That question matters more than a headline about new cards costing “double MSRP.” Retail pricing in the GeForce RTX 50-series has been sharply elevated in September 2026, but unevenly so. The RTX 5090 is the extreme case: it launched at $1,999 and has been reported above $5,000. The RTX 5060, introduced from $299, has been reported around $448. That is still a substantial increase, but it is not double MSRP. Reported average Blackwell-family price growth of about 60% likewise does not justify treating every current card as a 2×-MSRP product.
An expensive new-card market can make an older 24GB or 16GB card worth investigating. It cannot, by itself, establish that the RTX 3090, Radeon RX 7900 XTX, Radeon RX 6800 XT, or GeForce RTX 3060 12GB is a better gaming purchase than a current alternative. The necessary transaction and performance evidence for that conclusion is not established here. This guide instead sets out what must be true before one of those cards deserves your money.
Start with a reproducible used-price test
A used-GPU recommendation needs verified transaction data, not active listings and not opaque price-tracker estimates. The following is a reproducible U.S. method for evaluating the four cards as of September 14, 2026.
Use a 30-day window running from August 16 through September 14, 2026, inclusive. Count only completed U.S. sales with a visible final sale price. Record fixed-price, accepted-offer, and auction transactions only when the final transaction amount is visible. Do not use unsold listings, “best offer” asking prices, local-cash advertisements without a documented completed price, or listings that merely say an item was removed.
For all four cards, treat only retail desktop PCIe boards with their original air cooler as part of the standard board class. Keep reference designs and conventional air-cooled partner boards in the same broad class, but record the exact board maker and model so unusually premium coolers do not disappear inside a generic average. Do not combine different GPUs simply because their names are similar:
- RTX 3090 means the 24GB GeForce RTX 3090, not RTX 3090 Ti.
- RX 7900 XTX means RX 7900 XTX, not RX 7900 XT.
- RX 6800 XT means RX 6800 XT, not RX 6800.
- RTX 3060 means the 12GB model, not a differently configured RTX 3060 variant.
Exclude cards listed as defective, for parts, repair-only, untested, no-display, artifacting, intermittent, damaged, corroded, incomplete, or with missing fans or heatsinks. Also exclude OEM pulls, cards fitted with water blocks, cards sold without their original air cooler, engineering samples, and listings with undisclosed or clearly modified cooling. Those products may have legitimate uses, but they are not comparable to a normal retail card a typical Windows gamer can install and return with reasonable confidence.
For every included sale, calculate pre-tax delivered price: final item price plus mandatory shipping. Sales tax should be reported separately rather than folded into a national figure, because it changes by buyer location and checkout circumstances. The published result should provide the observation count, the dates covered, each included board model, the median pre-tax delivered price, and the low-to-high range.
That process has not been completed with an independently auditable transaction sample in the available evidence. Consequently, there is no verified observation count, dated median, range, or dollar buy ceiling for any of the four cards in this assessment. Commercial tracker numbers are deliberately excluded from the decision logic: their collection rules and condition treatment are not transparent enough, and the available figures materially conflict for some Radeon models.
The resulting condition-adjusted maximum-buy rule is necessarily firm, if not numerical: do not buy any of these cards as a value recommendation until a qualifying completed-sale sample establishes the clean-card median for the exact class above. A fully tested, complete air-cooled card can be compared with that median. An untested, modified, incomplete, OEM, or water-blocked card cannot be priced as though it were equivalent, regardless of how attractive its sticker price looks.
Gaming evidence must come before a shortlist
Memory capacity and board-power ratings describe important characteristics. They do not demonstrate gaming value, ray-tracing ability, frame-time consistency, image quality, or suitability for a particular Windows setup.
No current controlled Windows gaming comparison is established here for the RTX 3090, RX 7900 XTX, RX 6800 XT, and RTX 3060 12GB against relevant new cards. In particular, there are no comparable results identifying the game suite, Windows version, driver versions, CPU platform, resolutions, quality presets, raster results, ray-tracing results, or frame-time measurements. There are also no substantiated DLSS or FSR quality-and-performance comparisons across the four used candidates and current alternatives.
That absence prevents a responsible declaration that any of the four belongs on a general gaming shortlist. It also prevents claims that a larger memory allocation outweighs newer architecture, lower power requirements, more consistent frame times, or better ray tracing.
Before buying, a comparable test set should cover at least 1080p, 1440p, and 4K where the card class makes those modes realistic; use identical quality settings; separate native raster from ray-traced results; and document the exact graphics driver and game build. Average frame rate alone is insufficient. The review must also report frame-time behavior, because a card that produces an acceptable average can still deliver noticeable stutter or poor lows.
The comparison needs relevant new-card alternatives, too. For the 24GB RX 7900 XTX, the Radeon RX 9070 XT and RX 9070 are obvious current AMD reference points in memory-capacity terms, although their documented 16GB configurations do not establish their comparative gaming performance. For the RTX 3090, current sub-RTX-5090 GeForce RTX 50-series desktop cards are relevant capacity comparisons: they top out at 16GB or 12GB, while the RTX 5090 has 32GB. For the RX 6800 XT and RTX 3060 12GB, the correct new-card comparison depends on the buyer’s resolution, games, budget, and required features—not merely on VRAM.
