RTX 5090 laptop for LLMs: measured tok/s, power modes, what fits in 24 GB and what it costs

The RTX 5090 Laptop GPU is a GB203 chip with 24 GB of GDDR7 at 896 GB/s and a 95 to 150 W power limit (175 W with Dynamic Boost), so it has three quarters of the desktop card's memory and half its bandwidth. In llama.cpp's public scoreboard it decodes Llama 2 7B Q4_0 at 158 tok/s in a 175 W mode and 117 tok/s in a 95 W mode, against 300 tok/s on a desktop RTX 5090. A Qwen3.8-27B NVFP4 build ran at 39 tok/s, or 50 with multi-token prediction, in a September 2026 report. On 7 November 2026 Geizhals listed RTX 5090 laptops from €3,509, less than the cheapest desktop RTX 5090 card at €6,249.

What is the RTX 5090 Laptop GPU?

It is a desktop RTX 5080 chip with 24 GB of memory and a laptop power budget. Notebookcheck lists the GB203 with 10,496 of its 10,752 shaders enabled, GDDR7 at 28 Gbps on a 256-bit bus, and availability from April 2025. The desktop RTX 5090 uses the bigger GB202 with a 512-bit bus.

Laptop: NVIDIA's 50 Series laptop page and Notebookcheck. Desktop cards: Geizhals spec listings for the RTX 5090 and RTX 5080, read 7 November 2026. The laptop maker sets the power limit inside NVIDIA's range, so two "RTX 5090" laptops can differ.

How fast is an RTX 5090 laptop for LLMs?

About half as fast as a desktop RTX 5090, which matches the bandwidth ratio. All figures below are owner reports on GitHub; GPUwerk has published no RTX 5090 runs of its own.

The desktop RTX 5090 entry in the same CUDA scoreboard reads 14,970 tok/s prefill and 300.4 tok/s decode on Llama 2 7B Q4_0 with FA on. The 83.9 tok/s MTP figure depends on draft acceptance for the prompts used and is not comparable to plain decode.

The ratio is consistent. Decode on the 7B model is 53% of the desktop card's, prefill 51%, and bandwidth 50% (896 against 1,792 GB/s). For the 27B model, the bandwidth ceiling is 896 GB/s divided by a 15.5 GB file, about 58 tok/s; the measured 39.3 tok/s is 68% of it, a normal efficiency for llama.cpp. For what these models are and how they are configured, see the Qwen3.8-27B page.

How much does the power mode change LLM speed?

Enough to notice, and more for prompts than for generation. The same owner measured decode 26% slower and prefill 34% slower in the laptop's 95 W mode than in its 175 W mode. Prefill is compute work, so it tracks clock speed and power. Decode mostly waits on memory, which does not slow down as much.

Notebookcheck's 3DMark Speed Way chart shows the same slope for graphics: 87% of full performance at 135 W, 96% at 150 W, 100% at 175 W. Heat was not the limit in the one long-running LLM log I found. In llama.cpp issue #27330 (18 August 2026), an HP Omen Max 16 serving models drew 167 W of its 175 W limit at 82 °C, under the 87 °C target, with zero hardware thermal-slowdown time. That owner's actual bug was a CUDA graphs hang, worked around with GGML_CUDA_DISABLE_GRAPHS=1, which is worth knowing if you run long sessions on sm_120.

In practice: plug in, pick the vendor's performance profile, and check the power limit with nvidia-smi -q -d POWER before you compare numbers with anyone else.

What fits in 24 GB on an RTX 5090 laptop?

Models up to about 35B parameters at 4-bit with long context; not 70B, and not gpt-oss-120b. I added the weight file size, the KV cache for the context, and 1.5 GB of overhead (an assumption), and compared it with 24 GiB (25.77 GB).

KV bytes per token = full-attention layers x 2 x KV heads x head dim x bytes per value
Qwen3.6-35B-A3B: 10 full-attention layers of 40, 2 KV heads, head dim 256
                 10 x 2 x 2 x 256 x 2 = 20,480 bytes per token at FP16
at 128k tokens:  20,480 x 131,072 = 2.68 GB
need = 22.13 GB (UD-Q4_K_M) + 2.68 GB + 1.5 GB = 26.31 GB  -> too big at FP16 KV
with q8_0 KV (1.0625 bytes per value): 25.06 GB            -> fits, barely

Weights from the GGUF listings on Hugging Face (unsloth and ggml-org repos, for example Qwen3.8-27B), checked 7 November 2026. Layer and head counts from each model's config.json (Qwen3.6-35B-A3B). gpt-oss models use 128-token sliding windows on half their layers, which I counted as negligible. All sums are my calculations.

Bigger MoE models can run with their experts in system RAM, and one owner logged Qwen3.6-35B-A3B at about 30 tok/s that way with --cpu-moe on a laptop RTX 5090 (PR #23170). That needs lots of fast system memory, and the cheapest RTX 5090 laptops on Geizhals ship with 16 GB. The MoE CPU offload article covers the RAM math. To check a specific model, use the memory calculator.

