View all articles
newsopinionno-hypeai

No Hype: ASUS promises 32GB of VRAM for your local tasks

JG
Jacobo González Jaspe
|
No Hype: ASUS promises 32GB of VRAM for your local tasks No Hype

ASUS rates the card at up to 367 TOPS of INT8 performance. We cover Scale AI worker tracking complaints and a ChatGPT medical error lawsuit.

ASUS targets local workloads with 32GB VRAM

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. ASUS reports the UGen Arc Pro B70 provides 32GB of GDDR6 memory with 608GB/s of bandwidth (absolutegeeks.com). The card delivers up to 367 TOPS of INT8 performance and has a board power of 310W. ASUS states that performance can drop by 90 percent if a model spills into system RAM. ASUS quotes 43.4 tokens per second for an 80B-class model on a dual-card setup.

What we think. 32GB at 608GB/s is a real number; so is the 90 percent drop ASUS itself quotes when a model spills into system RAM, and the second number decides whether the first one matters. We read this spec sheet as a memory budget, not a speed promise; you check the VRAM calculator before you check the price.

The good bit. One card handles models under 27 billion parameters with a 128K-token context using Llama 3.1 8B (absolutegeeks.com).

What you can do. You can use llama.cpp on Intel Arc with SYCL to run models on this hardware.

Noyb files complaints against Scale AI for worker tracking

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Noyb filed five complaints against Scale AI in Austria, France, the Netherlands, and Italy (escudodigital.com). The group alleges the company used invasive biometric verifications and tracked European workers illegally. Meta acquired a 49% stake in Scale AI in June of last year. Noyb claims trackers are installed on devices when using the Outlier platform.

What we think. Five complaints in four countries make this a labour case before it is an AI case: checking that someone is who they say they are is not a licence to know which task they are on, or for whom. We treat identity as a check, not a tracker; you review what you collect from the people who work for you, then read the EU AI Act guide for where the line sits.

The good bit. The complaints target specific practices regarding employee data and identity verification (escudodigital.com).

What you can do. You check your local AI for developers environments for any unauthorised data collection.

Lawsuit alleges ChatGPT advice led to fatal medical error

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. A lawsuit by the parents of Bowen Wan claims ChatGPT-4o coached her to taper off anti-seizure medication (CNET). The chatbot stated on February 2 that reaching 0 milligrams of lamotrigine by February 15 was feasible. Wan was found dead on February 14. A separate lawsuit alleges ChatGPT-4o misdiagnosed a patient named Scott Winters. OpenAI research found 60% of American adults used AI for health information in a three-month period.

What we think. A chatbot that writes a tapering schedule with dates and a target dose is making a medical decision, and a medical decision needs someone who answers for it, not a text generator. We keep LLMs away from medication choices, on our machines and anyone else’s; you take them out of every health decision in your organisation.

The good bit. The lawsuit brings attention to the documented use of AI for health information (CNET).

What you can do. You review how much VRAM do I need for a local LLM? to ensure your local models run safely.

In one line

ASUS puts 32GB on one card; we treat the 90 percent spill penalty as the spec that matters, and you size the model to the memory before you buy.

Work with us

If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.

Diagram
Share: LinkedIn X
Veredicto semanal

Get new guides before anyone else

Subscribe and we tell you when new guides, templates and workflows go up. One email a week, no spam.

Already published: 98 guides and 25 templates. All free, no signup.

Bonus: the local-AI starter pack PDF when you subscribe
Once a week No spam Unsubscribe anytime

See what you get