RTX 5090 vs 5090D vs 5090D V2: What Are the Differences, and What Happens If You Buy the Wrong One?

2026-10-18 ｜ Comparison ｜ 18 min ｜ 作者：Display Card King Editorial

The RTX 5090, RTX 5090D, and RTX 5090D V2 are three versions of NVIDIA's Blackwell-architecture flagship graphics card. They share the same CUDA core count (21,760), the same base clock (2.01 GHz), the same boost clock (

Key Takeaways

The RTX 5090, RTX 5090D, and RTX 5090D V2 are three versions of NVIDIA's Blackwell-architecture flagship graphics card. They share the same CUDA core count (21,760), the same base clock (2.01 GHz), the same boost clock (2.41 GHz), and the same 575W total graphics power. The real differences are in just two places. First, AI compute: the global RTX 5090 delivers 3,352 AI TOPS, while both D models deliver only 2,375 AI TOPS, about 29% lower. Second, memory: the RTX 5090 and 5090D both carry 32GB of GDDR7 on a 512-bit bus with 1,792 GB/s of bandwidth, but the 5090D V2 is cut down to 24GB of GDDR7 on a 384-bit bus with 1,344 GB/s. For gaming, the three cards are very close. For AI work, rendering, and local large language model inference, the D models fall clearly behind. The easiest mistake to make is being sold a 5090D V2 as a 5090D — the model names differ by just two characters, "V2", but the memory differs by 8GB. Before paying, check the full model name on the box, the label on the card, the receipt, and four fields in GPU-Z. Warranty terms are as announced by the shop; confirm them in writing before you buy.

Introduction: Why Does One Card Have Three Versions?

If you have been shopping for a flagship graphics card in Hong Kong recently, you have probably seen the name "5090" appear in three different forms: RTX 5090, RTX 5090D, and RTX 5090D V2. The names look almost identical, and all three sit at the flagship tier. In reality, they are three distinct products built for different markets. Buying without understanding the differences is risky. For gaming, the mistake may cost you little. For AI or rendering, buying the wrong one can be expensive.

The three cards are broken down one item at a time below, based on NVIDIA's official specifications, press reporting, and benchmark data. Every rumor is labelled as a rumor. Anything that could not be verified is stated as unverified, not presented as fact.

1. Head-to-Head Specification Comparison

The figures above come from NVIDIA's official RTX 5090 specification page, the V2 specifications listed on NVIDIA's official website (as reported by Wccftech and Guru3D), and 5090D specification tables compiled by HotHardware and ComputerCity. One correction worth stating plainly: the V2 uses GDDR7 memory, not GDDR6. One online report mistakenly wrote GDDR6, contradicting every other source, and that should be treated as a typo in that report.

One naming note: before launch, leaks called the V2 the "RTX 5090DD". Its confirmed official name is "RTX 5090D V2". So an older leak mentioning "5090DD" refers to the V2, not to a fourth card.

2. Why Three Versions? The Export-Control Background

The standard RTX 5090 is not officially sold in China. In an official statement dated 15 October 2025, MSI confirmed that it has never officially sold or distributed the standard RTX 5090 in China. In China, MSI offers only the GeForce RTX 5090D and GeForce RTX 5090D V2 — versions designed to meet local regulatory requirements. In other words, standard 5090 cards found in China arrived through unauthorised channels as parallel imports, not through official distribution.

The 5090D was created to comply with US export controls. Its AI performance was restricted from 3,352 AI TOPS down to 2,375 AI TOPS, a drop of about 29%, while the CUDA cores, clocks, and memory used for gaming were kept identical to the standard card. This is confirmed by the official specifications and consistent reporting across multiple outlets.

Why did the V2 then cut memory from 32GB to 24GB and bandwidth from 1,792 GB/s to 1,344 GB/s? Multiple media outlets reported that the cut was made to satisfy export-control-related limits, and some reports mentioned a "1.4 TB/s bandwidth ceiling". That must be stated carefully: it is media reporting, not a published rule. GamersNexus has pointed out that the US Bureau of Industry and Security (BIS) has issued no specific guidance on memory bandwidth. So the claim that "a 1.4 TB/s bandwidth cap is a BIS rule" is a rumor, not an established regulation.

