RTX 5090 vs RTX Spark for Local LLMs: 32 GB GDDR7 vs 128 GB Unified (2026)

The two machines solve different problems. The RTX 5090 has fast memory but not much of it. The RTX Spark has a lot of memory, but it is much slower. Every benchmark in this guide follows from that. As of October 2026, the RTX Spark N1X ships in the Microsoft Surface Laptop Ultra, which went on sale on October 16, and in laptops from ASUS, Lenovo, Dell and others. NVIDIA calls it the first RTX platform with up to 128 GB of unified memory shared between an Arm CPU and a Blackwell GPU (NVIDIA blog).

Price matters here, and the starting price is misleading. Per Tom's Hardware's Surface Laptop Ultra review, the Surface starts at $2,599.99. That buys the cut-down 18-core chip with 24 GB of memory. The 20-core chip with 128 GB costs $5,899.99. So a reader with about $2,500-$3,000 is really choosing between an RTX 5090 desktop build and an entry RTX Spark laptop with less memory than the card. The 128 GB configuration, which is the reason to consider RTX Spark for local LLMs at all, costs about three times the RTX 5090's $1,999 launch MSRP.

This synthesis lines up the published per-model throughput for both platforms, works out which quantizations fit where, and ends with one pick per reader profile. Every figure links to its source, and where no source has published a run, the table says so.

Key Takeaways

• Bandwidth gap of about 6x on paper. The RTX 5090 has roughly 1,792 GB/s on a 512-bit GDDR7 bus (TechPowerUp). RTX Spark has up to 300 GB/s of LPDDR5X-9400 on a 256-bit bus (Wccftech).
• Capacity gap of 4x. The 20-core N1X takes up to 128 GB, and the GPU can be allocated up to 96 GB of it (StorageReview). The RTX 5090 has 32 GB.
• Same model, different speed. gpt-oss-20b runs at 298.2 tok/s on the RTX 5090 (Hardware Corner) and at 68.86 tok/s on a 64 GB RTX Spark laptop (Wccftech), a gap of about 4.3x.
• The 18-core chip is not a 128 GB machine. Per Tom's Hardware's launch report, the 18-core / 5,120-core N1X comes with 24-32 GB. Only the 20-core part goes to 128 GB.
• Power envelopes differ by an order of magnitude. The RTX 5090 board is rated at 575 W (TechPowerUp). StorageReview lists a 45-80 W laptop TDP for the N1X, covering CPU and GPU together.

Step 0: Which model sizes do you actually run?

Start with your models, not with the hardware. Decode speed depends mostly on how fast the weights can be read from memory, and whether a model runs at all depends on whether the weights plus the KV cache fit. That gives a three-way split:

1. Up to about 32B at Q4 (weights under ~20 GB): the RTX 5090 wins clearly. Hardware Corner measured dense Qwen3 32B at Q4_K_XL using 18.64 GiB of VRAM (Hardware Corner), with room left for long context. Every model in this range fits on the card and runs at its full bandwidth.
2. About 40-96 GB of weights (70B dense at Q4-Q8, gpt-oss-120b): the 128 GB RTX Spark wins, because the RTX 5090 cannot hold these models. Any layers that don't fit spill into system RAM over PCIe.
3. 14B and smaller: neither machine is needed. A 12-16 GB card runs these entirely in VRAM. Check the RTX 3060 benchmark page before spending flagship money.

If you don't know where your models fall, the per-model hardware picker and the Will-It-Run checker map individual models to memory requirements.

Spec delta: how do the RTX 5090 and RTX Spark compare on paper?

Sources: RTX 5090 specs from TechPowerUp. N1X core counts and memory ranges from Tom's Hardware's launch report. Bandwidth from Wccftech. GPU allocation and laptop TDP from StorageReview. Surface pricing from Tom's Hardware's review.

Two notes on this table. First, the 110 W figure: Microsoft told Tom's Hardware in a June hands-on that the Surface was designed around 110 W for the chip. Microsoft then added a statement that the Surface Laptop Ultra "is not targeting a 110w sustained TDP." Treat 110 W as an upper bound, not a sustained figure. Second, StorageReview's sources disagree with Tom's on the 18-core memory ceiling (StorageReview says 64 GB). Either way, the 18-core chip cannot be bought with 128 GB.

