eGPUs still bottlenecked: Thunderbolt caps cards to an eighth of desktop bandwidth. External graphics card enclosures still can't shake their oldest problem: the cable connecting them to a laptop simply can't carry as much data as a desktop's internal slot. A Thunderbolt 4 port moves 40Gbps total, and Intel's spec allots PCIe data 32 Gbps of it, an eighth of the bandwidth a desktop PCIe 4.0 x16 slot provides, according to Engadget's testing of current eGPU hardware. For anyone weighing a thin laptop plus a graphics enclosure against a basic prebuilt gaming desktop, the maths hasn't changed much in five years, and it still tilts toward the desktop. Why the Thunderbolt cable rarely gets saturated in games. The good news, such as it is, comes down to how a GPU actually uses that connection. Textures and shaders get loaded into VRAM once and remain there, meaning the cable only gets busy when a game pulls in new assets or sends finished frames back to the display. That's why an eGPU doesn't feel eight times slower than a desktop card, even though the bandwidth gap on paper is that large. Real-world numbers back that up, but the losses are still far from trivial. Independent testing lab Notebookcheck tested an RTX 2080 Ti enclosure back in 2021 and found deficits ranging from seven to 32 percent when compared to a desktop, though losses did shrink to single digits at 4K on an external monitor, with rendering to a laptop screen slower because finished frames travel back down the same cable. HP's own consumer guide arrives at a similar range for current hardware, noting that even Thunderbolt 5's 120 Gbps can't match a desktop PCIe x16 slot's 256 Gbps bandwidth, and estimating a 10-20% performance drop for gaming - and this may be even greater if you're using your laptop's built-in display. Faster gpus suffer worse, making mid-range cards the sensible pick. There's a counter-intuitive twist here: the more powerful the graphics card, the more of its potential gets strangled by the cable. That makes a mid-tier GPU, not a flagship, the more sensible choice for anyone building an eGPU rig, since a card that would dominate at native PCIe speeds simply can't stretch its legs through a four-lane Thunderbolt tunnel. Gamers chasing the absolute top-end cards are effectively wasting money on silicon the connection can't feed fast enough. Razer's revived Core X V2 illustrates the trade-offs. Razer's return to the eGPU market underlines how incremental the gains have been. The company revived its eGPU line with a Thunderbolt 5 dock in July 2025, launching the Core X V2 enclosure alongside a separate connectivity dock. Tom's Hardware reported the enclosure aims to consolidate display, data, and storage expansion into a single hub, touting up to 120Gbps burst bandwidth, with the dock delivering up to 140W of system charging via a single Thunderbolt 5 cable - but that dock is sold separately from the GPU box itself. Even with Thunderbolt 5's headline 80Gbps of bandwidth, the actual GPU connection is narrower: Razer's $350 Core X V2 connects the GPU through PCIe 4.0 x4, nominally 64 Gbps, giving an x16 graphics card four lanes instead of the 16 available in a suitable desktop slot. And buyers don't get a complete kit for that price. The Core X V2 comes with only the enclosure - no power supply, USB ports or Ethernet - meaning owners must add their own ATX PSU, with Razer selling the missing I/O separately as a Thunderbolt 5 dock starting at $390 for the Mercury version. PCWorld's review of the launch was blunter still, concluding that at this point, you may be wondering why you're spending $350 if you still have to fork over additional cash for an external GPU and a power supply. Framework's OCuLink kit chases a different fix for Laptop 16. Not every manufacturer is betting on Thunderbolt. Framework is building an eGPU kit for its Laptop 16, and the company has taken a markedly different technical route. Framework's own blog describes the upcoming OCuLink Dev Kit as an interface technology that allows bringing the internal PCIe out over a connector and cable to enable external peripherals like eGPUs without the overhead of Thunderbolt. The kit uses an OCuLink 8i interface (PCIe x8, up to 128 GT/s), that bypasses the bandwidth limitations of traditional external solutions like Thunderbolt or USB4 - twice the lane count of Razer's Thunderbolt-based enclosure. The catch is convenience. Framework itself frames the OCuLink kit as a tool for hands-on tinkerers rather than mainstream buyers, since the connector doesn't support the same plug-and-play hot-swapping that Thunderbolt users take for granted. It's a trade familiar to PC modders everywhere, including in Australia and New Zealand's enthusiast-building community, where OCuLink and DIY eGPU rigs have circulated in forums for years precisely because Thunderbolt's shared bandwidth frustrated performance-focused builders. Who should actually consider an eGPU. None of this makes eGPUs pointless - it just narrows who they're genuinely for. Someone who already owns a good ultraportable laptop and occasionally wants desktop-tier frame rates at home, on an external monitor, with a mid-range card, will likely be satisfied. But for anyone starting from scratch and asking whether to buy a laptop-plus-enclosure combo or simply build or buy a basic gaming PC, the enclosure, power supply, dock and graphics card can easily add up to desktop-PC money for a setup that still loses a meaningful chunk of that card's performance to a narrow cable. For Kiwi and Australian gamers pricing out a new setup, that calculation is worth doing carefully before committing to an eGPU enclosure, since importing Thunderbolt 5 docks and enclosures on top of a graphics card and PSU adds shipping and currency costs that a locally bought prebuilt desktop simply avoids. Sources. Article by News Desk. The GamesReviews News Desk covers the latest gaming news across PlayStation, Xbox, Nintendo, PC and esports. Reporting is AI-assisted and reviewed by its editorial team.

