RTX 50 Super and the RTX 5080 24GB: what is known, what is rumored, and what 24 GB would run

As of 7 November 2026, NVIDIA has announced no RTX 50 Super card. The newest report, from Taiwan's BenchLife via VideoCardz on 2 November 2026, describes a 24 GB RTX 5080 planned for Q1 2027 that would replace the 16 GB model, but says board partners had no design kits by the end of October. The reason for the long delay, per a VideoCardz source in July 2026, is memory: a 3 GB GDDR7 chip cost $60 to $70 against about $20 for a 2 GB chip. By my calculation, a 24 GB card would load a 27B dense model at 4-bit with about 119,000 tokens of FP16 KV cache. A 16 GB card cannot load it at all.

Rumors are labeled with the outlet that published them and its original source. Nothing in the "rumor" rows has been confirmed by NVIDIA.

Has NVIDIA announced the RTX 50 Super?

No. NVIDIA has said nothing official about a Super refresh, and the dates in the press have moved four times in a year. Here is the trail, oldest first.

• 10 November 2025. After Uniko's Hardware claimed the series was cancelled, TweakTown relayed HKEPC's report that NVIDIA told partners it was postponed to Q3 2026. Leaker MEGAsizeGPU said the same. BenchLife added that no partner had design information yet.
• 13 January 2026. The cards missed CES 2026. TweakTown cited the Chinese Board Channels forum: delayed indefinitely, because of memory costs, AI demand for GPUs and no new AMD competition expected in 2026.
• 5 June 2026. MEGAsizeGPU posted that the refresh was "back on track" and now included a 12 GB RTX 5060, per VideoCardz. He gave no window beyond "2026".
• 17 July 2026. VideoCardz reported that at least one partner physically had RTX 5080 Super, 5070 Ti Super and 5070 Super cards, and that NVIDIA had told partners to hold them over the price of 3 GB GDDR7.
• 12 October 2026. BenchLife, via VideoCardz, reported a 24 GB RTX 5080 for Q1 2027 that replaces the 16 GB card. Leakers hongxing2020 and MEGAsizeGPU claimed the RTX 5090 would be discontinued too. NVIDIA SVP Jeff Fisher denied that on 17 October, telling 4Gamers the 5090 "has certainly not been discontinued" while supply is "somewhat constrained" (VideoCardz translation).
• 2 November 2026. BenchLife says the 24 GB RTX 5080 design kits were due in October and had not arrived. VideoCardz notes CES 2027 runs 6 to 9 January and an announcement there would not need cards on shelves.

Neither outlet knows whether "RTX 5080 24GB" is the old RTX 5080 Super under a new name. TechSpot's Q4 2026 pricing update repeats the Q1 2027 rumor and adds that retail sources report rising wholesale costs for the current 16 GB card.

Why do 3 GB GDDR7 chips decide the RTX 50 Super?

Because the whole refresh is a memory swap: same GPUs, same buses, denser chips. Each GDDR7 chip sits on a 32-bit channel, so the RTX 5080's 256-bit bus (NVIDIA spec page) carries eight chips. Eight 2 GB chips give 16 GB. Eight 3 GB chips give 24 GB. The 192-bit RTX 5070 goes from 12 GB to 18 GB the same way.

The chips exist. Samsung announced the first 24-gigabit (3 GB) GDDR7 on 17 October 2024, at 40 Gbps, aimed at data centers and AI workstations as well as graphics cards. A 24 GB GB203 also exists already: the RTX 5090 Laptop GPU uses the GB203 chip with 24 GB on a 256-bit bus (Notebookcheck), which by the same arithmetic means eight 3 GB chips. So does a workstation card, the RTX PRO 4000 Blackwell, with 24 GB on a narrower 192-bit bus at 672 GB/s and 140 W (Geizhals spec listing).

