Moore Threads, the company we are familiar with that produces gaming GPUs, is on its way to a mega-IPO in Shanghai, and some call the Chinese firm the "NVIDIA" of the region.
Moore Threads: An Old Story Starting With Gaming GPUs That Match AMD/NVIDIA GPUs From Several Years Ago
With the IPO hype, many people have recently become aware of Moore Threads, but for those who have been following us consistently, it's clear what the Chinese GPU firm has accomplished over the years. We'll dive into that later, but for now, let's discuss Moore Threads' upcoming IPO in Shanghai, which is reportedly the largest in the domestic GPU space. It is reported that the IPO has drawn over 4,000 subscriptions from retail investors, and more importantly, the firm plans to raise $1.1 billion through its IPO at Shanghai STAR.
Of course, the magnitude of interest in the firm's IPO is so massive that it has been making headlines in the Western media, with CNBC calling Moore Threads 'NVIDIA of China', and claiming that it's Beijing's push towards being self-reliant in the race for AI chips. But, it's important that we dive into technicals, instead of sticking to headlines out there, since as far as what has been disclosed publicly, Moore Threads doesn't have a 'capable' AI chip solution to showcase to the world for now, and while there are reports of the firm developing a 7nm-based platform, we are yet to see its showcase.
What we do know is that Moore Threads has a reasonably extensive lineup of gaming GPUs to offer, with one of the prominent offerings being the MTT S80 and MTT S90, which are the company's consumer-focused options, similar to NVIDIA's RTX/GTX. Similarly, the Chinese firm also offers workstation offerings, such as the MTT S4000 and the MTT X300; the models mentioned above are those that the company has publicly disclosed. Moore Threads may have an AI chip platform under development, but for now, we are only aware of consumer/professional offerings.
| Specification | MTT S50 | MTT S80 | MTT S90 |
|---|---|---|---|
| GPU Architecture / Cores | 2,048 MUSA cores | 4,096 MUSA cores | Not disclosed |
| FP32 Compute | ~5.2 TFLOPS | ~14.4 TFLOPS | Not disclosed (rumored > RTX 4060) |
| VRAM & Memory | 8 GB GDDR6, 256-bit | 16 GB GDDR6, 256-bit | Not disclosed |
| Memory Bandwidth | ~448 GB/s | 448 GB/s | Not disclosed |
| PCIe Interface | PCIe 3.0/4.0 ×16 (varies) | PCIe 5.0 ×16 | Not disclosed |
| Power Consumption (TDP) | 85 W | 255 W | Not disclosed |
| Display Outputs | HDMI + DP x2, up to 8K | Multi-display, up to 8K | Not disclosed |
| Video Codec Support | AV1, H.264, H.265 | AV1, H.264, H.265, VP9 | Not disclosed |
| Form Factor | Single-slot | Dual-slot | Likely dual-slot |
| Target Market | Entry-level consumer GPU | Mid-range consumer/professional | High-end consumer GPU (rumored tier) |
When it comes to the performance of these GPUs, various benchmarks have surfaced over the years, examining individual models. However, it would be fair to say that the firm has failed to deliver impressive offerings, despite having 'decent' on-paper specifications. For example, the company's MTT S80 GPU features 4096 in-house MUSA cores, along with support for PCIe Gen 5 x 16, offering up to 128 GB/s bandwidth. The company was the first to adopt the PCIe Gen 5 standard; however, when it comes to performance, the MTT S80 lags behind NVIDIA's decade-old GTX 1050 Ti.
The MTT S80 does have a successor called the MTT S90, which features immensely better gaming performance, and rivals NVIDIA's RTX 4060 in terms of the benchmarks we have seen surfacing up, and while both GPUs have similar on-paper specifications, one of the main ways by which Moore Threads has brought in improvements is through the advancement of GPU drivers, which were in a terrible position during the start of the company's venture in the consumer GPU space.
Talking about the workstation lineup, you are looking at the MTT S4000, which is the 'closest' to what Moore Threads has to offer when it comes to catering to the DC markets. The S4000 offers 48 GB of GDDR6 memory, clocked at 16 Gbps, to provide 768 GB/s of bandwidth. The GPU features the latest MTLink 1.0 interface technology, enabling customers to run multiple cards simultaneously. Think of it as an NVLINK solution for Moore Threads GPUs.
According to the company itself, some of the compute figures shared showcase 25 TFLOPs of FP32, 50 TFLOPs of TF32, 100 TFLOPs of FP16/BF16, and 200 TOPS on INT8 performance. With the S4000, Moore Threads also offers a KUAE cluster, similar to NVIDIA's DGX system, by stacking eight MTT S4000 GPUs. It also comes with seamless expansion from a single machine to multiple cards and multiple AI systems. KUAE systems are claimed to scale up to 1,000 units of the MTT S4000, which the firm claims will allow running 'multi-billion' parameter models with ease.
| Specification | MTT S4000 | MTT X300 |
|---|---|---|
| Architecture | 3rd-Gen MUSA | 2nd-Gen MUSA |
| GPU Cores | Not disclosed | 4,096 MUSA cores |
| FP32 Performance | ~25 TFLOPS | 14.4 TFLOPS |
| Tensor / AI Performance | 50 TFLOPS TF32, 100 TFLOPS FP16/BF16, 200 TOPS INT8 | Not specified |
| VRAM Capacity | 48 GB GDDR6 | 16 GB GDDR6 |
| Memory Bandwidth | 768 GB/s | 448 GB/s |
| Memory Speed | 16 Gbps | Not specified |
| Memory Type | GDDR6 | GDDR6 |
| PCIe Interface | PCIe Gen5 ×16 | PCIe Gen5 ×16 |
| Multi-GPU Support | MTLink interconnect | Not specified |
| Display Outputs | 4× (supports 8K) | Up to 4× (supports 8K) |
| Video Capabilities | Multi-stream 96×1080p, 8K encode/decode | Professional video encoding/decoding |
| TDP | Not specified | ~255 W |
| Target Market | AI, HPC, Data-Center, LLM Inference | Workstation / Professional Graphics |
| Special Features | Multi-GPU scaling, high VRAM, AI acceleration | Cross-architecture support (x86, ARM, LoongArch) |
The above information has come to us based on public sources, and since developments around computation in China are often 'misreported' across several occasions, we have included the specifications only available by Moore Threads themselves. Now, this allows me to come to my conclusion, which is that the Chinese GPU manufacturer doesn't have an impressive software and hardware suite to offer, which is why calling it "NVIDIA of China" is a little far-fetched, at least when it comes to technical aspects.
Since the interest in pursuing domestic solutions in China is at its peak in recent times, companies like BirenTech, MetaX, Cambricon, and Moore Threads have seen massive retail interest, even if it doesn't justify the compute portfolio they offer. The Shanghai IPO will defintely aid Moore Threads in boosting R&D and manufacturing, but for now, the company is confined to the consumer GPU space, with few options for the DC/AI segment.
Follow Wccftech on Google to get more of our news coverage in your feeds.