How exactly is the AI restriction implemented — for example, whether it is a firmware lock on Tensor throughput or a hardware-level disabling of memory channels — is unverified. Different sources offer different speculation. The restriction mechanism is unconfirmed.

Two further rumors need both sides presented:

First, the discontinuation rumor. In early October 2026, reports circulated — originating from leakers including MEGAsizeGPU and picked up by several outlets — claiming NVIDIA had redirected all GB202 chips to RTX PRO professional cards and would no longer supply the GeForce RTX 5090 family, including the D and V2 models. NVIDIA Senior Vice President Jeff Fisher responded that the RTX 5090 is not discontinued, that supply is tight, but that shipments will continue. Fisher did not disclose shipment volumes or a timetable. Discontinuation is therefore unconfirmed and disputed. It should not be treated as discontinued.

Second, the reported China import ban on the V2. In May 2026, there were reports, attributed to HKEPC and relayed by several outlets, that Chinese customs had placed the V2 on a banned-imports list, refusing import or processing permits. But in June of the same year, Colorful still exhibited the V2 as a current product at Computex, and reporting quoted Colorful as saying the V2 was not banned in China. The reports conflict, the situation is unclear, and it should not be treated as a settled fact that the V2 has been banned in China.

3. The AI Performance Gap: 2,375 vs 3,352 TOPS

AI TOPS (trillions of operations per second) is the AI compute figure listed in NVIDIA's official specifications. The global RTX 5090 is rated at 3,352 AI TOPS. Both the 5090D and the 5090D V2 are rated at 2,375 AI TOPS, about 29% lower. This gap is confirmed and printed in the official specifications.

One important qualification: NVIDIA's official specifications list only the TOPS numbers. They do not publicly explain the restriction mechanism or say which frameworks are affected. No claim is made that any specific framework "does not work" on the D models, because that has not been verified.

What does a 29% theoretical gap mean in practice? The following results come from Expreview's benchmark testing, as reported by Wccftech, comparing the 5090D V2 against the 5090D:

• Local large language model (LLM) inference: the V2 was about 10% slower (test models included Llama 3.1 and Llama 2).
• Blender rendering: about 19% slower in the Junkshop scene, about 10.5% slower in Monster, and about 15.9% slower in Classroom.
• V-Ray rendering: about 7.5% slower in CUDA mode and about a fifth slower in RTX mode. 'About a fifth' is used below; the source text contains both 'about 20% slower' and a '24.8% gap' phrasing.

Note that these are results from a single testing lab, reported by two outlets — not a consensus across many labs. Benchmark data specifically measuring V2 AI training performance could not be found from a reliable source. No reliable data on the training gap.

If you are buying mainly for AI — local LLM inference, AI content creation, rendering — the D model already gives up about 29% of theoretical AI compute against the standard card, and the V2 adds a memory and bandwidth disadvantage on top. That is something to think through before choosing.

4. What the V2's 24GB / 384-bit Memory Really Means: Gaming vs AI

The V2's most consequential change is the memory cut: 32GB down to 24GB, bus width from 512-bit down to 384-bit, bandwidth from 1,792 GB/s down to 1,344 GB/s. How much does that matter? The answer is very different for gaming and for AI.

Gaming: Very Little Impact

According to Expreview's testing, as reported by Wccftech and PCGuide, the 5090D V2 and 5090D differ by only 1–2% in gaming and synthetic tests, with almost no difference in several games at native 4K resolution. The tested games included *Cyberpunk 2077*, *Monster Hunter Wilds*, *Marvel Rivals*, *Assassin's Creed Shadows*, and *Black Myth: Wukong*. 3DMark results (Fire Strike, Time Spy, Port Royal, Speed Way, Steel Nomad) were similarly close.

The conclusion: 24GB of memory remains enough for current games, even at 4K. One qualification: this conclusion rests on games available at the time of testing in August 2025. Future games with ultra-high-resolution texture packs may behave differently, and this should not be treated as a permanent guarantee.

What about the gaming gap between the 5090 and the 5090D? A precise figure is unverified — no directly comparable benchmark of equal quality could be found. Since the two cards share identical CUDA cores, clocks, and memory, gaming performance is expected to be very close, with the main difference coming from AI TOPS (which affects the theoretical ceiling of AI features such as DLSS). That is an inference, not a measured conclusion. No "zero difference" or percentage claim is made.