What changed is price. Using the chip prices VideoCardz's source gave in July 2026:

16 GB card:  8 chips x $20      = $160 of GDDR7  ($10.0 per GB)
24 GB card:  8 chips x $60-70   = $480-560       ($20.0-23.3 per GB)
extra cost per card                = $320-400 before any margin

That is my arithmetic on reported prices, not a bill of materials. TrendForce's 30 September 2026 forecast says graphics DRAM prices keep rising into Q4 2026 because GDDR6 supply is shrinking and GDDR7 output stays limited. The wider memory market is covered in the 2026 memory shortage article.

What would a 24 GB RTX 5080 run that a 16 GB card cannot?

Models in the 25B to 35B range at 4-bit, with long context. The table shows how much context each model has room for after its weights. I used the GGUF file sizes on Hugging Face, the attention layout from each model's config.json, and two assumptions: 1.5 GB of runtime overhead, and marketing GB meaning GiB (24 GiB is 25.77 GB).

KV bytes per token = full-attention layers x 2 (K and V) x KV heads x head dim x 2 bytes (FP16)
Qwen3.8-27B: 16 of its 64 layers use full attention, 4 KV heads, head dim 256
             16 x 2 x 4 x 256 x 2 = 65,536 bytes = 64 KiB per token
max context  = (VRAM - 1.5 GB - weights) / KV bytes per token
24 GiB: (25.77 - 1.5 - 16.46) GB / 65,536 B = about 119,000 tokens

Weights: GGUF files in the unsloth repos on Hugging Face (Qwen3.8-27B example), checked 7 November 2026. Layer counts and KV heads from each model's config.json (Qwen3.8-27B). q8_0 KV counted at 1.0625 bytes per value. Gemma 4's config.json sets attention_k_eq_v and uses different head sizes for its global layers, so I left its KV figure out rather than guess. All rows are my calculations, not measurements.

At 16 GB the only modern model in this list with room for real context is gpt-oss-20b. At 24 GB the 27B and 35B models become usable for long documents, while dense 32B stays cramped because its KV cache is four times larger per token than Qwen3.8's hybrid layout. For per-model detail on 16 GB cards, see the best LLMs for 16 GB of VRAM. For quantization formats, see the quantization guide.

How fast would a 24 GB RTX 5080 be for LLMs?

About as fast as a 16 GB RTX 5080 on the same model, which means fast enough for one user. Decode speed is capped by memory bandwidth divided by the bytes read per token. The RTX 5080 has 960 GB/s (30 Gbps on 256 bits, per Geizhals' spec listing). The rumored 32 Gbps memory would give 1,024 GB/s, 6.7% more.

ceiling tok/s = bandwidth / weight bytes per token (dense model)
Qwen3.8-27B Q4_K_M: 1,024 GB/s / 16.46 GB = 62 tok/s ceiling

The nearest measured hardware is the RTX 5090 Laptop GPU, the same GB203 chip with 24 GB at 896 GB/s. In a llama.cpp bug report dated 11 September 2026 (issue #28790), its owner logged Qwen3.8-27B in NVFP4 (a 15.5 GB file) at 39.3 tok/s decode without speculative decoding and 50.3 to 51.7 tok/s with MTP on official builds b10917 and b10889, at a 152,000-token context with q8_0 KV. That 39.3 tok/s is 68% of the laptop's bandwidth ceiling. A desktop card with more power and bandwidth should land above it; that is an inference, not a measurement.

What would it cost?

Expect well above the rumored $999 to $1,199. That range comes from leaks summarized in November 2025, before the memory price jump. Since then the 16 GB RTX 5080 has climbed: TechSpot found the cheapest in-stock US card at $1,400 in February 2026 and $2,200 in October 2026, 120% over its $999 MSRP, with Germany at €2,100 (TechSpot, 2 November 2026). On 7 November 2026 Geizhals listed it from €1,849.99 including VAT.

