However, beginners might not know where to start and experts will be aware that training Large Language Models (LLMs) in the wrong environment with the wrong hardware can lead to expensive cost blowouts, reliability issues, massive delays plus plenty of privacy, security and intellectual property concerns.

The good news is that Gigabyte has sought to significantly simplify all elements of the Ai workflow with its AI TOP technologies which combine software, hardware and automated Ai advice to facilitate the creation of Ai systems (consumer to business grade) at your own desk, efficiently and securely.

Models can be generated in just a few clicks using the AI TOP Utility interface that will be familiar to all gamers (it also allows experts to use console commands). It comes with the top 20 most popular Ai models (like Llama, GPT and BERT) included. There’s also a cloud-based AI TOP Tutor, called Airus. She can provide tech support, setup advice and even expert consultations.

All AI TOP Hardware components and pre-built systems will be familiar to PC gamers, but they’ve been augmented to work better with Ai workloads. Everything has been tested extensively to work in harmony, at peak performance, through long LLM processing workloads that can last many (potentially thousands of) hours.

Gigabyte’s extensive testing and Ultra Durable engineering processes means that all hardware also comes with an extended warranty.

The resulting PCs aren’t specialist workstations with expensive certifications, but they’ve nonetheless been extensively tested by Gigabyte to perform reliably for the required, extended periods. It’s also worth noting that they have no special power requirements and all components are also designed to work with a single PSU.

In addition to a variety of complete PC systems, that suit different industries and budgets, AI TOP hardware includes:

Motherboards

The Gigabyte TRX50 AI TOP is a foundational Ai motherboard with four, dual-slot PCIe 5.0 x16 slots and four SSD slots, that can run in RAID 0. It has two 10GbE LAN ports and room for 2TB of DDR5 RAM. The Gigabyte W790 AI TOP motherboard takes things up a notch by adding two Thunderbolt 4 ports and an additional two SSD slots.

Graphics Cards

Gigabyte’s AI TOP graphics cards come in two flavours. The first is Nvidia based – the GeForce RTX 4070 Ti Super AI TOP 16G – which is a dual-slot card, optimised for quad-card systems, with a full-copper heatsink, optimised airflow channels and turbo fans.

Gigabyte is also the only vendor to make graphics cards with AMD Radeon Pro series GPUs. The Gigabyte-constructed cards come in two flavours: the W7900 AI TOP 48G and the W7800 AI TOP 32G (which really do contain 48GB and 32GB of GDDR6 RAM, respectively). The huge amount of VRAM (potentially 192GB in a four-card PC) opens the door to even larger Ai models and significantly reduces the risk of LLM training failure.

SSDs

The AI TOP 100E is available in 1TB and 2TB flavours and allows 150x more Total Bytes Written than a regular SSD to maximise longevity when dealing with extended, write-rewrite, Ai-based workloads.

PSU

The Gigabyte UD1600PM PG5 AI TOP works with a single wall plug and can support systems using four graphics cards. It uses server-grade components and has been tested to handle ongoing, 24-hour Ai workloads. It has been designed to last twice as long as competitor products and has an 80 PLUS Platinum rating for power efficiency.

Whether you choose to build your own or buy a complete system, you’ll end up with an ultra-reliable PC that’s fully optimised to run Ai workloads for potentially thousands of hours at a time.

Finally, professionals working in the game industry can augment their workflows to become more efficient at their regular work while gaining the availability of a whole new toolset. The latter was recently on show at Computex when Gigabyte held an exhibition with world famous artists.

Titled, “Humanity X Art X Technology” the event, ‘explored the convergence of artificial intelligence with human creativity, artistic expression and technological advancements’ in Taipei. The collective used Gigabyte Ai PCs, powered by Nvidia GPUs, to create a variety of new-media artworks.

However, a particular highlight was made in collaboration with the Hello World Future Ltd collective which used the robust, high-performing Gigabyte Ai PCs to create a unique battle system by integrating the concept of classic fighting games with generative Ai in a nostalgic arcade atmosphere.

The exhibition showed gamers and a new generation of Ai-prompt engineers, just what could be created with the new hardware and software.

So, if you’ve been wondering how to break into Ai, or how to find hardware that won’t let your Ai workloads down, check out Gigabyte’s AI TOP Utility, Tutor and Hardware collection which has offerings to suit every budget and every industry. Find out more click here.