AI: A Big Impact, and a Hard Ceiling

Gaming differences are small. AI is a different story. The V2's 24GB is a hard ceiling, not merely "a little slower".

The following is general memory arithmetic, not V2-specific benchmark data:

• A 32B model at Q4_K_M quantisation has weights of about 20GB; a 70B model at Q4_K_M is about 40GB. Real use also needs KV cache and activation buffers, roughly another 15–20% for short contexts.
• Examples that fit in 32GB but not in 24GB: Qwen2.5 32B Q6_K (about 26–28GB), DeepSeek-R1 32B Q6_K (about 26–28GB), and Mixtral 8x7B Q4_K_M (about 26GB).
• A 24GB card running Qwen2.5-32B Q4_K_M is left with only about 4GB for KV cache, so large context sizes can easily run out of memory.

One point needs emphasis: "does not fit" does not mean "cannot run". Lower quantisation or CPU offload can make a model run, but more slowly. This matters because a buyer who gets a V2 by mistake for AI work may find that a model configuration they planned to use does not fit in memory at all — not that it runs 10% slower, but that it cannot run in the intended configuration.

In short: buy a V2 by mistake for gaming, and the impact is small. Buy a V2 by mistake for AI or rendering, and it is a clear downgrade.

5. How to Tell the Three Apart: Box, Model Name, and GPU-Z

The Correct Full Names

• `GeForce RTX 5090`
• `GeForce RTX 5090D`
• `GeForce RTX 5090D V2`

The key check: the box, the product listing, and the invoice or receipt must all show the complete model name. Writing just "5090" is not enough. "5090D" and "5090D V2" differ by two characters, but their memory differs by 8GB. "V2" is part of the model name and cannot be omitted.

As no side-by-side comparison of the three boxes is available, match the full model name on the box, the label on the card, and GPU-Z; do not judge by packaging appearance.

Another warning sign: the 5090D V2 has no Founders Edition. Anything sold as a "5090D V2 Founders Edition" is suspect.

Identifying with GPU-Z

GPU-Z is a free tool from TechPowerUp that shows a card's model name, GPU, memory, bus width, clocks, BIOS version, and more. To separate the three cards, check these fields:

Two points matter most:

First, 24GB / 384-bit means V2. That shows up immediately in GPU-Z. Second, the 5090 and 5090D have identical memory fields (32GB / 512-bit), so memory readings cannot separate the global card from the D model. You need the Name field plus the model name on the box. Any claim that "you can tell the D model from the global model by looking at memory" is wrong.

GPU-Z also has a fake-card detection feature: a fake card shows `[FAKE]` before the model name with a yellow exclamation mark. But be aware that GPU-Z readings can be falsified by a modified fake-card BIOS. GPU-Z is therefore a necessary but not sufficient check. It should be combined with real load testing (benchmarks, games, AI workloads) and with proof of purchase from the seller.

As no reliable source showing GPU-Z screenshots or a Device ID table for all three could be found, checking the four fields above is sufficient; no Device ID is needed.

Do Not Confuse These with Other Models

The RTX 5090 is not the same as the RTX 5090 Laptop GPU. The laptop version has completely different specifications, and desktop figures must not be applied to it. Laptop-version figures should be checked against the official specifications separately.

6. Consequences of Buying the Wrong One

Scenario 1: You Wanted a 5090 and Got a D Model

• Gaming experience: based on the confirmed specifications (identical CUDA cores, clocks, and memory), gaming performance is expected to be very close, and there should be no cliff-edge difference as a matter of inference. But the precise gap is unverified, so "zero difference" would be an overstatement.
• AI use: the roughly 29% lower AI TOPS is a confirmed, hard difference. If the card is for local LLM inference, AI creation, or rendering, the D model is clearly at a disadvantage.
• Origin and compliance: the standard 5090 is subject to export controls, and standard cards found in China are parallel imports (confirmed by MSI's statement). If you are offered a card described as a "5090" in the second-hand market, verify which version it actually is and which channel it came through; do not rely on verbal assurances. In the US, one retailer (Micro Center) has required some 5090 buyers to show photo ID and sign a no-export pledge. That is a retailer-level measure, not NVIDIA official policy, but it illustrates that the global card carries export-compliance risk.
• Resale value: unverified. No reliable source quantifying the second-hand price difference between the D model and the standard card could be found. No reliable data on the resale price gap.

