Why Your RTX 5090 Might Never Reach Its Full Potential

Usually because something else in the system sets the speed limit. The RTX 5090 is so fast that at 1080p and 1440p even a Ryzen 7 9800X3D can hold it back, and an older CPU costs it 20–30% at low resolutions. In AI work the limit is often memory: once a model or its cache spills out of the 32 GB into system RAM, speed collapses. Less often, it's starved of power by a half-connected adapter, throttled by heat, or held back by old software. A PCIe 4.0 motherboard, the worry most owners raise first, costs only about 1% in games. Here's how to find which limit you're hitting.

Where does the performance go?

Source: as listed. CPU results are four-game geometric means; TechPowerUp's PCIe figures are multi-game averages. AI figures depend heavily on host memory speed.

Is your CPU holding it back?

In games, almost certainly below 4K. With a 5090, the 9800X3D beat a Ryzen 5 7600X by 37% at 1080p Medium, 26% at 1080p Ultra and 11% at 1440p Ultra, narrowing to 8% at 4K Ultra. The 5090's lead over the RTX 4090 follows the same pattern: about 24–35% at 4K but roughly 12–20% at 1440p and 1080p, because the CPU, not the GPU, decides the frame rate. Some games stay CPU-bound even at 4K: Space Marine 2 still showed a 34% CPU gap at 4K Ultra.

Frame generation can disguise this. It adds frames on the GPU without new CPU work, so the counter rises, but responsiveness still follows the rendered frame rate. Hardware Unboxed recommends a base of 70–80 fps before turning it on. If you mainly play esports titles at 1080p, a 5090 with an older CPU is the most expensive way to be CPU-limited.

Does PCIe 4.0 matter?

Barely, for one card in games. TechPowerUp measured a 5090 at about 99% of its PCIe 5.0 x16 performance on PCIe 4.0 x16, and about 96–98% on PCIe 3.0 x16. Gamers Nexus "did not find a significant difference". Two cases are different:

• Slot sharing. On some boards, filling a particular M.2 slot drops the main slot from x16 to x8. ASUS's ROG Crosshair X870E Hero does this when M.2_2 is used. On PCIe 5.0 that's harmless; on an older PCIe 4.0 board it's the same as PCIe 3.0 x16.
• Bandwidth-hungry apps. Puget Systems found DaVinci Resolve about 10% slower at the 16 GB/s tier and about 25% slower at PCIe 3.0 x8 or 4.0 x4, while Blender and Octane were unaffected. Model loading and anything offloaded to system memory also crosses PCIe.

We go deeper in RTX 5090 on PCIe 5.0 vs 4.0.

Is it getting enough power and cooling?

The card can't run at full speed if it can't draw full power. NVIDIA's adapter needs four 8-pin inputs; with three, ComputerBase saw the card capped at 450 W and about 5% slower, and with two it didn't boot. NVIDIA asks for 1,000 W of system power.

Heat matters both ways. The Founders Edition's GDDR7 ran at 88–96 °C in reviews, and Tom's Hardware saw a Cyberpunk crash after about 20 minutes. Its exhaust also heats the CPU: ComputerBase measured a 9800X3D at 95 °C beside it at low fan speeds, and a CPU at its thermal limit loses boost clocks. Good intake airflow fixes most of it. See 575 watts for one graphics card and air vs liquid cooling.

A slow RTX 5090 is rarely a slow card. It's usually a fast card waiting on the CPU, the power cable, the airflow or system memory.

Is memory the limit for AI?

For language models, generating tokens is mostly limited by memory bandwidth, and the 5090's 1,792 GB/s is its biggest advantage. That only helps while everything stays in its 32 GB. Qwen3 32B at 4-bit fitted about 45,000 tokens of context on a 5090 in hardware-corner's tests; beyond that, the cache has nowhere to go. A model too big for VRAM runs from system RAM over PCIe, which is why gpt-oss-120b drops to roughly 10–30 tokens per second on a 5090 depending on system memory. The fix isn't a faster card but a smaller model, a quantised cache or more VRAM. See why 32 GB still frustrates AI developers.

Is your software holding it back?

• Old CUDA builds. Blackwell needs CUDA 12.8 or newer. PyTorch 2.7 added CUDA 12.8 wheels; tools that pin older PyTorch builds won't run properly on a 5090.
• The engine you choose. On identical RTX PRO 6000 cards, which use the same chip, switching LLM engine changed throughput 7.8 times in one published test (Helix). Try llama.cpp, vLLM or SGLang for your workload.
• Drivers. Blender's OptiX renderer needs driver 575 or newer. NVIDIA says either its Game Ready or Studio driver can run games and creative apps; Studio targets creative-app stability.
• Settings. Ollama's default 4,096-token context and unquantised KV caches limit what fits. In games, a frame cap or V-Sync can hold the card well below its potential.

How do you get the most out of it?

• Pair it with a current high-end CPU, ideally one with large cache such as a Ryzen 7 9800X3D, if you play below 4K.
• Use a native 12V-2x6 cable or all four adapter inputs, on a 1,000 W-plus PSU.
• Keep the main slot at x16 if your board is PCIe 4.0; check the manual's M.2 sharing table.
• Plan intake airflow for a card that dumps up to 575 W into the case.
• For AI, keep models and caches inside 32 GB, and keep CUDA and your inference engine current.

Buying in Singapore

RTX 5090 cards are not in stock at SourceIT at the time of writing; ask us about availability and a quotation. To remove the CPU bottleneck, the AMD Ryzen 7 9800X3D is S$756.38 and the Ryzen 9 9950X3D is S$1,108.02, and a 4K display such as the AOC AGON PRO AG276UZD 4K 240 Hz QD-OLED is S$1,295.00 (all listed 6 Nov 2026). Prices change; ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

What CPU should I pair with an RTX 5090? For gaming below 4K, the fastest gaming CPU you can afford. In TechSpot's tests the Ryzen 7 9800X3D was 37% faster than a Ryzen 5 7600X at 1080p Medium with a 5090, but only 8% faster at 4K Ultra.

Will a PCIe 4.0 motherboard slow down my RTX 5090? In games, by about 1% at PCIe 4.0 x16. Bandwidth-heavy apps such as DaVinci Resolve can lose more at reduced lane counts, and AI models offloaded to system memory are affected most.

Why is my RTX 5090 only slightly faster than my old RTX 4090? Probably because you're CPU-limited. At 1440p and 1080p the 5090's lead over the 4090 drops to roughly 12–20%, against 25–35% at 4K.

Does the adapter cable affect performance? Yes, if it isn't fully connected. With three of four 8-pin inputs the card was capped at 450 W and ran about 5% slower; with two it didn't start.

Why are my AI models slow on an RTX 5090? Usually because the model or its context doesn't fit in 32 GB and is spilling into system RAM, or because the software uses an older CUDA build. Check memory use, context length and your PyTorch or engine version.