TL;DR: Don't wait for RDNA 5 if you need a GPU today. It's likely 8–12 months out, won't launch at MSRP immediately, and AMD's software support for local LLM inference will still trail Nvidia at launch. If you want AMD, RDNA 4 (RX 9070 XT) is available now and genuinely competitive. If you want to wait for something, wait for RTX 5090 prices to normalize — that's a more concrete near-term opportunity.
"Should I wait for [next GPU generation]?" is the most common hardware question and almost always has the same answer: no. But the RDNA 5 question deserves a more specific response, because there are real factors on both sides.
Here's the honest breakdown.
What We Know About RDNA 5
AMD has confirmed that RDNA 5 is in development and on their roadmap. Based on AMD's historical release cadence and current fab timelines on TSMC N3/N2 process nodes, the earliest realistic RDNA 5 launch is late 2026 — Q4 at optimistic, 2027 at conservative.
What AMD has not confirmed publicly:
- Specific VRAM configurations (how much GDDR7 or HBM the cards will use)
- Target price points
- Architecture details for AI workload optimization
- Launch window with any precision
Every "RDNA 5 leak" floating around right now is speculation based on supply chain rumors. Treat it accordingly.
The Case for Waiting
There are legitimate reasons to consider waiting, but they're narrow:
If you specifically want AMD hardware: RDNA 5 should bring architectural improvements over RDNA 4. If you're committed to the AMD ecosystem for philosophical or practical reasons (Linux ROCm workflow, game performance priority, whatever), waiting for the generation-over-generation improvement makes sense — but only if you can genuinely wait 8–12+ months.
If you need 24GB+ VRAM and don't want Nvidia: AMD currently doesn't offer a consumer card with 24GB VRAM. Rumors suggest RDNA 5 may address this with higher VRAM configs on the top-tier cards. If that's your specific need and you're patient, waiting is rational.
If you're buying for reasons beyond LLMs: For gaming, rendering, or mixed compute workloads, RDNA 5 might offer meaningful improvements. But for pure local AI inference, the software gap with Nvidia means specs alone don't translate linearly to real-world performance.
The Case Against Waiting
Here's why "wait for RDNA 5" is bad advice for most people:
8–12 months is a long time. You could run your local AI workflow on a GPU you buy today for 8–12 months and have real productivity while waiting for a card that may or may not meet expectations at launch.
New GPU launches are never the cheapest moment to buy. RDNA 5 will launch at whatever MSRP AMD sets, in limited quantities, with early stock constraints. The RTX 5090 launched at $1,999 MSRP and immediately sold for $4,000+ on the secondary market. RDNA 5 won't be immune to this — if it's competitive with top-tier Nvidia, it'll face the same demand pressure.
AMD's software support gap won't disappear at launch. Even if RDNA 5 has superior hardware specs for AI inference, the ROCm ecosystem, llama.cpp HIP optimization, and community tooling will still be behind CUDA on day one. That gap has closed significantly with RDNA 4, but closing versus eliminating are different things.
RDNA 4 is already here and competitive. The RX 9070 XT at ~$599–649 is a real GPU you can buy now with 16GB GDDR6 and performance within 20–30% of Nvidia cards at similar price points. If AMD hardware meets your needs, it's already available.
What Actually Might Be Worth Waiting For (Near Term)
If you're going to wait on anything, these are more concrete opportunities:
RTX 5090 price normalization: The 5090 is the best single consumer card for local AI — 32GB GDDR7, 1,792 GB/s bandwidth. It's currently at $3,800–4,500 street price against a $1,999 MSRP. As Nvidia ramps production through 2026, prices should normalize. Waiting 3–4 months for the 5090 to hit closer to MSRP is a more tangible strategy than waiting 8–12 months for RDNA 5.
RTX 5070 stock at MSRP: The 5070 Ti and 5070 launched with stock shortages that pushed prices above MSRP. Inventory is building. If you want a 5070 Ti at $749 rather than $900, waiting 4–8 weeks for stock to normalize is sensible.
Used 40-series prices dropping further: The RTX 40-series selloff is ongoing as 50-series adoption grows. The RTX 4070 Ti Super and 4080 will likely continue declining through mid-2026. If you want a 40-series card, prices may be 10–15% better in 3–4 months.
The Specific Scenario Matrix
You need a GPU in the next 30 days: Buy now. RTX 5070 Ti if budget allows, RTX 3090 used if you need 24GB at lower cost, Arc B580 if budget is under $300.
You can wait 2–3 months: Buy now anyway, or wait for 5090 prices to normalize and 5070 Ti stock to settle near MSRP. RDNA 5 in 2–3 months is not realistic.
You can wait 6 months: RTX 5090 prices may be reasonable. 40-series used prices will be lower. RDNA 5 is still not realistic in 6 months.
You can wait 12+ months: RDNA 5 becomes possible in this window, but you're essentially waiting for an announcement that hasn't been made on hardware that hasn't been benchmarked.
You're an AMD-first person on Linux: RDNA 4 (RX 9070 XT) is the right buy right now. ROCm 6.x on Linux is solid and the hardware is genuinely good. Don't wait 12 months.
The Honest Summary
The GPU market always has something better on the horizon. RDNA 5 will be better than RDNA 4. The GPU after RDNA 5 will be better than RDNA 5. At some point you have to buy the hardware that exists and run your workflows on it.
For local LLM work in 2026, the hardware available right now is the best it's ever been and more affordable per TFLOP and per GB of VRAM than any previous generation. The case for waiting is theoretical; the case for buying now and running models today is practical.
If you're an AMD enthusiast specifically waiting for RDNA 5 to solve the software gap with Nvidia — that's not how software gaps work. ROCm improvement is an ongoing process that will continue regardless of which GPU generation ships. RDNA 5 hardware won't make llama.cpp's HIP backend suddenly equivalent to its CUDA backend overnight.
Buy the best card you can afford today. Run models. Upgrade when RDNA 5 ships if it's compelling. That's the honest answer.