For comparison, Nvidia's 96GB card uses GDDR7 for 1.8TB/sec bandwidth. This Huawei card is more in the league of an M4 Max bandwidth-wise.
Yes, similar to the slower M4 Max option. The Mac Studio with either M4 Max or M3 Ultra uses 8533 MT/s LPDDR5X.
Memory bandwidth for the M4 Max is 410 GB/s with the 32-core GPU or 546 GB/s with the 40-core GPU. It’s 819 GB/s in the M3 Ultra config.
But the price !
I wonder if Huawei are making any money off this, or are they trying to get a foot in the door first.
Which is?
I took a gander:
“Available for around $1,250 to $1,500 before shipping and potential import fees … if your primary bottleneck is VRAM for loading massive models … it is not an RTX 4090 replacement for high-speed inference on smaller models, but it is a VRAM barrier-breaker for running the giants of the open-source world.”