The 300 GB/s figure is a theoretical peak. StorageReview's STREAM Triad run on the Surface measured 122 GB/s from the CPU side. That is a CPU benchmark, not a GPU-path measurement, but it is a reminder that real unified-memory bandwidth comes in below the spec-sheet number.

Full benchmark tables for each chip: RTX 5090, RTX Spark N1X 20-core, RTX Spark N1X 18-core.

How fast is each one, model by model?

The table pairs runs of the same model where both platforms have one. Runtimes and quantizations differ between sources, so read each row as "same model, published conditions", not as a controlled A/B test.

Sources: Hardware Corner — RTX 5090 GPU LLM benchmarks, Hardware Corner — RTX 5090 context-depth tests, DatabaseMart — Ollama RTX 5090 benchmark, StorageReview — Surface Laptop Ultra, Wccftech — ASUS ProArt P16 RTX Spark.

18-core N1X rows. The cut-down chip has published runs from one source, StorageReview's Lenovo Pro 9N 15 review: Phi-4-mini at 95.2 tok/s in llama.cpp and 96.6 tok/s in a RAG answer stage, Llama 3.1 8B at 55.1 tok/s in llama.cpp, and MLPerf Client agent runs on Llama 3.1 8B at 46.9 tok/s (Data Analyst Agent) and 44.5 tok/s (SWE Agent). On small models the two N1X bins land within a few percent of each other, as expected when both share the same 256-bit memory bus.

What the rows show. On models that fit both machines, the RTX 5090 generates about 2.7x faster on Llama 3.1 8B, about 4.3x on gpt-oss-20b and about 3.4x on Gemma 4 26B, using the cited figures. Mixture-of-experts models narrow the gap in absolute terms because only a few billion parameters are active per token. That is why Qwen3.6 35B-A3B still reaches 89.2 tok/s on the Spark per StorageReview. The Gemma 4 31B row (61.1 vs 11.98) is the warning case: a dense 31B model reads every weight for every token, and that is where a 300 GB/s bus hurts most.

Which quantizations fit where?

Weight sizes below are arithmetic: parameter count times the bits-per-weight that llama.cpp publishes for each quant type (llama.cpp quantize README: Q2_K 3.16, Q3_K_M 4.00, Q4_K_M 4.89, Q5_K_M 5.70, Q6_K 6.56, Q8_0 8.50, F16 16.0). They exclude KV cache and runtime overhead, so leave 2-6 GB of headroom depending on context length. "Fits in 128 GB" uses StorageReview's 96 GB maximum GPU allocation, not the full 128 GB.

gpt-oss-120b's size comes from OpenAI's model card on Hugging Face, which says MXFP4 post-training lets it "run on a single 80GB GPU". That is well over 32 GB and well under 96 GB, so it is the clearest example of a model that only the Spark can run fully resident.

Prefill vs generation: why bandwidth decides decode speed and compute decides prompt processing

Running an LLM has two phases. Prefill (prompt processing) pushes the whole prompt through the model at once, so it depends on compute: CUDA and tensor cores. Decode (generation) produces one token at a time and reads the active weights for each token, so it depends on memory bandwidth. The RTX 5090 has about 3.5x the CUDA cores and about 6x the theoretical bandwidth of the 20-core N1X (per the spec table above), so it leads in both phases. The ratio differs by phase and by context length, though, so a single multiplier is misleading.

At 4K context (gpt-oss-20b): prefill is 9,443.8 tok/s on the RTX 5090 (Hardware Corner) against 3,070 tok/s on the 64 GB RTX Spark laptop (Wccftech), about 3.1x. Decode is 298.2 against 68.86, about 4.3x. Wccftech did not state its context length, so treat this as an approximate 4K-class pairing.

At 32K context (dense Qwen3 32B, RTX 5090 only): Hardware Corner measured decode falling from 61.38 tok/s at 4K to 43.82 tok/s at a 32K prefill, about 71% of the short-context speed. Prompt processing fell from 2,931 tok/s to 1,451 tok/s, about 50% (Hardware Corner). No source has published a matching 32K dense-model run on RTX Spark, so there is no ratio for this condition.

At 64K-128K context (gpt-oss-20b): StorageReview's Codex CLI agent runs on the Surface measured 58.0 tok/s at 64K and 59.2 tok/s at 128K in Ollama, which is nearly flat (StorageReview). Hardware Corner's extreme-context table lists gpt-oss-20b at 112.01 tok/s generation and 1,636.40 tok/s prompt processing at 131K on the RTX 5090. That works out to about 1.9x at 128K, much narrower than 4.3x at short context. The runtimes differ (llama-bench vs an Ollama agent loop), so treat 1.9x as directional.

