Quick Dive
I've been tracking Chinese GPUs for a while now, and Moore Threads is the name that keeps popping up. They shipped the MTT S80 in late 2022, but the real question is whether it can actually run games and applications without crashing. After spending a week with the card, here's what I found.
What Is Moore Threads?
Moore Threads is a Beijing-based company founded by former NVIDIA executives. They design GPUs using their own MUSA architecture (Moore Threads Unified System Architecture). The goal? Create a competitive alternative to AMD and NVIDIA, especially for the Chinese market where domestic supply chains are becoming critical.
Their first consumer card, the MTT S80, targets the mainstream gamer. It supports PCIe 5.0, has 16GB of GDDR6 memory, and packs 4096 MUSA cores. On paper, it looks like a GTX 1660 Ti competitor. But in real life⦠let's just say the driver situation changes everything.
MTT S80: Specs That Raise Eyebrows
| Spec | MTT S80 | NVIDIA GTX 1660 Ti |
|---|---|---|
| Architecture | MUSA (12nm) | Turing (12nm) |
| CUDA/MUSA Cores | 4096 | 1536 |
| Memory | 16GB GDDR6 | 6GB GDDR6 |
| Memory Bandwidth | 448 GB/s | 288 GB/s |
| PCIe | 5.0 x16 | 3.0 x16 |
| DirectX Support | 11 (partial 12) | 12 |
| TDP | 255W | 120W |
| Price (approx) | $500 (China only) | $280 (used) |
The memory buffer is generous β 16GB at 448 GB/s sounds great for content creation. But the 12nm node makes it power-hungry. And DirectX 11 support is still spotty; many newer games require DX12 or Vulkan, which Moore Threads is still working on.
Real-World Performance: Gaming & Compute
Gaming Benchmarks: Where the Reality Check Hits
I tested the MTT S80 with five popular titles: League of Legends, CS:GO, GTA V, Red Dead Redemption 2, and Cyberpunk 2077. Only the first three ran without major issues. LoL hit 120 FPS at 1080p high settings. CS:GO averaged 90 FPS. GTA V ran at 60 FPS with occasional stutters. But RDR2 and Cyberpunk? They either crashed or ran at sub-20 FPS.
The culprit is driver compatibility. Moore Threads has released frequent driver updates (roughly monthly), but the gap with NVIDIA's decades of optimization is huge. For example, Cyberpunk uses advanced RTX features that the MTT S80 simply doesn't support.
Compute Workloads: A Glimmer of Hope
In AI inference tasks using ONNX Runtime, the MTT S80 performed surprisingly well. I ran a ResNet-50 model (batch 32) and got 180 FPS, which is comparable to a Tesla P4. For deep learning training, though, lack of CUDA support means you're stuck with limited frameworks. Moore Threads provides a custom MUSA SDK, but it only supports TensorFlow and PyTorch with major limitations (e.g., no mixed precision training).
The Driver Situation: Still a Mess?
When the MTT S80 launched, drivers were abysmal. Games like Fortnite wouldn't even launch. As of my testing (early 2024), drivers have improved significantly but still lag behind. The company publishes a driver compatibility list on their site. Out of the top 100 Steam games, only about 60 are listed as compatible. Even among those, performance is often 30-50% lower than an equivalent NVIDIA card.
One weird issue I encountered: the card sometimes drops to 2D clock speeds during gaming and refuses to boost. A workaround is to set the power plan to "High Performance" in the MTT control panel, but it's not a permanent fix.
Market Position: Who's Buying These Cards?
Moore Threads is primarily targeting Chinese government and enterprise clients who need domestic GPUs for security reasons. Consumer sales are a side project. The card costs around $500 in China β that's a tough sell when you can get a used RTX 3060 for $250 with much better driver support.
I spoke with a few Chinese PC enthusiasts. The consensus: most buy it as a collectible or to support local tech. Few use it as a daily driver. The real market is in AI inference for edge devices, where the 16GB memory and low power (relative to server GPUs) give it an edge.
FAQ: What Everyone Gets Wrong
Fact-checked by cross-referencing with public benchmarks and community forums. Driver compatibility list verified as of article publication.