Scenario 2: You Wanted a 5090D (32GB) and Got a V2 (24GB) — the Easiest Mistake

This is the easiest trap to fall into, because the model name differs by only "V2", and a seller may omit "V2" deliberately or accidentally.

• Gaming: measured difference of only 1–2%. Buying a V2 by mistake for gaming has a small impact.
• AI / productivity: Blender up to about 19% slower, V-Ray RTX about a fifth slower, LLM inference about 10% slower. For AI or rendering buyers, the V2 is a clear downgrade.
• Memory ceiling: 24GB cannot hold model configurations that 32GB can (for example, 32B models at Q6_K quantisation). This is a hard ceiling, not "a little slower".
• Second-hand resale: buyers will check memory and bus width in GPU-Z, and the V2's 24GB / 384-bit shows up immediately. A seller passing off a V2 as a 5090D is misleading the buyer.

Scenario 3: Warranty Risk

The D and V2 models are China-market products. What warranty applies when a D or V2 card is bought in Hong Kong is unverified — no public warranty terms from the brands for D / V2 cards in Hong Kong could be found. No public warranty terms for Hong Kong are available.

Some general principles are confirmed, however:

• Warranty service is generally provided in the region of purchase. If you import a product from a different region, you may have to return it to the region of purchase for warranty service (ASUS official forum staff reply).
• MSI's official statement says that products bought through unofficial channels are not eligible for MSI warranty, repair services, or technical support in China, and that unofficial-channel products are unverified in source and handling and are not covered by MSI's quality-control or warranty policies. The situation in Hong Kong is as announced by each brand's Hong Kong agent.
• Warranty claims usually require the purchase receipt as proof of purchase date. Keep your receipt.
• A serial number that has been altered or removed, a damaged warranty seal, or unauthorised disassembly or repair may void the warranty.
• A shop's own "shop warranty" is a different thing from a manufacturer's warranty, and its terms are as announced by the shop.

Specific warranty durations and terms are as announced by the shop or agent.

A Note on Prices

On prices, only two confirmed official suggested prices are listed: the global RTX 5090 official suggested price of US$1,999 (NVIDIA's official suggested price), and the RTX 5090D V2 China official suggested price of ¥16,499 (the suggested price at its China launch on 12 August 2025 — not a Hong Kong street price). No other prices are stated. Actual selling prices are as announced by the shop; ask the shop before buying. Price-comparison-site listings, out-of-stock prices, and private-seller prices must not be treated as in-stock prices.

7. Pre-Purchase Checklist

Go through these one by one before paying. Do not be embarrassed to take your time:

1. Match the full model name: the box, the label on the card, and the receipt must all agree, and must show the complete name — "5090", "5090D", or "5090D V2". If the receipt says only "5090" with no suffix, ask for clarification on the spot.
2. Watch for a missing "V2": the product title, the receipt, and any verbal promise must all include "V2" for the card to be a V2. If "V2" is missing, do not assume the card is a V2 — and do not assume it is a 5090D either.
3. Ask to boot up and run GPU-Z: check Name, Memory Size, Bus Width, and Bandwidth. Take a screenshot and keep it. 24GB / 384-bit means V2. 32GB / 512-bit means you must then check the Name field to distinguish "5090" from "5090D". If the seller refuses verification, treat that as a warning sign.
4. Check the serial number and seals: look for any sign the serial number has been altered or scratched off, check that seals are intact, and keep the box, receipt, and warranty card.
5. Ask clearly about origin and warranty: which channel did the card come through? Who provides the warranty, in which region, and with what documents? Specific terms are as announced by the shop — get verbal promises in writing.
6. Ask about return arrangements: if the card you verify does not match the receipt, what happens? Settle this before buying, not afterwards.
7. Plan for power and case space: all three cards have a 575W total graphics power rating, and the power connector is a single 16-pin (12V-2x6). Power-supply recommendations vary by individual model — follow the manufacturer's recommendation for the specific model, as the Founders Edition and each partner model may recommend different ratings. Do not apply one number to all models. The 12V-2x6 connector must be fully and firmly seated, without excessive bending, and the case needs adequate airflow and cable-bend clearance.
8. Do not trust a single source for prices: for any "market price" claim, ask which channel it refers to and when it was quoted. Actual selling prices are as announced by the shop.