Four cards to assess conditionally
GeForce RTX 3090: 24GB is the reason to investigate
The RTX 3090 launched in September 2020 at $1,499. It has 24GB of GDDR6X memory, 10,496 CUDA cores, and a 350W graphics-card power rating. Its strongest present-day case is memory capacity. It exceeds every listed sub-RTX-5090 desktop GeForce RTX 50-series model in VRAM, though the RTX 5090 itself has 32GB.
That makes an RTX 3090 potentially relevant when 24GB is a hard requirement for game assets, creative work, or another workload. It is not proof that the card is the better gaming buy. A 350W rating also makes power-supply quality, available PCIe power connectors, case dimensions, GPU support, exhaust airflow, and noise tolerance part of the purchase decision.
DLSS 4.5 Super Resolution support extends across GeForce RTX GPUs, including RTX 30-series products. But NVIDIA says RTX 20- and 30-series cards lack native FP8 support and may take a greater performance hit with the newest DLSS 4.5 models. Compatibility is useful; it is not evidence of identical performance or behavior to newer hardware.
Radeon RX 7900 XTX: capacity without a low-power profile
AMD specifies 24GB of GDDR6 memory, 96 compute units, and 355W typical board power for the RX 7900 XTX. Like the RTX 3090, it is most interesting where 24GB is meaningful. It also has more memory than the documented released RDNA 4 gaming products, whose listed configurations reach 16GB.
Its 355W typical board power demands the same kind of practical planning as the RTX 3090. A low-quality or marginal power supply, poor case ventilation, or a cramped chassis can turn an apparently favorable used purchase into an expensive system rebuild.
For gaming, memory alone settles very little. Do not assume that a 24GB older card automatically beats a newer 16GB card in raster performance, ray tracing, frame-time stability, image reconstruction, or the particular games you play. It needs the controlled comparison described above and a completed-sales discount that reflects used-hardware risk.
Radeon RX 6800 XT: 16GB, 300W, and a future-software caveat
The RX 6800 XT has 72 compute units, 16GB of GDDR6 memory, and 300W typical board power. Its power class is lower than the two 24GB cards here, although 300W still calls for a sound power supply and sensible airflow.
AMD has renamed FSR 4 to FSR Upscaling. The company says ML-based FSR Upscaling support for RX 6000-series cards is planned for 2027. That is a forward-looking compatibility commitment, not a feature buyers can assume is available now, and there is no documented basis for calling the schedule “early 2027.” It should not justify paying a premium for an older card today.
The card can be a sensible candidate only if current game testing meets the buyer’s target resolution and if audited completed sales show an appropriate delivered-price advantage over suitable new alternatives. Neither conclusion follows from its 16GB specification alone.
GeForce RTX 3060 12GB: a narrower capacity-led proposition
The RTX 3060 12GB is not comparable in tier to the RTX 3090, RX 7900 XTX, or RX 6800 XT. Its distinctive trait is the 12GB GDDR6 configuration on a 192-bit interface. That can matter for a lower-cost build, but it does not erase the need to check actual performance at the desired settings.
It too supports DLSS 4.5 Super Resolution as part of the GeForce RTX family. The same FP8 qualification applies: NVIDIA says RTX 30-series products lack native FP8 support and can suffer a larger performance impact with the newest DLSS 4.5 models. For a card in this class, buyers should be especially careful not to interpret a supported feature label as a promise of smooth performance with demanding settings.
This is a conditional option only when the exact 12GB model is fully tested, its delivered price is supported by clean completed-sale evidence, and current testing shows it meets the target for the games and display in question.
Do not let an abnormal market lower your standards
Used hardware needs a stronger condition check than a new retail card. Ask for current proof that the GPU is recognized in Windows and can complete a graphics workload. Inspect photographs for damaged connectors, corrosion, stripped screws, missing fasteners, modified thermal work, bent heatsink fins, and fan damage. Confirm physical clearance, power connectors, and return terms before purchase.
Intel Arc is a separate case rather than evidence for or against these four cards. Intel has documented end-of-life status for the Arc A770 8GB bulk SKU, with final shipment dated December 31, 2025. That does not establish that every Alchemist board is discontinued: the Arc A770 16GB remains listed as launched. Intel also documents a hybrid DirectX 9 approach that uses translation layers where they offer a better experience. That design choice does not prove every older DirectX 9 game will have poor frame pacing; Windows gamers should check the specific older titles that matter to them.
Elevated retail prices make disciplined comparison more valuable, not less. A used high-VRAM GPU becomes attractive only after three tests are passed: a reproducible completed-sales price sample, controlled gaming evidence for the buyer’s actual resolution and games, and a condition inspection that accounts for power, cooling, and return risk. Until then, the four cards above are categories to research—not substantiated bargains to chase.