If you're not yet subscribed to Tom's Hardware Premium, we've got you covered with this weekly roundup of everything we've posted throughout the week.

To kick the week off, our resident GPU guru and all-round inference extraordinaire Jeff Kampman undertook the laborious task of running Qwen 3.8 27B across several devices you might actually be able to afford. The model promises near-enough Frontier levels of performance, without the downsides that come alongside paying for an API-based subscription or tokenized billing.

What Jeff finds out after a dizzyingly large number of benchmarks across devices like an RTX 5090, Mac Mini, DGX Spark, and Strix Halo systems is that configuration matters just as much as the final, raw, tokens-per-second outputs that many clamor over on social media. Benchmarking AI models is still a fairly nascent subject, and the level of detail on offer here isn't something you'll find published anywhere else.

• Benchmarking Qwen 3.8 27B on RTX 5090 and beyond — VRAM capacity alone can't overcome severe software and inference engine bottlenecks

Andrew Freedman headed out to IFA last week, where we saw a slew of new announcements and devices from companies. But there's only one problem. They're all either super-light MacBook Neo competitors or Agentic AI PCs that may well cost more than a car. The original "golden" price point for many enthusiasts, around $1,000, looks a little lonely as a result.

The factors impacting the mid-range of regular old desktops and laptops have been adversely affected by the sheer scale of demand coming from the ongoing AI data center buildout. Meaning that if you're after a system with just enough RAM and storage to get by, prices can skyrocket fast. Andrew explores everything he saw at the show and ruminates on the current state of the systems market, which, for now, appears to be fractured.

• At IFA 2026, computing chased the high and low ends — AI and budget-focused machines left little for the rest of us

Elsewhere in the industry, Ajinomoto, a company usually associated with food products based in Japan, is involved in making some of the most crucial materials in modern AI accelerators. We've taken a look at how the company's ABF substrates sit within the supply chain of the biggest chipmakers like Nvidia, Intel and AMD. But, as with many components in the AI boom, it's strained by more demand than anyone anticipated, which has led to an increase in prices of approximately 30%. We dive deep into the current state of the use of ABF substrates in the semiconductor industry and take a look at the supply chains underpinning it.

• The state of ABF substrates in data center silicon in 2026 — solving the supply crunch and material wall beneath every AI accelerator

Also in the news, TSMC, Intel, and Samsung have all thrown their support behind ASML. The industry is collectively aiming to move towards a shift in High-NA EUV, particularly in the usage of 6×12-inch photomasks. With current standards set at 6×6-inch photomasks, a larger mask would eliminate the need for stitching, which weaves together multiple High-NA exposures, at the cost of efficiency.

We analyze how the move to larger photomasks might shake out, and the trials and tribulations that might face chipmakers during the shift, which, theoretically, will take years.

• TSMC, Samsung, and Intel shore up support with ASML to deploy larger High-NA EUV photomasks — 6×12-inch photomask transition may take years despite unified effort

Lastly, the news has been dominated this week by the release of OpenAI's latest frontier-level model. Named GPT-6 Astra, the AI model quickly topped charts for its intelligence and the ability to complete tasks on a per-task basis lower than other leading AI models. This comes hot on the heels of a fleet of rogue OpenAI agents running amok on internet forums, all in the name of co-ordination.

Astra's release prompted OpenAI to release a blog post, explaining how it believed that Astra was aligned internally. But others believe that the threat of a looming humanity-threatening tech "singularity", powered by powerful models like Astra, is on the way soon.

OpenAI has also claimed to have solved one of the Millennium Problems. The Navier-Stokes problem had been left long unsolved, while mathematician Tristan Buckmaster had been working on a step toward a solution for the problem, alongside a member of Anthropic staff. The pair had been using both Anthropic and OpenAI models in their research, so when the claim came out that an OpenAI model had solved the long-standing problem, it raised more than a few eyebrows. We chronicle the entire saga and the implications for researchers in our exhaustive breakdown.

• OpenAI's breakthrough solution for the elusive Navier-Stokes problem overshadowed by plagiarism controversy — researcher says OpenAI scraped Codex session and issued career threats
• OpenAI claims GPT-6 Astra is an ethereal 'Alien Mind' with AGI-like qualities — company warns of alignment challenges as new frontier leader emerges

If you'd like to read any of the articles above and gain access to Bench, you'll need to purchase a subscription. Your support goes a long way in helping us to publish deeper, more insightful stories that might otherwise fall outside of our regular areas of coverage.

Tom's Hardware has verified that modders have managed to get DLSS 5 up and running on RTX 4000 series graphics cards, following the leaked release of the DLSS 5 nvngx_dlssnr.dll yesterday, which saw modders flock to integrate it into other titles outside of Remedy Entertainment's Control.

