Before anyone clears space in a Windows 11 tower for a 64GB unicorn, though, it pays to separate what has actually been announced from what is still creative arithmetic. Neither Moore Threads nor its product manager has published an official VRAM number, memory bus width, price, or firm retail launch date for Lushan. What we have right now is an informal social post, a moody mountain-themed teaser poster, and a looming year-end developer conference where the company has a lot to prove.
Separating the Official Teaser From the 64GB Math
The current buzz stems from two separate breadcrumbs dropped in mid-November 2026. First, on November 18, an MTT S80 product manager posting under the handle AIR_FORCE—reported on Bilibili by VideoCardz and on Weibo in coverage cited by TechPowerUp and 3dsinifi—asserted that Lushan would carry the highest video memory capacity ever offered on a consumer graphics card. Shortly afterward, Moore Threads released its first promotional poster for the 2026 MUSA Developer Conference (MDC 2026), showing a graphics card draped under a black cloth between two peaks—one rendered as a 3D wireframe mesh and the other as a neural network—while naming Lushan in the post title.
Notice what is missing from both the poster and the product manager's post: an actual gigabyte count. The widely circulated 64GB number comes from connecting two dots rather than reading a spec sheet:
- The baseline: Moore Threads' official specifications confirm that its previous consumer flagship, the MTT S80, carries 16GB of GDDR6 memory.
- The earlier promise: When Moore Threads first introduced its next-generation Huagang architecture at MDC 2025 in December 2025, it touted a 4x increase in GPU memory capacity over its prior generation.
- The arithmetic: Multiply 16GB by four, and you land on 64GB—double the 32GB of GDDR7 found on Nvidia's GeForce RTX 5090.
In other words, 64GB is a reasonable deduction if the 4x architecture claim applies directly to the flagship consumer Lushan SKU, and if Moore Threads' product manager is measuring against the 32GB RTX 5090. Until Moore Threads puts a formal spec table on stage, however, "more than 32GB" could technically mean 48GB, 64GB, or a multi-SKU lineup with only a halo prosumer model hitting the ceiling.
How Lushan Stacks Up Against Real Consumer and Workstation GPUs
Comparing VRAM numbers across the modern GPU landscape is tricky because "consumer," "prosumer," and "workstation" lines have blurred—especially when retail scarcity pushes standard gaming cards into luxury pricing. Tom's Hardware noted that while the 32GB GeForce RTX 5090 carries a $1,999 MSRP, real-world retail listings had climbed to $6,599 at Newegg as of November 15, reflecting intense demand and tight memory supply.
To keep the comparison honest, single-chip desktop cards need to be separated from enterprise workstation hardware, dual-GPU boards, and gray-market modifications:
| Graphics Card | Market Class | Memory Configuration | Pricing / Market Context |
|---|---|---|---|
| Moore Threads Lushan | Consumer / Desktop (upcoming) | Undisclosed (4x S80 claim implies up to 64GB) | Not announced; teased for MDC 2026 |
| Nvidia GeForce RTX 5090 | Consumer Flagship | 32GB GDDR7 | $1,999 MSRP; listed from $6,599 at Newegg (Nov. 15, per Tom's Hardware) |
| Moore Threads MTT S80 | Consumer (2022) | 16GB GDDR6 (256-bit, 448 GB/s) | ~2,999 yuan ($410) bundled at launch; 1,499 yuan (~$225) on JD.com (Nov. 21, per Tom's Hardware) |
| Lisuan 7G106 / 7G105 | Consumer / Professional | 12GB (7G106) / 24GB (7G105) GDDR6 | Domestic Chinese 6nm TrueGPU rivals |
| Maxsun Intel Arc Pro B60 Dual | Workstation (Dual-GPU PCB) | 48GB GDDR6 total (2x 24GB, bifurcated x8/x8) | Built for local AI inference; requires PCIe bifurcation |
| Nvidia RTX Pro 6000 Blackwell | Workstation | 96GB GDDR7 with ECC (512-bit, 600W) | Enterprise/professional tier, export-restricted |
| Modded RTX 5090 (China) | Gray-Market Modification | 96GB (as listed) | $3,888 listing cited by Tom's Hardware; unsupported mod |
That distinction matters. Nvidia's RTX Pro 6000 Blackwell Workstation Edition already packs 96GB of GDDR7 ECC memory, and Maxsun's Intel Arc Pro B60 Dual puts 48GB of GDDR6 onto a single PCIe slot—but the latter achieves it by bolting two separate GPUs onto one board via bifurcated PCIe 5.0 x8 links, meaning software sees two 24GB devices rather than one unified 48GB memory pool. Meanwhile, gray-market 96GB modded RTX 5090 boards circulating in China are bespoke hardware hacks, not standard Nvidia products. If Moore Threads actually ships a single-chip consumer card above 32GB—leaving domestic rival Lisuan's 12GB 7G106 and 24GB 7G105 behind—it would hold a genuine capacity crown in the standard desktop category.
