NVIDIA RTX 5090 D v2 (China Special Edition)#
Product Overview#
NVIDIA RTX 5090 D v2 is a consumer-grade flagship GPU customized for the Chinese mainland market by NVIDIA, officially launched on August 12, 2025, with official guide price ¥16,499. As the successor to the RTX 5090 D, the D v2 further reduces memory capacity to comply with US export control requirements, while retaining the full 21,760 CUDA cores and 575W TDP.
"D" Meaning: D = Downgraded, specifically designed for the Chinese mainland market. By reducing memory capacity and some AI performance metrics, it meets the US Department of Commerce export control regulations (10 CFR § 742.6) restrictions on "Total Processing Performance" (TPP) and "Performance Density".
Core Specifications#
| Item | RTX 5090 D v2 | RTX 5090 (Standard) | Difference |
|---|
| Architecture | Blackwell (GB202) | Blackwell (GB202) | Same |
| Process | TSMC 4N | TSMC 4N | Same |
| CUDA Cores | 21,760 | 21,760 | Same |
| Tensor Cores | 680 (5th-gen) | 680 (5th-gen) | Same |
| RT Cores | 170 (4th-gen) | 170 (4th-gen) | Same |
| Base Clock | 2.01 GHz | 2.01 GHz | Same |
| Boost Clock | 2.41 GHz | 2.41 GHz | Same |
| Memory | 24 GB GDDR7 | 32 GB GDDR7 | -25% |
| Memory Bus | 384-bit | 512-bit | -25% |
| Memory Bandwidth | 1,344 GB/s | 1,792 GB/s | -25% |
| AI Compute | 2,375 AI TOPS | 3,352 TOPS (FP4 sparse) | -29% |
| Ray Tracing | 318 TFLOPS | 318 TFLOPS | Same |
| FP32 Compute | ~105 TFLOPS | 104.8 TFLOPS | Basically same |
| TDP | 575 W | 575 W | Same |
| Power Connector | 1× 16-pin (12V-2×6) | 1× 16-pin (12V-2×6) | Same |
| Recommended PSU | 1,000W+ | 1,000W+ | Same |
| Display Outputs | DP 2.1b + HDMI 2.1b | DP 2.1b + HDMI 2.1b | Same |
| NVENC/NVDEC | 3× NVENC (9th-gen) + 2× NVDEC (6th-gen) | Same | Same |
| Launch Price | ¥16,499 | $1,999 | — |
| Launch Date | 2025-08-12 | 2025-01-30 | — |
Compliance with Chinese Mainland Restrictions#
| Metric | Control Threshold | RTX 5090 D v2 Actual | Compliance Status |
|---|
| TPP (Total Processing Performance) | ≤ 4,800 (restricted regions) | ~3,500 (estimated) | ✅ Compliant |
| Performance Density | ≤ 5.92 (restricted regions) | ~4.8 (estimated) | ✅ Compliant |
| Memory Capacity | ≤ 24 GB (restricted regions) | 24 GB | ✅ Compliant |
Note: NVIDIA has not publicly disclosed D v2's exact TPP data. The above are estimated values based on CUDA core count and clock speeds.
Comparison with RTX 5090 D (Previous-Gen China Special Edition)#
| Metric | RTX 5090 D (2024) | RTX 5090 D v2 (2025) | Change |
|---|
| Memory | 32 GB GDDR7 (512-bit) | 24 GB GDDR7 (384-bit) | -25% |
| Memory Bandwidth | 1,792 GB/s | 1,344 GB/s | -25% |
| Empty memory pads | None | 4 pads | Downgraded |
| Gaming Performance | ~98% of RTX 5090 | ~97% of RTX 5090 | -1% |
| AI Performance | ~90% of RTX 5090 | ~75-85% of RTX 5090 | -10-15% |
| Price | ¥16,499 | ¥16,499 | Unchanged |
AI Inference Scenarios#
- 70B LLM: After FP4 quantization ~40GB, 24GB memory cannot fully load, requires layer-wise loading or more aggressive quantization (INT4/INT2)
- 13B LLM: After FP4 quantization ~8GB, 24GB memory can run smoothly, supports inference + fine-tuning simultaneously
- Image generation (SDXL): 24GB memory can fully load SDXL + LoRA, basically no impact
Gaming Scenarios#
- 4K gaming: 24GB vs 32GB memory has minimal impact on frame rates (vast majority of games <16GB memory usage)
- 8K gaming: Some texture-intensive games may be affected, but RTX 5090 D v2's CUDA core count is not reduced, frame rate loss only 1-2%
Purchase Recommendations#
- Suitable for: Users in Chinese mainland who need Blackwell architecture + DLSS 4 + FP4 inference support, and primarily run ≤13B parameter LLMs
- Not suitable for: Local AI developers who need full 32GB memory to run 70B+ LLMs (recommend cloud compute solution)
- Alternative: RTX 4090 (24GB GDDR6X, better price-performance) or wait for non-Chinese mainland RTX 5090
Comparison with Competitors#
| Metric | RTX 5090 D v2 | RTX 4090 | AMD RX 9070 XT |
|---|
| Architecture | Blackwell | Ada Lovelace | RDNA 4 |
| Memory | 24GB GDDR7 | 24GB GDDR6X | 16GB GDDR6 |
| Memory Bandwidth | 1,344 GB/s | 1,008 GB/s | 640 GB/s |
| AI TOPS | 2,375 | 1,321 | ~800 (estimated) |
| TDP | 575W | 450W | 355W |
| Price | ¥16,499 | ¥12,999 | ¥5,999 |
Key Technical Features#
Blackwell Architecture New Features (Same as Standard Version)#
- 5th-gen Tensor Core: Native FP4 support, AI inference performance doubled
- 4th-gen RT Core: Path tracing performance improved 2×
- DLSS 4 Multi Frame Generation: AI generates up to 3 interpolated frames, frame rate improved up to 4×
- GDDR7 memory: 28 Gbps data rate, 384-bit bus width
China Special Edition Unique Features#
- Four empty memory pads: 4 memory chip positions reserved on PCB but not populated, physically limiting memory expansion possibility
- BIOS limitations: Some AI inference optimization features may be disabled or downclocked (NVIDIA not publicly disclosed)
Launch Date & Channels#
- Official Launch: August 12, 2025
- Sales Channels: JD.com, Tmall NVIDIA flagship store, Colorful, Gigabyte, and other AIB partners
- Recommended Retail Price: ¥16,499 (same as RTX 5090 D)
- Supply Status: Normal supply from August 2025
External Links#