RTX 5090 laptop vs desktop RTX 5090, MacBook Pro M5 Max and Strix Halo

On models that fit in 24 GB, the laptop RTX 5090 is the fastest laptop here. The Apple and AMD machines win on capacity. The same llama-bench test (Llama 2 7B, Q4_0) exists for all of them in llama.cpp's scoreboards, which makes a fair speed comparison, though the backends differ.

Speeds: CUDA scoreboard (FA on), Apple Silicon scoreboard (Q4_0 columns, commit c1d0e7a), Vulkan scoreboard (FA on), read 7 November 2026. Bandwidth: NVIDIA and Apple. Prices incl. 19% VAT from Geizhals searches for M5 Max 128 GB, Ryzen AI Max+ 395 128 GB and RTX Spark 128 GB. The DGX Spark price is the Founders Edition's lowest listing.

Three things stand out. The laptop RTX 5090 decodes 32% faster than an M5 Max and prefills 2.4 times faster, which matters for long prompts. The M5 Max holds more than five times as much memory, 128 GB against 24 GB. And NVIDIA's own RTX Spark laptops, with up to 128 GB of LPDDR5X and a 6,144-core Blackwell GPU, launched with events on 16 October; I found no llama-bench entry for them in the scoreboards I read, and NVIDIA lists no bandwidth. Desktop comparisons are on the DGX Spark vs Mac Studio and DGX Spark vs Strix Halo pages, and the Mac vs RTX 5090 article covers the desktop card against Apple.

What does an RTX 5090 laptop cost?

From €3,509 including VAT in Germany on 7 November 2026, which is odd but true: the whole laptop costs less than the cheapest desktop RTX 5090 card listed the same day. One plausible reason, which I have not confirmed: the desktop card needs the large GB202 chip, which NVIDIA also sells as RTX PRO cards, while the laptop uses the smaller GB203.

Source: Geizhals notebook search for "rtx 5090 laptop", sorted by price, 272 products, checked 7 November 2026. Prices include 19% VAT and exclude shipping; Geizhals warns prices may have risen since its last update. I did not check each offer's stock.

Should you buy an RTX 5090 laptop for local LLMs?

Buy one if you need a portable machine for models up to about 30B parameters and you will run them often. It is the fastest laptop option for that size, and at €3,509 it costs less than any desktop RTX 5090 card. The only cheaper new 24 GB NVIDIA GPU I found is the RTX PRO 4000 Blackwell desktop card, from €2,239.04 on Geizhals, with 672 GB/s against the laptop's 896 GB/s. Plan for 32 GB or more of system RAM if you want to try MoE offload.

Skip it if your target is a 70B dense model or a 100B-plus MoE: 24 GB will not hold them, and an M5 Max or Strix Halo laptop with 128 GB will, at lower speed. Skip it too if the machine will sit on a desk running a server; a desktop card or rented capacity does that without a 175 W laptop fan profile. GPUwerk rents a 32 GB desktop RTX 5090 at $0.89 per hour (RTX 5090 page), which is a cheap way to check whether a model you care about is fast enough at 24 to 32 GB before buying either. If a 24 GB desktop card is what you are waiting for, the RTX 50 Super article tracks those rumors. The GB10-class RTX Spark N1X laptops are compared with a desktop DGX Spark in DGX Spark vs N1X laptop.

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.

• GeForce RTX 50 Series Laptops, NVIDIA
• Nvidia GeForce RTX 5090 Laptop GPU benchmarks and specs, Notebookcheck
• Performance of llama.cpp on Nvidia CUDA (scoreboard), ggml-org/llama.cpp on GitHub
• RTX 5090 Laptop llama-bench at 95 W and 175 W, ggml-org/llama.cpp on GitHub, 2026-05-26
• HP Omen Max 16 RTX 5090 Laptop llama-bench on Fedora 44, ggml-org/llama.cpp on GitHub, 2026-09-16
• Eval bug: MTP causes ~57x prefill slowdown on a self-compiled Windows build, ggml-org/llama.cpp on GitHub, 2026-09-11
• Research: FR-Spec-style draft-vocab trimming for native MTP speculative decoding, ggml-org/llama.cpp on GitHub, 2026-07-01
• Eval bug: CUDA graphs hang the GPU channel on RTX 5090 Laptop / sm_120, ggml-org/llama.cpp on GitHub, 2026-08-18
• ggml: treat experts as cache residents during MoE offloading, ggml-org/llama.cpp on GitHub, 2026-05-16
• Performance of llama.cpp on Apple Silicon M-series, ggml-org/llama.cpp on GitHub
• Performance of llama.cpp with Vulkan, ggml-org/llama.cpp on GitHub
• MacBook Pro technical specifications, Apple
• NVIDIA RTX Spark, NVIDIA
• rtx 5090 laptop in Notebooks, sorted by price, Geizhals
• macbook pro m5 max 128GB in Notebooks, Geizhals
• ryzen ai max+ 395 128GB in Notebooks, Geizhals
• RTX Spark 128GB in Notebooks, Geizhals
• rtx 5090 in Grafikkarten PCIe, sorted by price, Geizhals
• Lenovo RTX PRO 4000 Blackwell 24GB price comparison, Geizhals
• unsloth/Qwen3.8-27B-GGUF file listing, Hugging Face
• Qwen3.6-35B-A3B config.json, Hugging Face