Long context also uses memory. Hardware Corner pushed Qwen3 30B-A3B to 147,000 tokens on the RTX 5090 with peak VRAM use of 31 GB and generation at about 52 tok/s. That is close to the card's ceiling. With a dense 32B model, the same source reached only about 45-47K tokens before running out of memory. If you need dense 32B-class models and 100K+ context together, capacity becomes the limit again, and that favours the Spark.

What happens when a model won't fit on the RTX 5090?

When a model is bigger than 32 GB, llama.cpp and Ollama keep as many layers as fit on the GPU and run the rest on the CPU from system RAM. Each token then waits for the slowest part, which is dual-channel DDR5 on most desktops. Throughput drops far below the card's in-VRAM numbers and starts to depend on the CPU and RAM speed. The gpt-oss-120b offload test on an RTX 3060 and a Ryzen 7 5800X shows this effect on a smaller card. The RTX 5090 can hold more layers, but the CPU side still sets the pace.

No source in this synthesis has published a fully-resident gpt-oss-120b run on an RTX Spark laptop. The nearest data point is NVIDIA's DGX Spark: a GB10 Grace Blackwell mini PC with 128 GB of unified memory and a 140 W SoC TDP (per Tom's Hardware). It runs gpt-oss-120b MXFP4 at 55.37 tok/s per the NVIDIA Developer Blog and at 60.57 tok/s in llama.cpp discussion #16578. An RTX Spark laptop with a lower power limit should not be assumed to match those numbers until someone publishes a run, but they show the scale: an MoE model with 5.1B active parameters runs at conversational speed when fully resident in a 128 GB pool.

Dense 70B is a harder case. LMSYS measured Llama 3.1 70B in FP8 at 2.7 tok/s on DGX Spark, because every token reads about 70 GB of weights through a unified-memory bus in the same few-hundred-GB/s class as RTX Spark. Capacity lets the model run, but it does not make a dense 70B model fast. For dense 70B, the best GPU for Llama 70B guide covers multi-GPU and workstation options.

Is a Strix Halo mini PC the cheaper 128 GB alternative?

For capacity on a desktop, AMD's Ryzen AI Max+ 395 ("Strix Halo") offers the same 128 GB unified-memory approach without a laptop's power limits. The BOSGAME M5 with Ryzen AI Max+ 395 and 128 GB and the Beelink GTR9 Pro with Ryzen AI Max+ 395 and 128 GB are two catalog examples. Per the community-maintained Strix Halo Guide benchmark log, a 128 GB Strix Halo system runs gpt-oss-120b at 56.14 tok/s in llama.cpp. That is in the same range as the DGX Spark figures above.

The trade-off is software. Strix Halo runs llama.cpp through Vulkan or ROCm instead of CUDA, and tools that assume CUDA (vLLM builds, some fine-tuning stacks) need more setup. The Strix Halo LLM benchmark roundup has the full per-model table. If you want 128 GB on a desk and don't need a laptop or CUDA, a Strix Halo box is the main alternative to the 128 GB Surface. Compare the live prices on each product page; both listings carry current Amazon pricing.

Perf-per-dollar and perf-per-watt

All dollar figures below are at MSRP, the launch list price, not today's street price. The only workload with published numbers on both a Surface RTX Spark and an RTX 5090 is gpt-oss-20b, so the math uses it.

RTX 5090 system at MSRP: the card lists at $1,999 (TechPowerUp). A card does nothing on its own, so add a host: CPU, motherboard, 64 GB of DDR5, a 1,000 W+ power supply, case and storage. As a stated assumption, say the host costs $1,500, for a total of $3,499. At 298.2 tok/s on gpt-oss-20b, that is about 85 tok/s per $1,000 of system cost.

Surface Laptop Ultra 128 GB at MSRP: $5,899.99 per Tom's Hardware. At StorageReview's 58.0 tok/s on gpt-oss-20b, that is about 9.8 tok/s per $1,000. The RTX 5090 desktop is roughly 8-9x better per dollar on a model that fits on the card. That ratio does not apply to gpt-oss-120b or 70B models, which the 5090 cannot run resident, so there is no fair per-dollar figure for those.