The 890M is the new iGPU to launch within the new AMD Ryzen AI 300-series. It includes 16 RDNA 3.5 CUs and will run at 2,900 MHz. The last-generation 780M, found in both 7000-series and 8000-series Ryzen mobile processors, contains just 12 RDNA 3 CUs at 2,800 MHz. From speeds and feeds alone, the 890M should be a decent step-up in performance, but we are now able to more confidently confirm that to be the case.

In new benchmarks from GPD (spotted by HXL on X) covering its latest GPD Duo device—a compact 2-in-1 laptop fitted with an AMD Ryzen AI 9 HX 370—we can now ascertain the rough improvement with the 890M versus the 780M.

At 54W, the GPD Duo reaches a score of 4,221 in 3DMark Time Spy. GPD's own benchmarks suggest the GPD Win Max 2, powered by the 780M, reaches a score of 3,218. That means the 890M is 31% faster than the 780M.

The 890M also performs better than a discrete RTX 2050 running at 45W, which scored 3,769 in the same test, and only a hair below an RTX 3050 at 50W. What's not clear is which devices these discrete GPUs were running within. Yet for an iGPU with a power budget and memory shared with its CPU component to be competing with a dedicated graphics chip, that's impressive.

We also ran a quick test on the ROG Ally at 30W, which comes with a semi-custom version of the Ryzen 7 7840U and the 780M, and it scored 3,180. That's a 33% increase for the 890M.

Though the AMD Ryzen AI 300-series also comes with both Zen 5 and Zen 5c cores. GPD shows these cores working well in Cinebench 2024 (R24)—a 124 single-thread score to pretty much match the Ryzen 9 7950X and a 1,525 multithreaded score to match the Ryzen 9 5950X.

Only a single source of benchmarks isn't much to go on, which means there's still more to look out for when laptops fitted with these chips actually land next month. Moreover, it'll be really telling if the next-generation of gaming handhelds opt for AMD's new 890M for a genuine uplift in performance with roughly the same battery life. That's certainly possible, looking at the benchmarks above, though there's also the 12-core 880M that may also be a possibility.

There are some guidelines for business cards entering into the contest: the card has to be roughly 3.5 x 2-inches and should "fit easily in a pocket". Beyond these few basic guidelines, the rules are open to interpretation. Though there are a few categories to strive for, including for the thinnest cards, utilitarian cards, and best-looking cards. 

The contest runs until July 2 but we've already various excellent entrants to gawk at.

One project, crafted by LambertTheMaker, ditches the usual information you'd expect to find on a business card, such as a phone number, email address, or any discernible company detail, for a basic 8x8 LED matrix screen, on which to play Snake. The business card, nee black PCB, also contains directional and A/B buttons to control it.

So, a games console, then? That's my kinda card.

There's one more like it, too. This card from Edison Science Corner is a mini tic-tac-toe console.

Another is a couple sheets of paper—seriously—and between which creator Lincoln Uehara has stuffed a circuit. With the correct cuts and folds, it's possible to fold the pieces of paper into a usable USB Type-A connector. In theory, it's then possible to use this to create some sort of usable flash drive, though as far as I can tell, the WIP project lights an LED.

This card from Martynas is handy in emergency situations. Well, sort of. It's a circuit able to generate an SOS signal in morse code.

Adam Billingsley's entry is a card capable of delivering a text-based payload to your PC when connected via USB, sharing your key information (and ideally not something nefarious), and which also sets off an awesome rainbow animation. One of the best bits about this entry, however, it reportedly only costs $4 to make.

Vying for the 'most expensive' category, if there is such a thing, is probably this one from Dimitar with an e-ink display embedded in the front of it. That means it also requires a fairly large lithium battery.

My absolute favourite design, however, is chaosneon's "persistence of vision" card, which you have to shake to reveal the text displayed on its handful of LEDs. This project, like some of the others, has been in the works for a good few years now—business card design is surprisingly popular on Hackaday.

The best bit about most of the designs—they're open for others to copy and develop themselves.

The practicality of any of these cards is, uh, dubious at best. I struggled to get my company to pay for a couple hundred cardboard business cards for a trip to Taipei—what will it think when I submit my invoice for 200 tiny games consoles? 

