Wink Pings

Moore Threads 'Lushan' GPU Revealed: 64GB VRAM, Twice That of the RTX 5090?

Moore Threads' new GPU, 'Lushan,' is rumored to come with 64GB of VRAM, which could set a new record for consumer GPU capacity. If inference performance can keep up, a single card may be able to run a 70B-class quantized model in full. Actual performance and pricing will have to wait until the December developer conference.

Moore Threads is teasing 'Lushan.' It is a new consumer graphics card rumored to have VRAM capacity of up to 64GB.

Moore Threads 'Lushan' GPU teaser screenshot

If true, this would set a new record for consumer GPU VRAM capacity. For comparison:

  • NVIDIA RTX 5090: 32GB GDDR7
  • Moore Threads 'Lushan': 64GB (to be confirmed)

The report comes from David Hendrickson of the X account TeksEdge. Moore Threads has previously claimed that its next-generation architecture aims for 4x VRAM capacity, 64x AI compute, 15x AAA gaming performance, and 50x ray tracing performance. These are the vendor's generational targets, not measured comparisons against the RTX 5090.

The previous-generation MTT S80 had only 16GB of VRAM. Now, a product manager at Moore Threads says that 'Lushan' may offer the largest VRAM capacity among consumer graphics cards.

Moore Threads has its own GPU compute platform, MUSA. llama.cpp is a commonly used local large-model inference tool, and there is already a MUSA backend for existing Moore Threads GPUs. This means that the software path for running local large models on Chinese-made graphics cards already exists.

If 'Lushan' really ships with 64GB and inference compatibility is well handled, then in theory a single consumer graphics card could fully run a 70B-class (roughly 70 billion parameter) quantized model in VRAM.

However, the key specs remain unknown:

  • Actual VRAM capacity
  • Memory bandwidth
  • Low-precision (FP4) inference performance
  • Generation speed in llama.cpp (tokens/sec)
  • Price and launch date

The Moore Threads Developer Conference on December 29-30 is expected to discuss this card.

Chinese-made GPUs are catching up in hardware specs, but the software ecosystem and actual inference efficiency are what determine whether it can be used. If the price and performance are competitive, would you consider using a 64GB Chinese-made graphics card to run local AI?

Source: David Hendrickson of TeksEdge

发布时间: 2026-11-22 14:35