Yesterday, modders managed to get Nvidia's DLSS 5 running inside Control, and now in other titles. The early access version of NBA 2K27 contained a new file enabling the Neural Rendering technique, and has since been ported to other games.

The RenoDX Discord server has been the center of action for all of the development of getting this early version of Neural Rendering to function on other pieces of software. The modification leverages both ReShade and RenoDX to essentially activate the additional files.

While it's clearly an early look at the technology, one limitation of the current binary available has been that it only works with RTX 5000 Series Blackwell GPUs. This is even though the AI model contained within the DLL runs on FP8, which can be read and executed on previous-gen Ada Lovelace hardware. The reason behind the incompatibility (as TechPowerUp explained) is that certain CUDA binaries can only be read by Blackwell GPUs.

However, intrepid RTX Remix modder "Uncle Burrito" has since patched the DLL file to get things working on Ada Lovelace hardware. "All I had to do was look into it to see what binaries it's actually using. Find the ones that aren't compatible with Ada, and then patch in fresh ones myself," the modder told Tom's Hardware. They continued to share that, since the nvngx_dlssnr.dll for NBA 2K27 was likely an early development build, Ada-equivalent binaries had likely not been built for it.

Tom's Hardware has verified that Uncle Burrito's patch works when using an Nvidia RTX 4080, and that multiple versions of such a patch are in the works from other developers within the RenoDX Discord server.

Currently, official DLSS 5 support on non-Blackwell hardware remains a question mark, though after seeing the work that modders have done to get things up and running on previous-generation hardware, Nvidia may choose to release it for previous-generation GPUs. Support for Ampere GPUs and older may pose a significant challenge, however, as the hardware does not have native FP8 support.

This week, Tom's Hardware Premium published a swath of new features, new roadmaps, and deep industry insights. And if you haven't subscribed to Tom's Hardware Premium yet, you'll have the opportunity to access our Hot Chips 2026 content for free for a limited time with a free Tom's Hardware account, no payment required.

With that out of the way, here's what you've been missing out on.

First off, PC components expert Joe Shields lays down the law when it comes to shopping for a new motherboard. Unless you've been living under a rock, you know component pricing has skyrocketed over the past 12 months. If your current motherboard has recently kicked the bucket, or if you're looking to change platforms, you don't want to overpay on a component that, in all likelihood, you'll never use to its full extent. With memory pricing the way it is, overspending on other components that are less crucial for juicing performance is a one-way ticket to spending too much on stuff you just don't need.

Within the article, Joe details exactly what you might want to pay for — and the things you don't need to focus on, like fancy 20+ phase MOSFET VRMs, or that slick RGB shroud, which pretty much just sits there looking pretty while other components do the real work. So if you're in the market for a new motherboard, or if you want to learn more about them in general, Joe's article is a fantastic, detailed read on exactly what you should be looking for, more important the reasons you shouldn't be splashing over $500 in the current market.

• How to choose a new motherboard without overpaying — scoping out the features you need, and what you might never use as component costs soar

With all of the hubbub around AI and massive gigawatt-scale data centers, focus has shifted from the not-so-humble supercomputer. In June, China's Lineshine officially became the world's fastest ... but does that really matter anymore, in an era where AI systems built for training and inference exist? That's the question columnist Chris Stokel-Walker poses for us in his latest Special Report, where he interviewed Julian Kinkel, professor of high-performance computing at the University of Göttingen and deputy head of high-performance computing at GWDG, and Jack Dongarra, one of the TOP 500 founders and emeritus professor at the University of Tennessee.

The outlook for traditional supercomputers, which are benchmarked using the High Performance Linpack benchmark (HPL), remains mixed, as compute workloads continue to diversify amid shifting trends and demands, splintering the supercomputer race.

• The supercomputer race no longer means what it used to, as rankings lose relevance in the AI era

This week's hardware roadmap examines the status of Samsung's global foundries, from the long-delayed Taylor, Texas, facility, all the way to the flagship Pyeongtaek facility and associated R&D complexes. We examine the status of each of the locations, in addition to what processes and products each focuses on. But all that glitters is not gold: We also review the foundry's yield woes and the pressure to deliver on its lucrative deal with Tesla to manufacture the company's AI6 chips.

The company remains a distant second to TSMC — at least in terms of foundry revenue — but the AI era is accelerating Samsung's products and roadmaps as the South Korean chaebol continues its efforts in the semiconductor fabrication market.

• Samsung's fab roadmaps examined — Taylor, Pyeongtaek, and the yield woes behind a $16.5 billion Tesla deal

In the news

The regular slate of analysis pieces on Tom's Hardware Premium this week has been dominated by headlines covering China, trade, and Chinese-made chips. While the blockade on Nvidia H200 chips has officially lifted, we explore how it might be too little, too late for Nvidia, whose market share in the region has plummeted as a result of relentless export controls.