Hey, some of the designs are cheaper than you'd expect, at least. This one from Dima I'd actually like for myself—it's a handy card for its included measurements on the exterior and it includes an NFC chip to put all your details or a link to your portfolio on, which is easily accessed by the scan of a mobile phone. At around $36 for 100 units, that's actually do-able! 

"Asetek A/S has received updated purchase forecasts from a number of the company’s largest OEM customers," the company said on June 11 (via Overclock3D). "Based on these new forecasts, the expected increase in demand in the second half of 2024 of the company’s liquid cooling products may not materialize."

The company has suspended its guidance for 2024, effectively throwing its hands in the air and saying it isn't sure what the future may hold.

The company is publicly traded on the Nasdaq Copenhagen. When the news reached the exchange, the company's share price fell 40% in a single day—from 5.02 DKK to 3.00 DKK. It's declined slightly further since. Today it's sitting at 2.62 DKK.

You might not know Asetek as a brand but you've likely used one of its products. The company invented the all-in-one (AIO) liquid cooler, and it's been responsible for making many popular models around today, including those from Thermaltake, Lian Li, NZXT, Phanteks, Asus ROG and more. It also works with system builders, including Alienware, iBuyPower, Falcon, and Overclockers UK.

It's these component makers and system builders that generate the demand for Asetek products, and they're just not biting right now. The PC market is in a bit of a slump, which most firms are hoping AI will drag them out of. However, most 'AI ready' devices today, such as Copilot+ PCs, are laptops today.

Asetek's liquid coolers are primarily aimed at desktops, though liquid cooling is also used in datacentres.

"The management team and board of directors of Asetek will consider the long- and short-term consequences of the weakened and uncertain market situation. The assessment today is that the negative trend in revenue may continue into 2025."

Though whether this is a general malaise for liquid coolers in the wider market or one more localised to Asetek is not entirely clear.

We spoke to one hardware manufacturer at Computex and were told it has invested in a new cooler without Asetek due to Asetek's prices. Though with customers of all sizes, prices likely vary a lot. Ultimately Asetek's cooling business is tied to PC demand, and that's been in the doldrums for a while.

A new generation could spark fresh interest in PCs, however, and we are expecting both AMD Ryzen 9000-series CPUs in the very near future and Intel's Arrow Lake CPUs later in the year. Also, Nvidia and AMD are surely, surely, working on new graphics cards to bless our gaming PCs with at some point in the near future. That's why this slump towards the tail-end of 2024 is a bit surprising—there's actually heaps of new hardware coming out to bolster demand. 

Or maybe everyone is just really excited for Noctua's pump-less AIO concept? Nah, that's miles off being a finished product yet.

Asetek also designs sim racing gear, its SimSports division, which it says is doing just fine and on track. So it's not all bad news.

But today, now that the early reviews are starting to trickle out—either in sponsored form or otherwise—the news isn't nearly so promising. And it's nothing to do with the brutal bout of Computex showvid-19 I came back with; it's all about the performance hit when it comes to gaming under that emulation layer. 

Dave2D's video offers up some okay benchmark figures at the 1080p Low preset level on a selection of games but goes on to highlight that "there's also occasional stutter in the games. I don't think it's stutter from a hardware thing, I'm convinced it's completely a software issue. Because the thermals were fine—and sometimes you'd go for a long stretch with a sustained frame rate—but eventually you get a stutter."

Even the sponsored Matthew Moniz video about the X Elite-powered Asus Vivobook S 15—paid for by Asus—notes that gaming is the real weak point of the Qualcomm chips. Fortnite refuses to load (which I believe is an EAC issue from talking to some Qualcomm folk around Computex) and they highlight Diablo 4 running fine for around a minute or so before suffering a catastrophic Prism crash.

Overwatch will work, but is seemingly locked to 1280 x 1080 and regularly stutters because "it feels like it's translating on the fly." Moniz goes on to state that this is the real issue with the ARM-based chips. "No matter what game you play on this, if it's been optimised for x86, you're going to take a performance hit… Now, if you're playing an Arm game, like World of Warcraft, it's good."

"But if you plan on doing any sort of gaming you're not going to have the best experience right now."