8. Frequently Asked Questions (FAQ)

Q: How much slower is the 5090D than the 5090 in games?

A: A precise figure is unverified — no directly comparable benchmark of equal quality could be found. Since the two share identical CUDA cores, clocks, and memory, gaming performance is expected to be very close, with the main difference coming from AI TOPS (which affects the theoretical ceiling of AI features such as DLSS). That is an inference, not a measured conclusion.

Q: Is the 5090D V2 much worse for gaming?

A: No. According to Expreview's testing, as reported by Wccftech and PCGuide, the V2 and the D model differ by only 1–2% in gaming tests, with almost no difference in several games at native 4K.

Q: Why is the V2 barely slower in games despite having less memory?

A: Because gaming performance depends mainly on CUDA cores and clocks, which are identical between the V2 and the D model. Memory capacity and bandwidth have not become the bottleneck in current games. This explanation is an inference based on the confirmed specifications.

Q: Which model should I buy for AI?

A: On official specifications, the global 5090 has the highest AI TOPS (3,352) and the largest memory (32GB). The D model is about 29% lower in AI TOPS, and the V2 adds a 24GB memory ceiling on top. If you are buying for local LLM inference or rendering, check how much memory the models you intend to run actually need before deciding. Actual selling prices are as announced by the shop.

Q: How can I tell a V2 at a glance?

A: In GPU-Z, look at Memory Size and Bus Width: 24GB / 384-bit means V2. The box and receipt must also show the complete model name "RTX 5090D V2".

Q: Does the 5090D V2 have a Founders Edition?

A: No. Anything sold as a "5090D V2 Founders Edition" is suspect.

Q: If I buy a D or V2 model in Hong Kong, is there a Hong Kong warranty?

A: Unverified. No public warranty terms from the brands for D / V2 cards in Hong Kong could be found. Before buying, get written confirmation of the warranty region and terms from the seller or the brand. Specific terms are as announced by the shop or agent.

Q: Has the 5090 been discontinued?

A: There are reports claiming GB202 chips were redirected to RTX PRO cards and the 5090 family would no longer be supplied, but NVIDIA Senior Vice President Jeff Fisher responded that the card is not discontinued, that supply is tight, and that shipments will continue. Discontinuation is unconfirmed and disputed; it should not be treated as fact.

Q: Has the V2 been banned in China?

A: The situation is unclear. In May 2026 there were reports, attributed to HKEPC, that the V2 had been placed on a Chinese customs banned-imports list; but in June of the same year, a manufacturer still exhibited it as a current product. It should not be treated as a settled fact that it has been banned.

Q: Is the 5090DD a fourth card?

A: No. "5090DD" was the leak name for the V2 before launch. Its confirmed official name is "RTX 5090D V2".

Q: Is checking with GPU-Z enough?

A: No. GPU-Z is a necessary but not sufficient check, because its displayed values can be falsified by a modified fake-card BIOS. Combine it with real load testing (benchmarks, games, AI workloads) and proof of purchase from the seller.

9. Conclusion

The relationship between the three cards can be summed up in one sentence: the gaming core specifications are identical; the differences are in AI compute and memory. The global RTX 5090 is the most complete specification. The 5090D keeps the gaming specifications and 32GB of memory but gives up about 29% of AI TOPS. The 5090D V2 cuts memory further to 24GB and bandwidth to 1,344 GB/s — a 1–2% difference in gaming, but roughly 10–25% slower than the D model in AI and rendering, and some large-model configurations simply will not fit in 24GB.

The three most important things before buying: first, check the complete model name, especially the two characters "V2". Second, boot up and run GPU-Z, and check Name, memory, bus width, and bandwidth. Third, get the origin and warranty terms in writing — specific terms are as announced by the shop. Do those three things, and a model name missing two characters will not cost you 8GB of memory.

