The short version
- The RTX 5090 offers 32 GB of GDDR7 and a launch price of $1,999, according to NVIDIA's product page. NVIDIA lists 3,352 AI TOPS, which is a peak vendor figure and not a tokens-per-second result.
- The RX 7900 XTX offers 24 GB of GDDR6. AMD's launch press release put its suggested price at $999, with availability from AMD.com starting December 13, 2022.
- AMD's ROCm 10.1 is documented as covering the RX 7900 family. That is a compatibility statement, not a performance claim.
- No dated, same-model, same-quantization benchmark of these two cards turned up in the verified material. Anyone who tells you one is "faster for local AI" is extrapolating.
Why VRAM matters more than TOPS
Local language models have to fit in video memory. If the weights plus the context cache don't fit, the runtime spills into system RAM and speed drops sharply. The 5090's extra 8 GB therefore buys longer context windows, higher-precision quantizations, or room for a second small model alongside the main one.
A common rule of thumb puts 4-bit weights at roughly half a byte per parameter. That makes a 32-billion-parameter model about 16 GB of weights before overhead. Treat that as a planning estimate only. Quantization format, context length, runtime and cache overhead all move the real number. A 24 GB card running tight under one context setting says nothing about how fast a 32 GB card would run the same job.
AMD's own RX 9070 XT page lists less memory than either card, per the research behind this piece. For local-model capacity it is the wrong card to buy, regardless of how it games.
Software: CUDA versus ROCm
NVIDIA. NVIDIA's developer archive lists CUDA Toolkit 13.4.3 (October 2026) as the latest release, which differs from the 13.4.2 figure some summaries quote. Toolkit version numbers don't measure framework compatibility. They do show that the NVIDIA stack is actively maintained, and most local-AI tools target CUDA first. That is general industry knowledge, not a measured claim.
AMD. AMD published ROCm 10.1.0 on October 5, 2026. Its release notes describe a focus on inference, developer tooling and profiling across Instinct, Radeon and Ryzen AI platforms. Several details matter to Windows users:
- AMD's Windows Adrenalin Edition for ROCm 10.1 notes list the Radeon RX 7900, 7800, 7700 and 7600 families as compatible.
- WSL2 support enters technical preview in 10.1. AMD says the ROCm runtime detects WSL2 automatically. The preview targets Ubuntu through .deb packages and Python wheels. Profiling, debugging and KFD-dependent tooling are not supported at launch.
- The release notes say a Windows fix resolves deadlocks and invalid values in multi-process workloads, along with issues seen running LLMs on multi-GPU Windows setups. A separate fix covers an out-of-memory problem on Strix Halo APUs.
- AMD's ROCm Linux documentation lists the RX 7900 XTX as RDNA3, target gfx1100, officially supported. That page has since been marked as moved to a newer compatibility matrix, so check the current matrix before installing.
One correction to circulating coverage: AMD's ROCm 10.1 blog text describes the bundled ROCm CLI as version 0.2.0 in the copy retrieved. Another AMD page reportedly says 1.0.0. Don't repeat either number as settled. Check the version you actually install.
The honest limit: AMD's documentation shows hardware and OS support. It does not show that every consumer LLM framework runs as smoothly on a 7900 XTX as on CUDA. Check your chosen tool's own support matrix.
The price picture
The one hard, dated number comes from PC Games Hardware. On October 15 the cheapest available RTX 5090 in its German comparison cost €6,699. By the publication's account, the cheapest model now costs €6,699, up from €5,999 on October 2. January's low was €3,400, and NVIDIA's recommended Founders Edition price is €2,099. That is a German retail snapshot. It is not a global price or an official NVIDIA price change.
The supply story is more serious than a price blip. PCGH reports that NVIDIA is no longer allocating new GB202 GPUs to board partners for GeForce cards, citing multiple independent sources. It also says existing GPUs and completed graphics cards can still be sold. PCGH says NVIDIA had not responded to its October 15 request for comment. Whether NVIDIA treats this as a formal end of life or a capacity reallocation is unclear.
NVIDIA's public position, per Tech-Insider's account of a Taipei interview, is that the RTX 5090 "has definitely not been discontinued" and shipments would continue, with supply-chain capacity acknowledged as constrained. That secondary account sits uneasily beside PCGH's sourcing. The two aren't strictly contradictory. A card can be "not discontinued" while receiving no new chips. For a buyer, the practical outcome is the same: scarce stock and high prices.
US figures are noisier. Tech-Insider reports Amazon resale listings around $8,500, and PCGH, citing TweakTown, notes some marketplace sellers asking over $10,000. These are asking prices from third-party sellers, not transaction data.
AMD's side is murkier. PC Gamer reported on October 13 that AMD had reportedly raised GDDR6 prices for add-in-board partners, effective October 1. The report relies on Boardchannels and Benchlife. It is not an AMD-confirmed retail increase, and it doesn't tell you what any specific 7900 XTX costs today.
I could not confirm current 7900 XTX street prices from comparable, dated retailer data. So I won't call the Radeon the best value at today's prices. At launch pricing the arithmetic does favor it: $999 for 24 GB against $1,999 for 32 GB. Whether that gap persists on shelves is something you have to check in live listings.
How to decide
- Do you need 32 GB on a single card? If yes, the 5090 is the only consumer option in this comparison. Be prepared to pay a heavy premium, or to wait.
- Do you run 14B to 32B models at 4-bit? A 24 GB card can plausibly handle these with moderate context, going by the rule of thumb above. Verify with your own model and context length.
- Is your toolchain CUDA-first? If you depend on niche repositories or custom CUDA kernels, expect friction on a Radeon. If you use mainstream runtimes with a documented ROCm or Vulkan backend, the gap is narrower.
- Do you game on the same machine? Let the gaming choice lead, then confirm the AI stack covers your tools.
- Is the workload temporary? At these prices, renting cloud GPU time may beat buying.
Measure it yourself
The fairest comparison is your own workload. Use the same model file, the same quantization, the same context length and the same runtime on each machine. Record these separately:
- prompt-processing speed
- token-generation speed
- peak VRAM
- wall power
Prompt processing and generation stress different parts of the GPU, so one card can lead one phase and trail the other. If you use llama.cpp's bundled benchmark tool, build it with the CUDA or HIP backend as that project's current documentation describes. Don't copy build flags from an old guide.
Bottom line
The RTX 5090 wins on capacity and software breadth on paper. The RX 7900 XTX has documented ROCm 10.1 Windows support and 24 GB at a much lower launch price. The 5090's real-world availability is the problem. Prices are volatile, the supply picture is contested, and no controlled benchmark settles the speed question. Test your workload, check live listings, and treat any single street price as a snapshot.
References
- RTX 5090 vs RX 7900 XTX for Local AI: 32GB Wins (2026) - shattered.io shattered.io · 2026-10-23T02:45:32+00:00
- NVIDIA Home nvidia.com
- Redirecting… — ROCm installation (Linux) rocm.docs.amd.com