Now, none of this should come as a surprise really. When it was first announced Qualcomm was making a PC chip and would be releasing laptop silicon to run Windows with an emulation layer for x86 applications, not a single one of us was like 'yeah, that's going to be great for PC gaming.'

And yet Qualcomm, bizarrely, has made it a part of every recent event based around its new processor range. There was no need, more so if you knew you were going to be leaving a ton of relative gaming performance on the table as a result of the emulation layer. Maybe it was surprised that so many games did just work, and wanted to shout about that, but if the performance is half that of an equivalent x86 machine it's a tough one to get too excited about right now.

I will say I was kinda hoping for something more exciting when it came to the battery life, too, especially given how efficient the X Elite architecture was supposed to be, but the comparison with the latest Intel and AMD x86 chips isn't exactly night and day. The Vivobook S15 does have a smaller battery than the Intel and delivers longer uptime under load, but it's not exactly revolutionary. If I was going to suffer a performance hit I'd at least want something in return.

What Dave2D does say, however, is the sleep performance of the X Elite processor is very impressive, and that is absolutely something you cannot say for any laptop running Windows in an x86 environment. Shut the lid of your laptop and you can be pretty confident after a day of sleep it will have run the battery dry. The X Elite, however, reportedly will only lose something like 1 or 2% of its battery life overnight.

So it ain't all bad, and the thing to note is that this is early doors for Qualcomm's new ARM-based PC chip adventure. Reports are the hardware is very good, but it's all about the software and the software emulation layers. And those are things that have a good chance of being improved over time.

We haven't yet had more than a hands-on experience with a few different Snapdragon X Elite machines so far, and they are still notable for their absence from our labs. But we hope to have our own reviews live soon.

How weird would that be in your actual day-to-day? Well, it's real, and a bank in the United States is rolling out this sort of system to its customer call centres, as American Banker's Penny Crosman reports.

First Horizon Bank is opting for this system as a way to keep its call centre agents relaxed over long shifts of dealing with the public—I get it, people can be horrible. The bank hopes the system could help deal with burnout among agents.

It all relies on Cisco's AI model for call centres, Webex Contact Center—after all, only AI could make something so intrinsically weird. It requires monitoring of an agent's stress levels through various markers, including responses to customers. Cisco says this information isn't retained. 

Altogether, the system is able to piece together a picture of an agent's day-to-day stress level.

"Different people basically break at different breakpoints," Aruna Ravichandran, SVP and customer officer at Webex by Cisco, says to American Banker.

When an agent is close to their breaking point, the system recognises this and runs—I can't believe this is actually what this is called—a Thrive Reset.

A Thrive Reset is a montage of personal photos, set to music, with inspirational quotes. Thankfully, agents pick their own photos and songs—it might be a little too bleak for an AI to generate some fake family vacation snaps set to Tiny Tim's Tiptoe Through the Tulips. 

Have you seen the Apple TV original Severance? It's the equivalent of the egg bar social.

At this point in writing this story, I assume this is entirely made up. There's no way it's called a Thrive Reset, I tell myself. Alas, the videos are made by very real 'productivity and health platform', Thrive Global. And yep, Thrive Global integrates with Webex Contact Center. And just to be extra sure, Ravichandran is a real person, too.

And, apparently, the system works. That's the thing I can't really get my head around. It sounds awful, but in a number of trials, First Horizon saw at least double digit reductions in levels of burnout in agents. In other pilot schemes for so-called Thrive Resets, agents preferred this sort of break to just having some free time. Customer satisfaction also increased a couple of percentage points.

Trialled earlier in the year, and with plans to roll out the system even further by March, First Horizon Bank should have the system handing out montages to 3,000 agents by now.

So, if agents are happier and customers are happier, who am I to turn my nose up at this burgeoning AI use case? Nope, sorry, this is all too alternate timeline Corporate America for me. I prefer getting my workplace rage out the traditional way: bottling it up until I can buy a new overpriced item of clothing or gadget to give me a fleeting sense of control over my own life. 

Phew, that's better.

At Computex 2024 I got the chance to get hands-on with the latest version, the Frore AirJet Mini Slim. I also had the opportunity to chat to Frore Systems CEO, Dr. Seshu Madhavapeddy, about scalability, use-cases, and exactly what sort of performance the new models can deliver.