Sources

• NVIDIA official RTX 5090 specifications: https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
• NVIDIA official press release (5090, 92B transistors, 3,352 TOPS): https://nvidianews.nvidia.com/_gallery/download_pdf/677c9a9c3d6332e22e6c8571/
• HotHardware (5090D official page, 2,375 TOPS): https://hothardware.com/news/nvidia-geforce-rtx-5090d-china-crypto-ai-locks
• ComputerCity (5090D specification table): https://computercity.com/hardware/video-cards/geforce-rtx-5090-d-specs
• Wccftech (V2 official launch, NVIDIA website specifications): https://wccftech.com/nvidia-rtx-5090-d-v2-launched-officially-at-16499-yuan-gets-listed-on-hong-kong-retailer-for-cheaper-than-msrp/
• Wccftech (V2 launch, three-model comparison table): https://wccftech.com/nvidia-geforce-rtx-5090-d-v2-gpu-launch-china-12th-august-21k-cores-24-gb-memory/
• Guru3D (V2 launch date, GB202-240, 24GB / 384-bit): https://www.guru3d.com/story/nvidia-geforce-rtx-5090-d-v2-launch-date-confirmed-for-china-in-august-2025/
• Guru3D (V2 official release, 680 Tensor / 170 RT, benchmark deltas): https://www.guru3d.com/story/geforce-rtx-5090-d-v2-officially-released-specs-performance-and-pricing/
• Wccftech (Expreview benchmarks: gaming 1–2%, Blender, V-Ray, LLM): https://wccftech.com/china-exclusive-rtx-5090-d-v2-benchmarks-reveal-2-slower-performance-than-rtx-5090-d/
• PCGuide (V2 comparison table, Expreview benchmarks): https://www.pcguide.com/news/chinas-new-rtx-5090-with-less-vram-performs-pretty-much-the-same-for-4k-gaming-benchmarks-reveal/
• Hardware Busters (V2 launch, benchmark deltas, AI TOPS): https://hwbusters.com/news/nvidia-released-the-geforce-rtx-5090d-v2/
• TechSpot (MSI statement: China sells only D / V2; standard cards are parallel imports): https://www.techspot.com/news/109868-msi-pallets-rtx-5090-cards-seen-china-came.html
• MSI official statement (global card not officially sold in China; unofficial-channel warranty position): https://www.msi.com/news/detail/MSI-Clarifies--GeForce-RTX-5090-Is-Not-Officially-Sold-in-China--Non-Authorized-Units-Identified-as-Parallel-Imports-147226
• GamersNexus (no specific BIS memory-bandwidth guidance): https://gamersnexus.net/gpus-deep-dive-news/nvidia-ai-gpu-black-market-investigating-smuggling-corruption-governments
• Tech4Gamers (China V2 import-ban report, citing HKEPC): https://tech4gamers.com/china-reportedly-blocks-entry-of-nvidia-geforce-rtx-5090d-v2-graphics-cards/
• DigitalCitizen (Colorful says V2 is not banned in China): https://www.digitalcitizen.life/colorful-says-rtx-5090-d-v2-is-not-banned-in-china-as-it-shows-new-arcanix-gpu-and-white-x870e-vulcan-board/
• VideoCardz (NVIDIA says RTX 5090 is not discontinued): https://videocardz.com/newz/nvidia-says-rtx-5090-is-not-discontinued-but-good-luck-finding-one-for-1999
• ComputingForGeeks (local LLM memory arithmetic): https://computingforgeeks.com/build-local-ai-workstation-llm/
• InsiderLLM (models that fit in 24GB vs 32GB): https://insiderllm.com/guides/rtx-5090-local-ai-worth-it/
• HEXUS (GPU-Z fake-card detection, [FAKE] + yellow exclamation mark): https://hexus.net/tech/news/software/123200-gpu-z-version-2120-flags-fake-nvidia-gpus/

常見問題

關於Comparison選購嘅常見疑問，歡迎致電 +852 2186 8010 查詢，或親臨新界沙田石門達利廣場1209室門市。

相關文章

• RTX 5060 Ti 8GB vs 16GB: Which One Should You Buy? Gaming and AI Compared Comparison2026-10-18
• RTX 5070 vs RTX 5070 Ti: Is the Price Difference Worth It? Performance, Price, Power and Full Comparison Comparison2026-10-17
• 2026 香港顯示卡購買指南：5 步揀啱你嘅 GPU 購買指南2026-10-13

本文由 Display Card King Editorial 撰寫，報價數據來自 Display Card King 每日全港監測，只供參考。