RTX 5090 vs RTX 5090D for AI: what is the difference?

Short answer: The RTX 5090D is a version of the RTX 5090 that NVIDIA released for a specific market under export rules. The 32 GB RTX 5090D has the same 32 GB of GDDR7, 512-bit memory bus, 21,760 CUDA cores and 575 W board power as the RTX 5090; the difference is Tensor Core throughput, which NVIDIA rates at 2,375 AI TOPS on the 5090D against 3,352 on the RTX 5090 (NVIDIA RTX 5090 page, flopper.io RTX 5090 D 32GB entry, read 2026-11-04). The same models fit on both. How much the lower rating changes your job depends on the workload, so time a short run on each card.

Specs side by side

RTX 5090 RTX 5090D (32 GB)
Architecture Blackwell Blackwell
VRAM 32 GB GDDR7 32 GB GDDR7
Memory interface 512-bit 512-bit
CUDA cores 21,760 21,760
AI TOPS (5th-gen Tensor Cores) 3,352 2,375
Total graphics power 575 W 575 W
Compute capability 12.0 not on NVIDIA’s CUDA GPUs list; check with nvidia-smi

Sources: NVIDIA’s RTX 5090 page and CUDA GPUs list; RTX 5090D figures from the flopper.io spec entry, which attributes the 2,375 AI TOPS rating to NVIDIA (all read 2026-11-04). NVIDIA later released an “RTX 5090 D v2” with 24 GB on a 384-bit bus; the RTX 5090D on CloudGPU is listed at 32 GB in /gpu-prices.json. Check what your instance reports:

nvidia-smi --query-gpu=name,memory.total,compute_cap,power.limit --format=csv

What the lower AI TOPS rating means

  • What fits is the same. Both cards have 32 GB, so the model and context limits are identical: FP8 FLUX.1-dev and SDXL fit; the full 16-bit FLUX.1-dev pipeline (about 31.4 GB loaded in BF16, Hugging Face Diffusers blog) is at the limit; a 32B LLM fits at 4-bit (about 19-20 GB of weights) but not at FP8 (Qwen3-32B-FP8 is 34.3 GB on Hugging Face). See how much VRAM an LLM needs.
  • AI TOPS is a Tensor Core figure. It is a theoretical Tensor Core peak, not a measured speed. Work that keeps the Tensor Cores busy, such as diffusion sampling, long prompt prefill, batched inference and training, is where a lower rating can show. NVIDIA does not publish how each workload is affected, so this page does not quote a percentage.
  • Memory bus and capacity are the same. Single-user LLM token generation is usually limited by memory bandwidth rather than Tensor Core peak, so it is the case where the two cards are most likely to be close. Measure it rather than assume it.
  • Software is the same. Both are Blackwell GeForce cards. If nvidia-smi reports compute capability 12.0, you need a PyTorch build for CUDA 12.8 or newer, as on the RTX 5090; see how to fix CUDA “no kernel image” on the RTX 5090.

Benchmark both in a few minutes

Rent each card for a short session and run the same three checks. Billing is per minute with a 1-minute minimum, so a test like this costs a few minutes of each card’s hourly rate.

1. Raw Tensor Core throughput (PyTorch BF16 matmul):

# matmul_bench.py — measures sustained BF16 matmul TFLOPS
import time, torch
n = 8192
a = torch.randn(n, n, device="cuda", dtype=torch.bfloat16)
b = torch.randn(n, n, device="cuda", dtype=torch.bfloat16)
for _ in range(10):
    a @ b
torch.cuda.synchronize()
iters = 100
t = time.time()
for _ in range(iters):
    a @ b
torch.cuda.synchronize()
dt = time.time() - t
print(torch.cuda.get_device_name(0), f"{2 * n**3 * iters / dt / 1e12:.1f} TFLOPS (BF16)")

2. LLM prefill and generation speed (llama.cpp): build llama.cpp as in how to build llama.cpp with CUDA, then:

# pp = prompt processing (compute-heavy), tg = token generation (bandwidth-heavy)
./build/bin/llama-bench -m ~/models/Qwen3-32B-Q4_K_M.gguf -p 2048 -n 256 -ngl 99

3. Your real job: time one image at your resolution and step count in ComfyUI, or one epoch of your fine-tune, and divide the hourly rate by the work done per hour.

Rates

At 2026-11-04 19:05 (UTC+8): RTX 5090D $0.65/hour, RTX 5090 $0.69/hour. A 15-minute benchmark costs $0.163 on the RTX 5090D and $0.172 on the RTX 5090. Billing is per minute, rounded up, 1-minute minimum; destroying the instance stops billing. Current rates are on /pricing, and which cards can be started now is shown on /deploy, where each card also shows its vCPU and memory.

How to decide

  1. If your job is memory-bound (single-user chat with a quantised LLM, an Ollama endpoint), run llama-bench’s tg test on both and compare tokens per second per dollar.
  2. If your job is Tensor Core-bound (FLUX or Wan sampling, batched vLLM serving, LoRA training), compare images, requests or steps per dollar from a short run on each card.
  3. If the model does not fit in 32 GB on either card, neither is the answer: see which 48GB GPU to rent or the RTX 4090 vs 4090 48G vs 5090 comparison.

Frequently asked questions

Is the RTX 5090D slower than the RTX 5090?

On Tensor Core work it is rated lower: 2,375 AI TOPS against 3,352. Memory size, memory bus and CUDA core count are the same. How much slower a given job runs depends on how much it relies on Tensor Core throughput, so time your own workload.

Does the RTX 5090D have less VRAM?

The 32 GB RTX 5090D has the same 32 GB of GDDR7 as the RTX 5090. A later “RTX 5090 D v2” has 24 GB; the RTX 5090D on CloudGPU is listed at 32 GB. Check nvidia-smi on your instance.

Can I run the same models on both?

Yes. With the same 32 GB, the same models, precisions and context lengths fit. The software setup is also the same: Blackwell needs a PyTorch build for CUDA 12.8 or newer.

Why is it called “D”?

NVIDIA released D versions of some GeForce cards for a specific market to comply with export rules. The 5090D keeps the RTX 5090’s memory and CUDA cores and has a lower AI TOPS rating.

Where are the prices?

This page shows the live rates when it is rendered; /pricing has the current rate for every card.

Sources (read 2026-11-04, Asia/Shanghai)

The Deploy page lists the cards you can start. Rates on this page are read from CloudGPU’s live price list when the page is rendered; /pricing has the current rate for every card.

Last updated: 2026-11-04 (Asia/Shanghai).