Surrounded by various devices jury-rigged with AirJet Minis, one of my first questions was about noise. A pet theory, perhaps, but an early assumption I'd made about the lack of adoption was how much noise a jet of air would make, and whether a high pitched whine or similar issue might be holding it back. I asked Dr Madhavapeddy if I could hear one, up close, moving air at full-tilt.

"Of course", he replies. "You're surrounded by them right now."

While the various Macbook Airs, iPhones and similarly thin devices around me were wearing transparent cases to show the AirJets inside, I'd assumed they were idling, given the lack of fan noise in the room. I picked a laptop up and held it to my ear, and sure enough, it was pushing a decent amount of hot air while running a benchmark—and the noise was only audible a few inches away. Even then? Surprisingly quiet.

Granted, this was in a conference room at the Nangang Exhibition centre, so background noise was absolutely a factor. However, with no distinct whine or noisy whooshing of air from even a close distance, the AirJets I saw were running much more quietly than I'd expect at max capacity.

Beyond that, the benchmarking graphs show significantly better performance results for chips fitted with AirJets compared to models featuring traditional cooling systems, or passive solutions. So if the performance is impressive and my personal noise theory quashed, why haven't we seen the AirJet feature in more devices?

"Doing something as disruptive as we are doing—basically we are creating the entire supply chain, and the manufacturing line for this—this is a very complex product that we've built using entirely new materials, and manufacturing processes," explains Dr. Madhavapeddy.

"So, you know, I would urge you to be a little patient."

Fair enough. However, the products I was shown looked—and felt—like production units, not prototypes, and the room was full of them. How many are Frore ready to produce, I asked?

"Millions," Madhavapeddy replies, with a grin. 

Holding the AirJet Mini Slim in my hands and looking through media detailing its internal mechanism, it isn't difficult to see why such a small unit with such a radical approach to moving air would take time to scale for mass production.

The internal piezoelectric membrane vibrates at 24 kHz, and pulls air in from the top of the device, disrupting the "boundary layer" of hot air above the chip being cooled underneath it, before firing it out of a thin vent at one end. This mechanism is said to be key to the AirJets efficiency in such a small device, compared to a similarly-sized conventional fan.

That fast moving membrane, however, strikes as an easy point of failure. I asked if Frore had conducted any long-term durability testing, and the expected lifespan of a single unit:

"We've gone through numerous reliability tests," says Madhavapeddy. "We have been able to confirm that it would last for about 10,000 hours of continuous operation. That should be about five years in a typical use case scenario."

While fans die all the time, I'd say it'd be reasonable to expect a conventional fan to last for over five years of operation. A Noctua fan is rated at over 150,000 hours, for example. Still, with the AirJet acting as a single, self-contained unit—and providing a device using one could be opened easily—replacing one when it fails doesn't seem like it'd be too taxing of a task, and potentially easier than a conventional fan/heatsink combo in a regular laptop.

Looking around the room, most of the demonstration devices had one thing in common: very thin, very light, and difficult to design around conventional fans. Given the claimed efficiency of the device, I wondered out loud about eventually seeing several strapped to a desktop CPU. At this suggestion, Dr Madhavapeddy seems to slightly recoil:

"For desktops, I think right now, with this generation, we could help with spot cooling. For Gen 5 SSDs, if your main fan-based cooling system is taking care of the CPU cooling, you can put one of these on top of a Gen 5 NVMe controller, and get relief."

So, it seems that as things currently stand, even several AirJets don't seem capable of handling something like a high-powered modern desktop CPU any better than a good old-fashioned heatsink and fan. Still, the suggestion of a thin, efficient Gen 5 SSD cooler compared to some of the sizeable models I saw at this years show seems like a tangible use case for the tech as it exists today.

To this point, I was shown an external SSD with two AirJet Minis integrated into the chassis. A passively cooled version dropped to a 1,061.77 MB/s write speed over sustained usage, whereas the AirJet-cooled sample was shown to maintain a consistent 3,153.98 MB/s rate.

Roughly three times the performance is not to be sniffed at, although the form factor could do with some work. The prototype I was shown had the output vents situated on the side, meaning it was blowing uncomfortably hot air right where you'd put your hand to pick it up. A production model was suggested with the vents on the end of the device, although I think it'd take some getting used to an external drive venting large amounts of hot air, no matter how quietly or efficiently.

