NVIDIA GeForce RTX 4090 D (China, AD102-250)
nvidia ada (AD102-250) 2023
Fields
id
accelerators/nvidia-rtx-4090-dVendor
nvidia
Architecture
ada (AD102-250)
Release year
2023
Process (nm)
4
Form factors
pcie
VRAM (GB)
24
Memory type
gddr6x
Memory bus (bit)
384
Memory bandwidth (GB/s)
1,008
Bandwidth basis
VideoCardz reproduction of the NVIDIA China spec table: 24 GB GDDR6X at 21 Gbps on a 384-bit bus, bandwidth printed as 1008 GB/s. Arithmetic checks: 21 x 384 / 8 = 1008.
Peak FLOPS
precisionfp32tflops73.54denseyesvendor_claimnosource_idvideocardz-rtx-4090d-launchbasisderivedbasis_detailNot printed by NVIDIA. 14592 CUDA cores x 2 FLOPs x 2520 MHz boost = 73.54 TFLOPS. The core count and boost clock come from the NVIDIA China table as reproduced by VideoCardz; the multiplication is mine. FP32 has no sparse path, so dense:true.unitTFLOPS
TDP (W)
425
Unified memory
no
Consumer part
yes
notes
CHINA-ONLY CUT OF THE RTX 4090, shipped December 2023 to fit US export rules after the full AD102-300 4090 was restricted. 114 SMs / 14592 CUDA cores / 456 Tensor cores against 128 / 16384 / 512 on nvidia-rtx-4090, same 2520 MHz boost, same 24 GB GDDR6X at 21 Gbps (1008 GB/s), TGP cut from 450 W to 425 W, and no factory overclock allowed. Memory bandwidth is unchanged, so decode tok/s on a bandwidth-bound model should land close to a 4090; prefill, which tracks compute, should not. A used '4090 D' listing is not a 4090.
Notes
CHINA-ONLY CUT OF THE RTX 4090, shipped December 2023 to fit US export rules after the full AD102-300 4090 was restricted. 114 SMs / 14592 CUDA cores / 456 Tensor cores against 128 / 16384 / 512 on nvidia-rtx-4090, same 2520 MHz boost, same 24 GB GDDR6X at 21 Gbps (1008 GB/s), TGP cut from 450 W to 425 W, and no factory overclock allowed. Memory bandwidth is unchanged, so decode tok/s on a bandwidth-bound model should land close to a 4090; prefill, which tracks compute, should not. A used '4090 D' listing is not a 4090.