TechSpot adds a theory, which it labels as such: price rises on the 16 GB card could reset the market so that a 24 GB model's launch price looks acceptable. Add the $320 to $400 of extra memory cost from the calculation above and my estimate is that a 24 GB RTX 5080 will not sell below today's 16 GB street price. If you need 32 GB instead, the RTX 5090 is the next step up, and it is in worse shape: $9,400 at its cheapest in the US in October per TechSpot, and $5,599.99 for the cheapest in-stock listing on PCPartPicker when I checked on 7 November 2026.

Should you buy now or wait for the RTX 50 Super for local AI?

Do not wait if your models already fit in 16 GB. Wait only if you specifically want 24 GB on a desktop card under about €2,500 and can live with the risk that it slips again or launches higher.

• Your models fit in 16 GB today (gpt-oss-20b, 8B to 14B models): buy the card you need now. A Super adds capacity, not speed.
• You want 24 GB now: two new GB203 products have it. The RTX PRO 4000 Blackwell desktop card listed from €2,239.04 on Geizhals on 7 November 2026, with ECC but only 672 GB/s, so expect roughly 70% of a 5080's decode speed by the bandwidth ratio. The RTX 5090 Laptop GPU has 896 GB/s; its measured speeds and prices are in the RTX 5090 laptop article.
• You want 32 GB or more: a desktop RTX 5090 is expensive and scarce, and the DGX Spark vs RTX 5090 comparison covers when 128 GB of slower memory is the better trade. The memory calculator checks a specific model.
• You only need the capacity for a project: rent. GPUwerk rents a 32 GB RTX 5090 at $0.89 per hour and a 128 GB DGX Spark at $0.79 per hour, billed per minute (pricing).

My pick for one person running local models in November 2026: buy nothing on a Super rumor. If 16 GB is too small for the model you actually use, test that model on rented 32 GB hardware first, then decide after CES 2027 whether a 24 GB RTX 5080 exists and what it costs.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• NVIDIA RTX 50 SUPER launch on hold over 3GB GDDR7 pricing, VideoCardz, 2026-07-17
• NVIDIA RTX 5080 24GB reportedly launching in Q1 2027, board partners still waiting for details, VideoCardz, 2026-10-12
• NVIDIA reportedly hasn't provided RTX 5080 24GB design kits to board partners yet, VideoCardz, 2026-11-02
• NVIDIA GeForce RTX 50 SUPER is 'back on track', RTX 5060 12GB reportedly planned, VideoCardz, 2026-06-05
• NVIDIA says RTX 5090 is not discontinued, but good luck finding one for $1,999, VideoCardz, 2026-10-17
• NVIDIA told AIB partners it has NOT canceled RTX 50 SUPER plans, but postponed until Q3 2026, TweakTown, 2025-11-10
• GeForce RTX 50 SUPER Series 'delayed indefinitely' as NVIDIA informs its partners, TweakTown, 2026-01-13
• GPU Prices Are Spiraling Out of Control: Q4 2026 Update, TechSpot, 2026-11-02
• Samsung Develops Industry's First 24Gb GDDR7 DRAM for Next-Generation AI Computing, Samsung Newsroom, 2024-10-17
• AI Server Demand Sustains Memory Contract Price Increases in 4Q26, While Consumer-Side Pressure Persists, Says TrendForce, TrendForce, 2026-09-30
• GeForce RTX 5080 specifications, NVIDIA
• INNO3D GeForce RTX 5080 X3 OC White price comparison, Geizhals
• Lenovo RTX PRO 4000 Blackwell 24GB price comparison, Geizhals
• Eval bug: MTP causes ~57x prefill slowdown on a self-compiled Windows build (RTX 5090 Laptop measurements), ggml-org/llama.cpp on GitHub, 2026-09-11
• Nvidia GeForce RTX 5090 Laptop GPU benchmarks and specs, Notebookcheck
• GeForce RTX 5090 video cards sorted by price, PCPartPicker
• Qwen3.8-27B config.json, Hugging Face
• unsloth/Qwen3.8-27B-GGUF file listing, Hugging Face