ASUS ProArt GR1X: The 128GB AI Memory Trap Nobody Talks About

The trap is this: 128 GB lets the ASUS ProArt GR1X load a large AI model, but its speed depends on memory bandwidth, not capacity. ASUS quotes "up to 300 GB/s" for the NVIDIA RTX Spark N1X chip inside it. That is far less than a high-end graphics card's 1,792 GB/s or a Mac Studio M5 Ultra's 1.2 TB/s. When a model generates text, the chip has to read its active weights from memory for every token. So a model that "fits" can still crawl. The GR1X shines on mixture-of-experts (MoE) models such as gpt-oss-120b, which read only a small slice of their weights per token. It struggles with large dense models such as a 70B at 8-bit.

Why doesn't 128 GB make AI fast?

Text generation has two phases. Prefill reads your prompt and is limited mostly by compute. Decode produces the answer one token at a time, and for each token the GPU must stream the model's active weights out of memory. That makes decode a bandwidth problem. A useful rule of thumb: the theoretical ceiling for single-user tokens per second is roughly memory bandwidth divided by the bytes read per token. Real software reaches perhaps 60–85% of that.

Capacity decides whether a model loads. Bandwidth decides how fast it answers. The GR1X has plenty of the first and a modest amount of the second. That's not a flaw unique to ASUS: every RTX Spark N1X system shares the same memory design, whether it's a laptop, the MSI EdgeMesa or the GR1X.

How fast can a 300 GB/s machine generate text?

Source: our arithmetic (bandwidth ÷ bytes read per token, single user, theoretical ceilings, not measurements). Bandwidth figures from ASUS (GR1X "up to 300 GB/s"), Apple and NVIDIA spec pages. Model sizes from Hugging Face file listings. gpt-oss-120b's active read assumes about 5.1 billion active parameters at roughly 4.25 bits per weight. NVIDIA itself hasn't published an N1X bandwidth figure; reviewers' calculations also land near 300 GB/s.

Published results on the same chip line up with this. No GR1X benchmark had been published at the time of writing, so these come from RTX Spark laptops:

• gpt-oss-120b: Techtesters measured 63.3 tokens per second on an MSI Prestige N16 Flip AI+ 128 GB and 58.9 on an ASUS ProArt P14 (16 Oct 2026, tool and quantisation not published). That's about 55–60% of the ceiling above, and quick enough to feel responsive.
• Qwen3 32B: XDA measured 11.8–12.1 tokens per second in Ollama on a Lenovo Yoga Pro 9n 64 GB (26 Oct 2026). That's usable, but closer to reading speed than to instant.
• Qwen3.5-122B-A10B, a larger MoE: Techtesters measured 30.6 tokens per second on the MSI laptop.

The GR1X runs at a 140 W sustained power level, against 45–80 W for the laptops. That may help prefill, image generation and rendering. It's unlikely to raise decode speed much, because the bandwidth is the same. Nobody had tested this on the GR1X when we wrote this.

128 GB of unified memory decides which models you can open. Roughly 300 GB/s decides how long you wait for each answer.

Do you even get all 128 GB for AI?

Not by default. On RTX Spark laptops Windows splits the unified memory between the CPU and GPU. PCWorld reported a default split of about 50/50 on the ProArt P16 128 GB, adjustable to about 96 GB for the GPU in MyASUS (29 Oct 2026). ASUS hadn't documented the same control for the GR1X at the time of writing, so treat it as likely rather than confirmed. On 64 GB machines, XDA found Windows exposed at most 55.4 GB to the GPU. That's why gpt-oss-120b, with about 65 GB of weights, won't load on a 64 GB N1X.

The model also isn't the only thing in memory. Each conversation keeps a KV cache that grows with context length. For Llama 3.3 70B it's about 0.33 MB per token at 16-bit, so a 32K-token conversation adds roughly 10.6 GB and a 128K one about 42 GB. Big contexts and several users eat the "spare" memory quickly. We go deeper on that in whether the GR1X is powerful enough for multiple users.

Which models suit the GR1X, then?

• Good fit: MoE models with a small active size, such as gpt-oss-120b, Qwen3.5-122B-A10B and the many 30B-A3B-style models. They use the capacity without paying the full bandwidth cost per token.
• Fine: 8B–32B models at 4-bit for one person, plus image generation with FLUX-class models, where the GR1X's GPU compute matters more.
• Slow: dense 70B models, especially at 8-bit or with long contexts. They load, but single-user speeds of roughly 4–7 tokens per second (theoretical) feel sluggish for chat.
• Not a fit: models larger than about 100 GB at their chosen precision, or serving a team of users at once. Those call for more bandwidth, more memory, or both.

NVIDIA's "120B parameters with 1M context" headline is a capacity statement. It doesn't say which quantisation or how fast. Our explainer on what NVIDIA's 120B and 200B model claims really mean covers that wording.

How do you avoid the trap?

1. List your actual models, with their size at the precision you'll run. Divide about 300 GB/s by the bytes read per token to get a ceiling, then expect roughly 60–85% of it.
2. Prefer MoE models if you want large-model quality on this class of machine.
3. Budget memory for context, not just weights, and remember the default GPU split.
4. Compare bandwidth, not just gigabytes, when you look at alternatives. A Mac Studio M5 Ultra or an RTX PRO 6000 costs more but generates text several times faster. Our GR1X vs RTX 5090 post shows the other side of the trade-off: a fast GPU with too little memory.
5. Choose 64 GB or 128 GB on workload, not marketing. Our 64 GB vs 128 GB guide walks through which models need which.

Buying in Singapore

The ProArt GR1X is listed at SourceIT as price on request in its ASUS ProArt GR1X collection, with 64 GB and 128 GB configurations. If you'd rather compare against the same chip in a laptop now, the ASUS ProArt P16 RTX Spark 128 GB is S$10,773.20 and the MSI Prestige N16 Flip AI+ 128 GB is S$10,854.12 (listed 8 Nov 2026). Prices change, so ask for a current quotation. For higher bandwidth, the NVIDIA DGX Spark collection also carries Mac Studio configurations on request. Send us the models and context lengths you plan to run and we'll suggest a configuration. To get a quote, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Is 128 GB on the ASUS ProArt GR1X all usable for AI? Not by default. On RTX Spark laptops Windows gives the GPU about half the memory, adjustable to about 96 GB in MyASUS on the ProArt P16. ASUS hasn't documented this for the GR1X yet, and some memory always goes to Windows and to the KV cache.

How fast will a 70B model run on the GR1X? Theoretically about 7 tokens per second at 4-bit and about 4 at 8-bit for a single user, based on roughly 300 GB/s of bandwidth. Real speeds are usually lower than the ceiling. No GR1X benchmark had been published when we wrote this.

Why is gpt-oss-120b fast on RTX Spark when it's so large? It's a mixture-of-experts model that reads only about 5 billion active parameters per token. RTX Spark laptops with 128 GB have measured around 59–63 tokens per second on it.

Does the GR1X's 140 W power make text generation faster than a laptop? Probably not by much. Text generation is limited by memory bandwidth, which is the same on every N1X system. Higher sustained power is more likely to help prompt processing, image generation and rendering, but this is untested on the GR1X.

What should I compare instead of memory size? Compare memory bandwidth and the bytes your model reads per token. The N1X has about 300 GB/s, a Mac Studio M5 Max 614 GB/s, an M5 Ultra 1.2 TB/s and an RTX PRO 6000 1,792 GB/s.