Bold Architecture Claims Meet the MTT S80's Rocky History
Capacity is only half the story, and Moore Threads knows its reputation in PC gaming has some bruises to heal. Built on the new Huagang graphics architecture—alongside a data-center AI sibling called Huashan—Lushan was first previewed at MDC 2025 with eye-popping generational multipliers over the MTT S80:
- 15x AAA gaming performance
- 50x ray-tracing performance
- 64x AI compute throughput
- 16x geometry processing (plus 4x texture fill rate and 8x atomic memory access)
To back up those figures, Moore Threads says Lushan introduces a dedicated hardware ray-tracing engine, its UniTE rendering architecture, native DirectX 12 Ultimate support, and AI generative rendering—which Chinese tech outlet MyDrivers compared to Nvidia's DLSS and AMD's FSR for AI upscaling and multi-frame generation.
Still, veteran PC builders have seen impressive slide decks before, and the MTT S80's history is a textbook reminder that silicon specs do not write Windows drivers. When the S80 debuted in November 2022—initially sold only in a 2,999-yuan bundle with an Asus motherboard—it looked ambitious on paper: 4,096 MUSA cores at 1.8 GHz, 14.7 TFLOPS of FP32 compute, 16GB of GDDR6, a 255W power envelope, and the distinction of being the first GPU with a PCIe Gen5 x16 interface.
Once independent reviewers tested the card in 2023, reality hit hard. Japanese outlet PC Watch found that DirectX 12 and Vulkan games routinely failed to run, leaving testers stuck with older DirectX 11 titles and legacy benchmarks like 3DMark 06, while early performance trailed even low-end cards like Nvidia's GeForce GT 1030. Moore Threads responded with a steady drumbeat of monthly driver updates and claimed in December 2025 that the S80 benchmarked 3.4x higher than it did at launch. As Tom's Hardware pointed out, Lushan's 15x gaming target looks slightly less mythical when you remember just how low the S80's starting floor was—yet it remains an untested company projection until third-party reviewers get hardware on a test bench.
What a 64GB Desktop GPU Actually Means for Windows Gamers and Local AI
Assuming Moore Threads delivers a 48GB or 64GB consumer card, who actually benefits? Paradoxically, pure PC gamers are probably not at the top of the list.
For Windows 11 gaming, VRAM capacity acts like a warehouse, not a forklift: if a game needs 12GB or 16GB of textures and frame buffers at 4K, jumping from 32GB to 64GB does not magically double your frame rate. Real-world playability on Lushan will rise or fall on shader efficiency, memory bandwidth (which depends heavily on whether Moore Threads uses GDDR6 or costly GDDR7), and whether its DirectX 12 Ultimate drivers and AI generative rendering actually work cleanly across modern titles without stuttering or visual glitches.
Where a 64GB desktop card would immediately turn heads is local AI and creative workstations. Running large language models locally, experimenting with heavy image and video generation workflows, or loading massive 3D and architectural visualization scenes hits a hard wall the moment a model spills out of VRAM into system RAM. With RTX 5090 street prices inflated well above MSRP and workstation cards like the 96GB RTX Pro 6000 costing enterprise budgets, an affordable high-VRAM card looks tempting on paper.
Even there, Windows users face two major hurdles:
- The software stack: Modern local AI tools on Windows—from PyTorch and ComfyUI to LM Studio and Ollama—are overwhelmingly tuned for Nvidia's CUDA ecosystem or standard DirectML/Windows ML paths. Moore Threads relies on its own MUSA compute platform and PES Control Center. Without broad framework support outside China, raw VRAM capacity can sit stranded behind compatibility workarounds.
- Availability and sanctions: Moore Threads was placed on the U.S. Department of Commerce's Bureau of Industry and Security Entity List in October 2023. Even if Lushan ships before the end of 2026 as previously indicated, North American and European buyers should not expect it to appear casually on standard retail shelves alongside GeForce and Radeon cards.
What to Watch at MDC 2026
We will not have to wait long to see whether Lushan's memory capacity lives up to the hype. Moore Threads is holding MDC 2026 on December 29–30, 2026, at the Zhongguancun International Innovation Center in Beijing, where the company has promised more than ten new products, technologies, and solutions. Founder and CEO Zhang Jianzhong—a former Nvidia China general manager who founded Moore Threads in 2020—will deliver the opening keynote at 9:30 a.m. Beijing time on December 29 (8:30 p.m. ET on December 28).
A developer conference reveal is not the same thing as retail stock on a shelf, and a teaser poster hidden behind mountain wireframe art leaves plenty of room for surprises. When Zhang takes the stage, keep an eye out for three concrete details: the exact VRAM size and memory type, independent DirectX 12 Ultimate gaming demos rather than synthetic multipliers, and whether Moore Threads is positioning Lushan as a genuine gamer's card—or a local-AI workstation bargain wearing gaming clothes.
References
- Moore Threads' next 'Lushan' gaming GPU could pack a record 64GB of VRAM Tom's Hardware · 2026-11-22T12:30:00+00:00
- Lushan GPU Could Set VRAM Record, Executive Claims · 3dsınıfı 3dsinifi.com
- videocardz.com videocardz.com