So then, the million dollar question: Given the inevitable R&D and manufacturing setup costs, how much does an AirJet Mini eventually cost per unit?

"It's reasonable," says Dr Madhavapeddy. But reasonable in comparison to the costs of manufacturing a good fan?

"No. No it's not. But reasonable compared to the overall cost of the product that you purchase as a consumer. We fit into the structure of the costs that typically can be borne by anybody who's making a device of this nature.

"But if we compare head to head with a fan, we will be more expensive. [The price] would come down with scale. It will also come down because our second generation, third generation are going to be more high performance.

"We wouldn't expect the cost to go up at the same rate as the performance is going up," Madhavapeddy continues. "So the performance, the cost per watt, will come down with subsequent generations. And also, in any given generation, the cost comes down with volume."

Speaking of future developments, the company has a credo that, on the surface, seems somewhat outlandish. It's called "Frore's Law", a play on Moore's law, which in this case is the goal to double the thermal performance of the unit every two years.

"We're not bound by the limitations of mechanical devices for how we can make it more and more powerful," says Madhavapeddy. "We could increase the performance by increasing the frequency of vibration, we can also increase the performance by increasing the amplitude (of the vibrations).

"So we have several levers that we can continue to improve on to significantly improve the thermal dissipation capability of the chip"

That's an ambitious goal, and one that seems more than a little unfeasible. That being said, the units I was shown seem much more ready for prime-time than I was expecting. If they really can be scaled up in performance at that kind of a rate, it does seem like more adoption would follow.

While the AirJet Mini Slim might not be the solution for all cooling needs as it currently stands, what it does seem to do well is move significant amounts of air in form factors that are unsuitable for a traditional cooler.

You won't be seeing one in a desktop PC anytime soon, at least on something substantial like a GPU or CPU. But in ultra-slim laptops, tablets, phones and the like, or atop an SSD? The AirJet, at least in Mini Slim form, does seem like a product that's ready to go. Whether mass-adoption follows, however, we'll just have to be a little more patient to find out.

Asus ROG Zephyris G16 | RTX 4070 | Core Ultra 9 185H | 16-inch | 1600p | 240 Hz | OLED | 16GB LPDDR5X-6400 | 1TB SSD | $1,999.99 $1,599.99 at Best Buy (save $400)
While it's a bit pricey for an RTX 4070 laptop, that 16-inch 240 Hz OLED panel is stunning to look at and super smooth to game on. And the hardware behind it all is more than up to the job.View Deal

In the vast ocean of Intel Raptor Lake-powered gaming laptops, it's a rare sight to see one sporting the latest Meteor Lake processors. This 2024 edition Asus ROG Zephyrus G16 boasts a Core Ultra 9 185H chip, with six P-cores, eight E-cores, and 22 threads in total. With a max power consumption of 115 W, it's easier to keep cool than the Raptor Lake laptop processors, too.

It would be nicer to have more P-cores, but they use lots of power, and six is enough for most games. There's even a little NPU (neural processing unit) inside the CPU, to help with basic AI tasks like Microsoft's Copilot, but it's not powerful enough to meet the Copilot+ PC requirements.

Besides, most of the performance will be controlled by the RTX 4070 graphics chip and even though it's limited to a maximum of 105 W, that's fine for 1080p gaming.

Yes, the display has a resolution of 2560 x 1600, so you're more likely to be playing at 1440p+ resolutions but don't worry—the latest games support the use of upscaling and many have a frame generation option, as well. That RTX 4070 is DLSS 3.5 compatible, so you get the full gamut of Nvidia's AI performance-boosting tech.

We tested an RTX 4090 version of this laptop and the GPU in that was massively underused, as the cooling system just couldn't cope with that chip. Here, the compact RTX 4070 will be far easier to manage, so you should be able to squeeze the last drop of rendering power out of that chip.

The real star of the show is the 240 Hz OLED panel—"a beautifully deep and vibrant viewing experience," we said in our review. Like all such displays, it's not super bright, so you'll need to avoid using the laptop in a very well-lit room, but in a darkened gaming den, it's an absolute joy to behold.

