Asus ROG Flow Z13 GZ302EA Convertible Review - AMD's Strix Halo GPU is neck-and-neck with the RTX 4070 Laptop

[just at least 512-bit bro]

128GB RAM in an (expensive) 13" 2-in-1 is not for everyone (it even rhymes, so it must be true).

• Strix Halo memory bus width: 256-bit, up to 128GB RAM
• Medusa Halo (Strix Halo successor) (rumor): 384-bit, up to 192GB RAM

If Medusa Halo had up to 256 GB RAM, it could fit a decent real DeepSeek-R1 / V3 quant (not the fake "Distill" ones):
huggingface.co/search/full-text?q=+DeepSeek-R1-GGUF&type=model
huggingface.co/unsloth/DeepSeek-V3-0324-GGUF

Code Select Expand

1.78bit IQ1_S 173GB Ok
1.93bit IQ1_M 183GB Fair
2.42bit IQ2_XXS 203GB Recommended
2.71bit Q2_K_XL 231GB Recommended
3.5bit Q3_K_XL 320GB Great
4.5bit Q4_K_XL 406GB Best

[Benchmark eg Llama-3 70B]

AIDA64 / Memory Read
Asus ROG Flow Z13 GZ302EA-RU073W
AMD Ryzen AI Max+ 395, Radeon RX 8060S   
121931 MB/s

Lets add what has been said in notebookchat.com/index.php?topic=226695 ("HP ZBook Ultra G1a 14 workstation review: MacBook Pro alternative for gamers")
Strix Halo is a 256-bit chip, so its theoretical bandwidth is:

Code Select Expand

memory bandwidth [GB/s] = 8000 * 64 (bit per channel) * 4 (channels) (aka quad-channel) (= the 256-bit memory bus width) / 8 (bit to Byte) / 1000 (MB to GB)
                        = 256 GB/sThe practical benchmark value is often 70-80% of the theoretical value, so:

Code Select Expand

256 GB/s * 0.75 = 192 GB/s

Code Select Expand

LLM tokens per second = 192 GB/s / 39.6 GB (Llama-3.3-70B-Instruct-Q4_K_M.gguf)
                      = 4.85
Obviously, the measured 121931 MB/s don't align with what should be measured as something like 192 GB/s. Notebookcheck, could you run a e.g. Llama-3.3-70B-Instruct-Q4_K_M.gguf LLM and see how many tokens per second you get?