Uploaded September 2026 | Updated November 2026, 4 days ago
I ran Qwen 3.8 27B on a MacBook Pro M5 Max and on an RTX 5090. The 5090 was more than 4 times faster. Here is why.
▸ RTX 5090: 130 tokens per second. MacBook Pro M5 Max: 30 tokens per second.
▸ Every new token reads the whole model from memory once
▸ So LLM speed depends on memory bandwidth: how many GB per second the chip can read
▸ Think of filling a bucket through a pipe. A wider pipe fills it faster.
▸ RTX 5090 reads almost 1,800 GB per second. M5 Max reads 460 GB per second.
▸ That is almost 4 times more bandwidth, and that is where the gap comes from
▸ Comparing a Mac and a GPU for local LLMs? Check memory bandwidth, not just the chip speed.
Were you checking only the chip speed, or the memory as well? Tell me in the comments 👇
Full benchmark (exact numbers, settings, prompts): kgptalkie.com/tutorials/llm-benchmarking/qwen-3-8-27b-on-macbook-m5-max
Full video: youtube.com/watch?v=c9VPDM1TYcs
Follow for more local LLM benchmarks.
#M5Max #RTX5090 #LocalLLM #Shorts
I ran Qwen 3.8 27B on a MacBook Pro M5 Max and on an RTX 5090. The 5090 was more than 4 times faster. Here is why.
▸ RTX 5090: 130 tokens per second. MacBook Pro M5 Max: 30 tokens per second.
▸ Every new token reads the whole model from memory once
▸ So LLM speed depends on memory bandwidth: how many GB per second the chip can read
▸ Think of filling a bucket through a pipe. A wider pipe fills it faster.
▸ RTX 5090 reads almost 1,800 GB per second. M5 Max reads 460 GB per second.
▸ That is almost 4 times more bandwidth, and that is where the gap comes from
▸ Comparing a Mac and a GPU for local LLMs? Check memory bandwidth, not just the chip speed.
Were you checking only the chip speed, or the memory as well? Tell me in the comments 👇
Full benchmark (exact numbers, settings, prompts): kgptalkie.com/tutorials/llm-benchmarking/qwen-3-8-27b-on-macbook-m5-max
Full video: youtube.com/watch?v=c9VPDM1TYcs
Follow for more local LLM benchmarks.
#M5Max #RTX5090 #LocalLLM #Shorts










