K4.8 tok/s57 sPDF: Falcon 40B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Falcon 40B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB6-bit4K6.4 tok/s29 sPDF: Falcon 40B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit4K5.0 tok/s29 sPDF: Falcon 40B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit4K8.9 tok/s14 sPDF: Falcon 40B on M4 Max 128 GB report
M5 Max 128 GB8-bit4K10.0 tok/s4.1 sPDF: Falcon 40B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit4K11 tok/s8.0 sPDF: Falcon 40B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit4K11 tok/s8.0 sPDF: Falcon 40B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit4K11 tok/s8.0 sPDF: Falcon 40B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit4K16 tok/s2.5 sPDF: Falcon 40B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit4K3.5 tok/s8.2 sPDF: Falcon 40B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit4K3.5 tok/s8.2 sPDF: Falcon 40B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit4K4.7 tok/s2.9 sPDF: Falcon 40B on RTX Spark report
DGX Spark8-bit4K4.3 tok/s2.9 sPDF: Falcon 40B on DGX Spark report
Dual DGX Spark8-bit4K4.3 tok/s1.6 sPDF: Falcon 40B on Dual DGX Spark report
DGX Station GB3008-bit4K94 tok/s0.2 sPDF: Falcon 40B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit4K27 tok/s0.4 sPDF: Falcon 40B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit4K27 tok/s0.2 sPDF: Falcon 40B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit4K27 tok/s0.1 sPDF: Falcon 40B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit4K31 tok/s1.8 sPDF: Falcon 40B on A100 80 GB SXM report
H100 SXM8-bit4K49 tok/s0.6 sPDF: Falcon 40B on H100 SXM report
4× H200 NVL8-bit4K67 tok/s0.2 sPDF: Falcon 40B on 4× H200 NVL report
8× H200 NVL8-bit4K67 tok/s0.1 sPDF: Falcon 40B 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/falcon-40b-i1-GGUF
vLLM
vllm serve tiiuae/falcon-40b

Questions

How much VRAM does Falcon 40B need?

Falcon 40B needs 23.6 GB for its weights at 4-bit, 24.9 GB with an 4K context and 24.9 GB with 4K. At 8-bit, 42.7 GB.

What is the minimum VRAM for Falcon 40B?

About 20.3 GB at 3-bit with an 4K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the Radeon RX 7900 XTX.

How big is Falcon 40B?

41.8B parameters: 77.9 GB of weights at 16-bit, 23.6 GB at 4-bit.

Can Falcon 40B run on an RTX 4090?

Yes. At 3-bit it holds 4K of context and writes about 33 tokens a second.

More models