Skip to content

PowerMoE 3B VRAM requirements

How much VRAM PowerMoE 3B needs
4-bit weights: 1.9 GBWith 4K context: 2.7 GB8-bit, 4K context: 4.2 GB
Parameters
3.4B
Active per token
958M
Longest context
4K
Released
Aug 2024
License
Apache 2.0
Strongest at
Not rated

How much VRAM PowerMoE 3B needs

Memory PowerMoE 3B needs at each precision and context length
PrecisionWeightsWith 4K
16-bit6.3 GB7.1 GB
8-bit3.3 GB4.2 GB
6-bit2.6 GB3.4 GB
5-bit2.2 GB3.1 GB
4-bit1.9 GB2.7 GB
3-bit1.5 GB2.4 GB
2-bit1.2 GB2.0 GB
1.6-bit664 MB1.5 GB

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

GPUs and Macs that run PowerMoE 3B

The best way PowerMoE 3B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit4K138 tok/s0.5 sPDF: PowerMoE 3B on RTX 4060 report
RTX 50608-bit4K194 tok/s0.4 sPDF: PowerMoE 3B on RTX 5060 report
RTX 30708-bit4K194 tok/s0.3 sPDF: PowerMoE 3B on RTX 3070 report
RTX 3080 10 GB8-bit4K261 tok/s0.2 sPDF: PowerMoE 3B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit4K168 tok/s0.6 sPDF: PowerMoE 3B on RTX 3060 12 GB report
Arc B5808-bit4K130 tok/s0.4 sPDF: PowerMoE 3B on Arc B580 report
RTX 40708-bit4K208 tok/s0.2 sPDF: PowerMoE 3B on RTX 4070 report
RTX 4070 Super8-bit4K208 tok/s0.2 sPDF: PowerMoE 3B on RTX 4070 Super report
RTX 50708-bit4K245 tok/s0.2 sPDF: PowerMoE 3B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit4K143 tok/s0.3 sPDF: PowerMoE 3B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit4K194 tok/s0.3 sPDF: PowerMoE 3B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit4K146 tok/s0.3 sPDF: PowerMoE 3B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit4K181 tok/s0.2 sPDF: PowerMoE 3B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit4K245 tok/s0.2 sPDF: PowerMoE 3B on RTX 4070 Ti Super report
RTX 4080 Super8-bit4K257 tok/s0.1 sPDF: PowerMoE 3B on RTX 4080 Super report
RTX 5070 Ti8-bit4K282 tok/s0.2 sPDF: PowerMoE 3B on RTX 5070 Ti report
RTX 50808-bit4K291 tok/s0.1 sPDF: PowerMoE 3B on RTX 5080 report
Radeon RX 7900 XTX8-bit4K215 tok/s0.3 sPDF: PowerMoE 3B on Radeon RX 7900 XTX report
RTX 30908-bit4K288 tok/s0.2 sPDF: PowerMoE 3B on RTX 3090 report
RTX 40908-bit4K297 tok/s0.1 sPDF: PowerMoE 3B on RTX 4090 report
Radeon AI PRO R97008-bit4K181 tok/s0.2 sPDF: PowerMoE 3B on Radeon AI PRO R9700 report
RTX 50908-bit4K366 tok/s0.1 sPDF: PowerMoE 3B on RTX 5090 report
2× RTX 30908-bit4K288 tok/s0.1 sPDF: PowerMoE 3B on 2× RTX 3090 report
2× RTX 40908-bit4K297 tok/s0.1 sPDF: PowerMoE 3B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit4K333 tok/s0.1 sPDF: PowerMoE 3B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit4K291 tok/s0.1 sPDF: PowerMoE 3B on RTX 6000 Ada report
L40S8-bit4K277 tok/s0.1 sPDF: PowerMoE 3B on L40S report
2× RTX 50908-bit4K366 tok/s0.1 sPDF: PowerMoE 3B on 2× RTX 5090 report
4× RTX 30908-bit4K288 tok/s0.1 sPDF: PowerMoE 3B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit4K366 tok/s0.1 sPDF: PowerMoE 3B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit4K67 tok/s4.5 sPDF: PowerMoE 3B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit4K67 tok/s4.5 sPDF: PowerMoE 3B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit4K67 tok/s4.5 sPDF: PowerMoE 3B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit4K111 tok/s2.2 sPDF: PowerMoE 3B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit4K111 tok/s2.2 sPDF: PowerMoE 3B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit4K111 tok/s2.2 sPDF: PowerMoE 3B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit4K144 tok/s1.1 sPDF: PowerMoE 3B on M4 Max 128 GB report
M5 Max 128 GB8-bit4K149 tok/s0.3 sPDF: PowerMoE 3B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit4K156 tok/s0.6 sPDF: PowerMoE 3B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit4K156 tok/s0.6 sPDF: PowerMoE 3B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit4K156 tok/s0.6 sPDF: PowerMoE 3B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit4K173 tok/s0.2 sPDF: PowerMoE 3B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit4K105 tok/s0.6 sPDF: PowerMoE 3B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit4K105 tok/s0.6 sPDF: PowerMoE 3B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit4K148 tok/s0.2 sPDF: PowerMoE 3B on RTX Spark report
DGX Spark8-bit4K138 tok/s0.2 sPDF: PowerMoE 3B on DGX Spark report
Dual DGX Spark8-bit4K138 tok/s0.1 sPDF: PowerMoE 3B on Dual DGX Spark report
DGX Station GB3008-bit4K471 tok/s0.1 sPDF: PowerMoE 3B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit4K366 tok/s0.1 sPDF: PowerMoE 3B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit4K366 tok/s0.1 sPDF: PowerMoE 3B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit4K366 tok/s0.1 sPDF: PowerMoE 3B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit4K380 tok/s0.1 sPDF: PowerMoE 3B on A100 80 GB SXM report
H100 SXM8-bit4K425 tok/s0.1 sPDF: PowerMoE 3B on H100 SXM report
4× H200 NVL8-bit4K450 tok/s0.1 sPDF: PowerMoE 3B on 4× H200 NVL report
8× H200 NVL8-bit4K450 tok/s0.1 sPDF: PowerMoE 3B 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 tensorblock/PowerMoE-3b-GGUF
vLLM
vllm serve ibm-research/PowerMoE-3b

Questions

How much VRAM does PowerMoE 3B need?

PowerMoE 3B needs 1.9 GB for its weights at 4-bit, 2.7 GB with an 4K context and 2.7 GB with 4K. At 8-bit, 4.2 GB.

What is the minimum VRAM for PowerMoE 3B?

About 2.4 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 RTX 4060.

How big is PowerMoE 3B?

3.4B parameters (958M active per token): 6.3 GB of weights at 16-bit, 1.9 GB at 4-bit.

Can PowerMoE 3B run on an RTX 4090?

Yes. At 8-bit it holds 4K of context and writes about 297 tokens a second.

More models