Perf-per-watt (rated power, not measured draw): 298.2 tok/s over the 5090's 575 W board rating is about 0.52 tok/s per watt, before counting the host CPU. 58.0 tok/s over the top of StorageReview's 45-80 W N1X range is about 0.73 tok/s per watt. Rated power is a ceiling, and real draw during decode is usually lower on both, so read these as rough bounds. The Spark's efficiency lead is real but modest. Its clear advantages are that it is quiet and portable.

StorageReview also reported a caveat on its Surface unit: a sustained all-core double-precision workload made the laptop cut power three times in one evening, after an earlier review unit failed the same way. The reviewer's log evidence pointed at power delivery or firmware protection, not heat. If you plan to run long batch jobs, watch for a firmware fix.

Verdict matrix

Get the RTX 5090 if…

• Your models are 32B or smaller at 4-bit, or MoE models whose quantized weights fit in about 28 GB.
• You care about tokens per second and prompt-processing speed more than portability.
• You already own a desktop that can supply 575 W and has room for a large card.

Get the RTX Spark (20-core, 128 GB) if…

• You need gpt-oss-120b or other models with 40-96 GB of weights fully resident, on a laptop.
• You want CUDA tooling on unified memory and can accept 300 GB/s decode speeds.
• Your budget covers the 128 GB configuration. The 24-32 GB entry models give up the capacity advantage entirely.

Get a Strix Halo box if…

• You want 128 GB of unified memory on a desk at a lower system price than the 128 GB Surface.
• You are comfortable with llama.cpp on Vulkan or ROCm instead of CUDA.
• Your main large models are MoE (gpt-oss-120b, Qwen 30B-A3B class), where unified-memory boxes do best.

Recommended pick

For most local-LLM builders running models up to 32B, buy the ASUS TUF Gaming GeForce RTX 5090 32GB. On published runs it is 2.7-4.3x faster than RTX Spark on shared models, and at MSRP it delivers several times the throughput per dollar. If you want the same chip with a larger cooler and a white finish, the ASUS ROG Astral GeForce RTX 5090 32GB White OC uses the same 32 GB GDDR7 configuration. For a laptop user who needs 70B-120B models resident, the 20-core, 128 GB RTX Spark Surface Laptop Ultra is the only pick in this comparison that does it. For a desk-bound user who wants 128 GB for less money, the BOSGAME M5 is the Strix Halo pick.

Live price comparison

The RTX 5090 vs RTX Spark comparison page tracks current pricing side by side. For the RTX 5090, the ASUS TUF RTX 5090 product page carries the live Amazon price and buy button. The RTX Spark has no catalog listing yet, since it is sold as complete laptops. For a 128 GB unified-memory machine with a live price, use the Strix Halo listings above or see the DGX Spark review for the GB10 desktop option.

Related guides

• RTX 4090 vs RTX 5090 for local LLM inference
• gpt-oss-20b on the Surface Laptop Ultra: RTX Spark vs DGX Spark
• Qwen 3.6 35B-A3B on the RTX Spark Surface: memory configs and quant table
• NVIDIA DGX Spark review: Grace Blackwell performance
• Per-model hardware picker, 7B to 70B

Citations and sources

Accessed October 24, 2026.

• TechPowerUp — NVIDIA GeForce RTX 5090 specs
• Tom's Hardware — Microsoft Surface Laptop Ultra review: RTX Spark
• Tom's Hardware — Surface Laptop Ultra targets 110W TDP for RTX Spark
• Tom's Hardware — RTX Spark N1X launches in October, 18 or 20 CPU cores
• NVIDIA blog — Local AI with RTX Spark
• StorageReview — Microsoft Surface Laptop Ultra for Business review (RTX Spark N1X)
• StorageReview — Lenovo Pro 9N 15 review (RTX Spark N1X 18-core)
• Wccftech — ASUS ProArt P16 RTX Spark laptop review
• Hardware Corner — RTX 5090 GPU LLM benchmarks
• Hardware Corner — RTX 5090 LLM benchmarks at context depth
• DatabaseMart — Ollama GPU benchmark: RTX 5090
• llama.cpp — quantize README (bits per weight)
• OpenAI — gpt-oss-120b model card on Hugging Face
• NVIDIA Developer Blog — DGX Spark performance
• llama.cpp discussion #16578 — DGX Spark results
• LMSYS — NVIDIA DGX Spark benchmarks
• Strix Halo Guide — benchmark log

This piece is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.