Overview OVERVIEW
RTX 4090 D ("D" = Dragon, per the official release) is NVIDIA's Ada-compliant flagship for the Chinese mainland market
(January 2024 release, official figures): AD102 core, 14592 CUDA 、 24GB GDDR6X · 1008GB/s 、TGP
425W , official MSRP ¥12,999 . It is a compliance-trimmed version of the RTX 4090 (16384 CUDA/
450W) — the China-market Variant supply model shaped by export controls (per public reporting), sitting slightly below the 4090 in performance.
Its 24GB of large memory makes it the go-to consumer card for running local large models (estimated figures).
History HISTORY
- 2023-10 : US export controls on China tightened, and the 4090 was placed on the control list (per public reporting).
- 2024-01 : RTX 4090 D launched (official figures) — the compliant flagship filled the China supply gap, with an official MSRP of ¥12,999.
- 2024-2025 : The D variant became the mainstay consumer AI card in Chinese retail channels (market figures); the H20 data-center compliant line was supplied in parallel.
- 2025 : The 50-series D variants followed for China compliant supply (per public reporting) — the D production line became the standard for consumer supply to China.
Compliant Variant Comparison GENERATIONS
| Dimension | RTX 4090 (global version, official figures) | RTX 4090 D (this entry, official figures) |
| CUDA | 16384 | 14592 (roughly -11%) |
| Memory | 24GB GDDR6X · 1008GB/s | 24GB GDDR6X · 1008GB/s (same spec) |
| TGP | 450W | 425W |
| Positioning | Global consumer flagship | China mainland compliant supply (per public reporting) |
Note: both columns are from official pages (verified 2026-11); the compliance-trim background is per public reporting. Memory/bandwidth are the same spec — the performance gap comes mainly from the CUDA trim.
Technical Specifications SPECS
| Item | Specs (NVIDIA official) |
| Architecture | Ada Lovelace · AD102 |
| CUDA | 14592 |
| Memory | 24GB GDDR6X · 384-bit · 1008GB/s (official figures) |
| Power / Form Factor | TGP 425W · triple-slot class · 12VHPWR power |
| Interface | PCIe 4.0 x16 · DisplayPort 1.4a + HDMI 2.1 |
| Codec | AV1 dual NVENC (8th gen · official figures) |
| AI stack | DLSS 3 · full CUDA ecosystem (official figures) |
| Launch | 2024-01 · ¥12,999 official MSRP (release figures) |
Model Fit MODEL FIT
| Use cases | Fit estimate for 24GB GDDR6X (estimated: weights × precision + KV/framework overhead) |
| 7B-32B-class inference | FP16 full precision (7B-14B) / INT4 Quantization (32B) fit — the mainstream consumer capacity tier (estimated figures) |
| 70B-class | INT4 Quantization sits at the limit (~35GB weights excluding KV) — dual cards or step down to 32B (estimated figures) |
| Image generation/video | Plenty of headroom for high-resolution SD/DiT; dual AV1 encoders are friendly to creative pipelines (ecosystem figures) |
| LoRA fine-tuning | 7B-14B-class LoRA runs on a single card (estimated figures) |
Limitations and Boundaries LIMITS
- Performance tier : CUDA trimmed by roughly 11% — a clear performance gap versus the 4090 (official comparison figures); choose according to needs.
- 425W power delivery : 12VHPWR carries high current at Full Load — cable crimp quality and full seating are safety prerequisites (per public reporting).
- Compliance volatility : Shifting export-control policy affects supply and pricing (per public reporting) — verify current channel compliance documentation before purchasing.
FAQ FAQ
What's the difference between the 4090 D and the 4090?
Official figures: CUDA 16384→14592 (roughly -11%), TGP 450W→425W, with the same 24GB GDDR6X 1008GB/s memory — the D variant is a compliance trim for China export (per public reporting), and ranks below the 4090 in performance.
Can the 4090 D run large models?
Single 24GB card: 7B-14B-class FP16 full precision fits, 32B-class INT4 Quantization fits, 70B-class INT4 sits at the limit (estimated figures) — 24GB is the sweet spot for consumer local LLM workloads; dual cards (48GB) extend to LoRA scenarios.
Is there a D variant of the 50 series?
The 5090 D line is supplied in China mainland in compliant form (per public reporting; see media and channel listings for variants like 5090D/5090DD) — this entry uses the 4090 D as the anchor for the D production line; for 50-series D specs, refer to the official current pages.
References REFERENCES
- NVIDIA: GeForce RTX 4090 D page (China region, verified 2026-11) · Original
- NVIDIA Newsroom (January 2024 launch of the RTX 40 SUPER series and 4090 D) · Original