There's only 16GB of RAM, unfortunately, but at least it's LPDDR5X-6400, so performance won't be an issue. There's a decent-sized 1TB NVMe PCIe 4.0 SSD inside, so you're unlikely to run out of storage space for a while, though it's not clear if there's a second M.2 to add another drive.

Overall, this is a fine gaming notebook—a little bit on the pricey side compared with some other RTX 4070 laptop deals and it doesn't sport the most powerful CPU and GPU combination you can have—but the balance of components is spot on. You'll certainly love staring for many a long hour at the OLED display, that's for sure.

The only blot in the ROG Zephyrus G16 exercise book is the fact that it's an Asus laptop and the company's management of warranty claims and RMAs has triggered an investigation by Gamers Nexus. Hopefully, the outcome of that will ensure that if you do pick up this great deal, you'll have a solid warranty behind it.

Mesh systems work by having more than one router, or node, scattered around your property. They all broadcast the same SSIDs, and communicate with one another over Ethernet, if your house is wired for it, or using a high-capacity wireless network separate from the usual ones you'll use for streaming and that Wi-Fi kettle you bought on a whim. That can certainly be more convenient as long as you get a good connection. Placing mesh nodes, however, especially if you're trying to eliminate Wi-Fi dead spots caused by quirks of architecture or distance, is an artform in itself.

The thing about the Eero Max 7, though, is that one is probably enough. There's definitely a case for mesh if you're connecting an office environment, a house split over multiple floors, a thick-walled basement or a farm, but Wi-Fi 7 itself, with its extra speed and ability to cut through interference, is pretty good on its own. That said, though, put it in a mesh system and you've pretty much got something that can make itself heard anywhere.

The Eero Max 7 node comes as a smooth and shiny white rectangle that wouldn't be out of place as an Ikea wall light or Apple product from the time of the Anglepoise iMac. The Eero logo is picked out in silver, with a single coloured LED to tell you the router's status, and it's a minimal, classy design you won't mind having on display.

There are no visible antennas or multiple blinking lights, it just sits there being completely silent and unobtrusive—which is what you want if you're planning on dotting a few of them around your enormous property. It is kind of large though, which is something Wi-Fi 7 routers all seem to have in common. If space is a concern Eero has much smaller Wi-Fi 6 mesh nodes available.

The back is more interesting, although there's still a lot of smooth white plastic on show with cooling vents at the top and bottom. It's where the Ethernet ports are and Ethernet ports are interesting, dammit, especially when they're as fast as these. Eero has gone for some definite future-proofing here (though with Wi-Fi 7 we wonder how many cables these sockets will actually see, beyond the one that hooks up to the internet in home use) with a pair of 2.5Gb ports next to two more 10Gb connections. That's a lot of networkingness, and you'll need to make sure any cables and hubs you're using are rated for the same speeds to get the most out of it. 

But you don't buy a Wi-Fi 7 router just to use Ethernet cables, and the Wi-Fi options here are comprehensive. Setup is straightforward, but assumes you're tech savvy enough to find and download the phone app yourself—there's no printed guide in the box or QR code on the node itself. 

Once you've got it hooked up there aren't a lot of fine-grained controls on offer, which is generally fine as you can just leave it to get on with the job of running itself and transferring your data. The Max 7 comes with a trial subscription to the Eero Plus security package, which adds a VPN, MalwareBytes protection, content filters, ad blocking and enhanced support to the package. It costs $100 a year, though, and provides many things you could do yourself. Still, with the Eero's commitment to simplicity and its hands-off approach to router management, I can see it being popular.

Wireless speeds are, of course, promised to be spectacular. The Max 7 can theoretically handle up to 574 Mbps on the 2.4 GHz band, up to 8,677 Mbps on the 5 GHz band, and up to 11,529 Mbps on the 6 GHz band. That's a grand total of just over 20,000 Mbps if you take your shoes and socks off and add them all together. Though it's obviously something you'll never achieve in reality, it's still a lovely big number. 

This, of course, assumes perfect conditions and no inconvenient walls, but while even the 2.4GHz band, which is usually the longer-ranged of the trio thanks to the way it penetrates solid objects better, can outpace the average internet connection it doesn't offer the same (claimed) high speeds as the Netgear Nighthawk RS700S.

