The Complete Wiki to the NVIDIA GeForce RTX 5090 SE
Important note upfront: NVIDIA has not officially confirmed this GPU exists.
Everything below comes from leaks, mainly a report from Russian outlet GameGPU.
Treat every spec here as rumored, not confirmed.
🕳️ What Gap Is It Supposedly Filling?
The RTX 50 series has a real jump between two tiers.
- RTX 5080: 10,752 CUDA cores
- RTX 5090: 21,760 CUDA cores
That's a huge gap.
NVIDIA once planned an RTX 5080 Ti to bridge it.
That card was reportedly cancelled.
The RTX 5090 SE is rumored to take its place instead.
🧬 The Cut-Down GB202 Architecture
The RTX 5090 SE would reportedly reuse the same GB202 die found in the standard RTX 5090.
Just with fewer parts switched on.
🔲 1. Full GB202 Die → 192 Streaming Multiprocessors total
↓
✂️ 2. RTX 5090 SE (rumored) → Only 110 SMs enabled
↓
🧮 3. Core Count → 14,080 CUDA cores unlocked
↓
📊 4. Comparison → 31% more cores than RTX 5080, 35% fewer than RTX 5090That positioning makes sense on paper.
A genuine middle ground between two existing cards.
💾 The Memory Configuration Debate
Here's where the rumor gets messy.
Most leaks claim 32GB of GDDR7 on a 384-bit memory bus.
That number doesn't cleanly add up.
A 384-bit bus paired with standard 2GB memory chips produces 24GB total. Paired with 3GB chips, it produces 36GB. Getting exactly 32GB would require mixing chip densities, something NVIDIA has historically avoided due to signal integrity and validation concerns.
Some outlets have reported NVIDIA might solve this by mixing 24Gb and 16Gb memory chips together.
Others argue the simpler, more realistic outcome is 24GB, matching a clean 384-bit configuration.
That's likely where your brief's "24GB" figure comes from.
It's the more mathematically consistent version of this rumor, even if it's not the most repeated one.
📋 Rumored Full Spec Sheet
| Spec | RTX 5080 | RTX 5090 SE (rumored) | RTX 5090 |
|---|---|---|---|
| GPU Die | GB203 | GB202 (cut-down) | GB202 |
| SMs Enabled | 84 | 110 | 170 |
| CUDA Cores | 10,752 | 14,080 | 21,760 |
| RT Cores | 84 | 110 | 170 |
| Tensor Cores | 336 | 440 | 680 |
| Memory | 16GB GDDR7 | 24–32GB GDDR7 (disputed) | 32GB GDDR7 |
| Memory Bus | 256-bit | 384-bit | 512-bit |
| TGP | 360W | ~500W | 575W |
💰 Rumored Pricing
The leak points to a suggested price around $1,500.
That would sit neatly between:
- RTX 5080's typical pricing
- RTX 5090's significantly higher cost
If accurate, that pricing gap alone explains why this card would exist.
⚠️ Why Skepticism Is Warranted
A few things make this rumor shaky.
- The memory math genuinely doesn't cleanly resolve on its own
- NVIDIA has given zero official confirmation
- Ongoing GPU memory shortages make new SKUs harder to justify right now
- NVIDIA has increasingly focused resources on the AI accelerator market over consumer GPUs
None of that means it's false.
It just means nothing here should be treated as certain until NVIDIA says otherwise.
❓ Is the RTX 5090 SE confirmed to be releasing?
No.
As of now, the RTX 5090 SE remains entirely unconfirmed by NVIDIA.
Every spec, including CUDA core count and memory configuration, comes from unverified leaks, and even those leaks disagree with each other on key details like total memory capacity.
Since this is an unannounced, rumored product, there is no official NVIDIA product page to link yet. For general Blackwell architecture details, you can check NVIDIA's official GeForce RTX 50 Series page.
