NVIDIA GeForce RTX 5090D V2 24GB Blower (China Edition)

A China-market variant of the RTX 5090 with 24GB GDDR7 (instead of 32GB) on a 384-bit bus at 1,344 GB/s (instead of 1,792 GB/s) to comply with US export bandwidth restrictions. Same 21,760 CUDA cores as the standard 5090. AI Tensor TOPS firmware-limited to 2,375 (versus 3,352). Launched August 2025 in China. The blower variant at MillionMiner is an aftermarket conversion for multi-GPU rack deployment.

Two reasons. Less VRAM (24GB versus 32GB reduces manufacturing cost and falls under export restrictions). China-market pricing (launched at 16,499 yuan, approximately $2,250 USD). At $2,950, MillionMiner's price includes the aftermarket blower conversion markup and international availability.

At $2,950: the 5090D V2 delivers 33 percent more CUDA cores (21,760 versus 16,384), 33 percent more bandwidth (1,344 versus 1,008 GB/s), Blackwell versus Ada Lovelace architecture, fifth-gen FP4 versus fourth-gen FP8 Tensor Cores, and PCIe Gen 5 versus Gen 4. Same 24GB VRAM. The 5090D V2 is the better card at equivalent pricing on every metric.

NVIDIA firmware-limited the Tensor Core throughput to 2,375 AI TOPS (versus 3,352 on the standard 5090, a 29 percent reduction). This affects AI training and inference workloads that rely heavily on Tensor Core precision (FP8, FP4, INT8 operations). General CUDA compute and gaming are unaffected since those use CUDA cores, not Tensor Cores for primary computation.

No. GPU mining uses CUDA cores for hashing algorithms, not Tensor Cores. Mining hashrate on the 5090D V2 should be within 5 percent of the standard 5090 on core-intensive algorithms (KHeavyHash, KawPow, Autolykos2). Memory-intensive algorithms (Ethash) will see approximately 25 percent lower hashrate due to the bandwidth reduction.

24GB handles: 7B to 13B models at FP16, 25B at INT8, stable diffusion standard workflows, LoRA fine-tuning on 7B to 13B. Needs 32GB: 13B+ at FP16 with large batches, 25B+ at FP16, stable diffusion with very large batch sizes, workloads approaching or exceeding 24GB VRAM. If your work fits in 24GB, the 5090D V2 at $2,950 is the smarter buy than the standard 5090 at $5,000.

Expect similar core-intensive hashrates to the standard 5090 (same 21,760 CUDA cores at same clocks). Ethash and other memory-bandwidth-dependent algorithms will see approximately 25 percent lower hashrate due to the 1,344 versus 1,792 GB/s bandwidth reduction. Net mining profitability should be similar to the RTX 4090 or slightly better given the architectural improvements.

Same CUDA cores, same clocks, same architecture, same blower form factor. The 5090D V2 trades 8GB VRAM and 25 percent bandwidth for roughly 40+ percent lower cost. For workloads that fit in 24GB and are compute-bound: the 5090D V2 delivers nearly identical performance at significantly lower price. For workloads needing 32GB or maximum bandwidth: the standard 5090 Blower is required.

Same dual-slot form factor as the 5090 Blower: up to 8 per server chassis with appropriate infrastructure. Eight cards: 192GB combined (versus 256GB on standard 5090 blower), approximately 838 TFLOPS FP32 aggregate (same compute), 4,600W GPU draw.

The 5090D V2 at $2,950 is the strictly better card at the same price: more cores, more bandwidth, newer architecture, better Tensor Cores, PCIe Gen 5. The only reason to choose the 4090: if you specifically need NVLink (4090 does not have it either, but the 3090 at $950 does), or if you prefer factory NVIDIA hardware over an aftermarket blower conversion.