In our testing, which involves transferring a gig of data at varying distances over the Wi-Fi and working out the transfer rate using our fingers and toes, the Eero Max 7 put up a good fight, giving an average transfer rate over Wi-Fi 7 of around 570 Mbits/s whether it was in the same room or had a floor and a few walls in the way. This rate wandered, however, being as high as 624 Mbits in one test and as low as 499 Mbits in another. Its fastest rate was lower than that of the Netgear RS700S, but its slowest rate was better. It's the networking equivalent of a score draw.

Such is its commitment to simplicity that the Max 7 combines all its SSIDs into one, and doesn't allow you to separate them, relying on the devices to negotiate the best connection they can. There's a guest network, however, so you can hive some users off that way.

There's nothing else on the Max 7 beyond a USB-C power connection and a WPS button, but what is noticeable by its absence is the otherwise ubiquitous USB port, which allows you to set up simple file sharing through a flash drive (or USB hard drive for a multi-device backup solution), or share a USB printer, just by plugging it in. It's a feature that probably doesn't get a lot of use, especially as network storage and Wi-Fi printers are now commonplace, but having had USB ports on routers since the days before Wi-Fi 4, it seems wrong not to see one. 

Being an Amazon device—the well-known online grocer is Eero's parent company—there's plentiful IoT support in the Max 7. It will pair with your Echo and contains the necessary radios to control Zigbee and Matter devices, sidestepping the need for an extra hub. 

The question that arises with these high-speed, high-capacity routers, however, is just who are they for? These kinds of network speeds, especially over wired connections, are only really useful if you're transferring a lot of data to and from a server. Perhaps one day when we're all streaming 8K and have AI edge servers running our homes it'll be different, but a PC gamer wanting to share a sub-gigabit internet connection between their rig, phone, laptop and tablet is going to get the same sort of results from a much cheaper router, and might get a USB port too. 

With its speed, capacity, future-proofing and the ability to add extra nodes if your house suddenly gets bigger, the Eero Max 7 is one of the best Wi-Fi routers it's currently possible to buy, but you'll need to pay for the privilege.

Demakes of the latest popular 3D games are often designed to emulate how they might look on hardware a few years old, sometimes a couple of decades. Taking Baldur's Gate 3 and getting it to run on something that's over 40 years old, though, is on a whole new level—but that's exactly what one ingenious software engineer has managed to do.

Strictly speaking, the project isn't an exact clone of BG3, nor could it be, as the hardware in question is a TRS-80 Model 100 computer—arguably the world's first laptop. With an 8-bit processor running at 2.4 MHz, a few kB of RAM, and a built-in 240 x 64 LCD display, there's absolutely no chance of running one of the most hardware-demanding games out there.

Alex Bowen created Mol (as reported by Hackaday) as a piece of 'interactive fanart', a 1980s-era tribute to the first act of Baldur's Gate 3. So it's not a line-by-line perfect recreation of the game but if you watch the video above, many of the main events in the first act have been recreated.

Despite the limitations of the display, there's an undeniable charm to the simplistic sprites, and I'm sure BG3 fans will easily recognise the various locations, thanks to handy (or should that be Tandy?) text descriptions alongside them.

As someone who got into games and programming in the early 1980s on a Z80-powered home computer, I immediately got floods of nostalgia, especially seeing the assembly code used to create the demake. Where my first system only had 1kB of RAM, Bowen needed 24kB for the engine, basic sprites, and text.

Even that wasn't enough, so some ingenious tricks needed to be employed to get BG3's mountain of information, descriptions, and dialogue into a compact enough format to squeeze into the TRS-80's RAM. Today's games are hugely complex and require thousands of kB of memory to run, but most gaming PCs sport far more RAM than they'd ever need.

Forty years ago, that definitely wasn't the case, especially so when you're talking about a laptop. The TRS-80 Model perhaps wasn't the very first laptop you could buy but it was certainly the first that was genuinely usable and it sold in great quantities.

Bowen developed Mol on the back of another project, Dungeon Delver Engine, which lets you create RPGs based on the 5.1 version of Dungeons & Dragons OGL-SRD. The code for both projects is hosted on GitHub, so if you just happen to have an old TRS-80 lying around, and you're familiar with zasm, then dust off your programming chops and get stuck in.