Framework is refunding customers who paid inflated prices for RAM in its model 13 Pro laptop. In July, the company reduced RAM allocations when memory costs more than doubled unexpectedly, forcing some customers who ordered 64GB to receive only 32GB. The laptop maker has now secured a limited quantity of Micron 32GB and 64GB LPCAMM2 memory modules at lower prices. Customers who purchased when memory costs peaked will receive refunds for the price difference. Those with pending orders will see price reductions and can upgrade their RAM at the new, lower rates. Framework joins other tech companies, including HPE and Everpure, in absorbing increased memory costs. The company specialises in highly repairable laptops.

Ryzen AI Max+ PRO 495 Mini PCs Offer 192GB RAM for Local AI. Acemagic, GMKtek and Framework are launching mini PCs built on AMD's Ryzen AI Max+ PRO 495 chip with up to 192GB of unified memory, enough to run 300-billion-parameter AI models entirely on-device. What happened. A wave of mini PC makers, including Acemagic, GMKtek and Framework, are launching compact desktop systems built around AMD's new Ryzen AI Max+ PRO 495 processor, configured with up to 192GB of unified memory. The memory capacity is large enough to run local large language models with roughly 300 billion parameters entirely on-device, without relying on cloud infrastructure. These systems represent a new category of small-form-factor "AI workstations" that bring capabilities previously reserved for data-centre GPUs or high-end workstations into a desktop unit small enough to sit on a shelf. The high memory ceiling is the standout feature, as it allows the chip's integrated graphics and AI processing units to hold and run very large models locally rather than offloading inference to remote servers. Why it matters. The emergence of mini PCs capable of running 300-billion-parameter models locally signals a meaningful shift in where AI inference can realistically happen. Until now, models of this scale have typically required cloud-hosted GPU clusters or specialised on-premises hardware. Compact machines with this kind of memory headroom point toward a future where organisations - and even individual developers - can experiment with, fine-tune or deploy substantial AI models without routing data through third-party cloud services. For technology and digital transformation leaders, this matters less because of the specific hardware brands involved and more because of what it makes possible: greater control over data residency, latency and cost for AI workloads that would otherwise depend on external infrastructure. Local-first AI capability at this scale could reshape decisions around where sensitive or performance-critical inference tasks are run, particularly for regulated industries or regions prioritising data sovereignty. By the numbers. * 192GB of unified memory available in the top configuration, enabling large-model workloads to run without external GPU support. * 300 billion parameters is the scale of large language model these systems are reported to be capable of running locally. The renascence take. The headline number here is memory, but the more interesting story is about where AI processing power is heading - away from centralised cloud dependency and toward the edge, even down to desk-side hardware. That has implications well beyond hobbyist and developer circles. Most coverage will focus on the novelty of "supercomputer-grade" mini PCs, but the real signal is behavioral: when powerful AI capability becomes physically local and personally owned, trust and adoption dynamics change. Teams are more willing to experiment with AI when it feels contained, private and under their control rather than routed through an unseen cloud vendor. Organisations exploring AI-enabled service design should watch this shift closely - not to buy the hardware, but to understand that "local AI" is becoming a credible design constraint customers and employees will start to expect, particularly around data privacy and responsiveness. This briefing was written by its Newsdesk, synthesising reporting from the outlets below. Follow the links for the original coverage. Questions Renascence get on this topic. With up to 192GB of unified memory, these mini PCs can run large language models of roughly 300 billion parameters entirely on-device, without relying on cloud servers. More in AI Stay ahead of CX Get the signal, not the noise. The stories shaping customer experience - plus the Journal and Experience Loom - in your inbox.