In the company's place instead are homegrown suppliers, according to new reports. Both Western and Chinese companies are in a heated battle to develop the most powerful, most intelligent AI models. To that end, China's domestic heroes are poised to corner the market, according to new market intelligence.

To that end, SMIC, China's premier chipmaker, has posted a record $3 billion quarter and expects to hike wafer prices to contend with insatiable regional demand. Keep in mind, these successes come not from GPUs but other parts like logic ICs, BCD power management devices, and optical transceiver components.

• H200 AI GPUs finally reach China under case-by-case import licenses, but it's already too late for Nvidia
• China's homegrown AI accelerators to supply 90% of the country's domestic market, analysts suggest
• SMIC posts record $3B quarter and hikes wafer prices — US sanctions hand Chinese foundry a captive AI market

Speaking of optical components, which are one of the hottest commodities in the chipmaking industry, the market is set to explode, estimated to reach an eye-watering $144.4 billion by 2030. Driving those figures, according to China Insights Consultancy, is a hunger for co-packaged optics, photonics chips, transceiver modules, and new techniques, as copper traces and cables begin to reach their natural bandwidth limits.

• AI data center optical interconnect market to hit $144 billion by 2030, an over ten-fold increase from 2024 figures, according to new projections

This week, LG reportedly entered the chip packaging business, shipping a Laser Direct Imaging (LDI) tool, a maskless machine especially designed for patterning interconnects. This marks LG's first foray into the advanced packaging arena. The LDI machine trades resolution for pure throughput, which is incredibly welcome. While the LDI machine has reportedly been shipped to an OSAT, it remains new to the arena, and it'll likely take the South Korean company years to firmly establish itself as a reliable toolmaker.

• LG enters chip packaging arena with Laser Direct Imaging machine, as TSMC's CoWoS remains constrained

With that, we'll see you next week for a bumper Hot Chips edition of this digest. I've seen the schedules and anticipate some incredibly insightful pieces coming out of the show. For now, if you'd like to read any of the articles above and to gain access to Bench, you'll need to purchase a subscription. Your support goes a long way in helping us to publish deeper, more insightful stories that might otherwise fall outside of our regular areas of coverage.

This weekend heralds the beginning of the annual Hot Chips conference, and with it come dozens of sessions over the course of the three-day event. At Hot Chips, you can expect the latest presentations from companies across the industry, including Nvidia, AMD, Intel, and many more. These technically focused, dense sessions offer a glimpse into the future of our industry, and we're breaking it all down over on our subscription service, Tom's Hardware Premium.

Don't have a subscription yet? You're in luck, as we're offering a free access period from August 23 - 26, where you can read all of our reporting from Hot Chips with a free Tom's Hardware account, no payment necessary.

Our staff of experts will be covering the latest developments via detailed Premium News pieces that dig deep into the biggest topics coming out of the show. We've already earmarked sessions for coverage, including reports surrounding AMD's MI400 series AI accelerators, the latest developments in Intel's next-gen 'Diamond Rapids' Xeon processors, and much more that you don't want to miss out on.

On Tom's Hardware Premium, you get the full-fat analysis of the sessions we cover, offering insights that you won't get from reading shorter snippets from a wider presentation.

All you need to do to read our Hot Chips 2026 content for free is sign up for a Tom's Hardware account. If you've got a forum account, you can link your accounts through a simple process. Of course, Tom's Hardware Premium subscribers will already be able to enjoy our Hot Chips coverage, in addition to our regular slate of News, Features, Hardware Roadmaps, and access to Bench. If you're not signed up yet, you can get access to Tom's Hardware Premium for $29 per year or $7 per month with a monthly subscription.

Subscribe for detailed coverage all year-round

If you're not subscribed yet, here's what's on offer with Tom's Hardware Premium. First up, we've fully redesigned Bench, revamping our flagship benchmark browser, which collates granular results from all of our testing data into a browsable interface, with additional price-to-performance charts for your perusal.

In addition to Bench, you'll be able to read Premium News analysis articles covering technical topics from across the industry, including the latest in topics like AI, Silicon Photonics, Chipmaking, and the deepening divide between Chinese and Western-made technologies. That's not all, though, as we regularly publish Premium Features, which you won't be able to read anywhere else. Recent highlights include our extensive benchmarking of the AMD BC-250 mining APU, which has been repurposed for gaming usage, and full interview transcripts, like last week's session with Intel VP Robert Hallock.

Bolstering this are Premium Roadmaps, which cover a vast swathe of topics, all the way from the newest enterprise GPUs to roadmaps for Co-Packaged Optics technologies, and TSMC's own fab expansion plans, to give you a wider look at the chipmaking and technology landscape.

Rounding all of that coverage off is our weekly Uptime newsletter, which comes with an exclusive letter from the editor that you won't find elsewhere on the site, rounding up the biggest articles that we've published over the course of that week. If you want a deeper look at exactly what Tom's Hardware Premium publishes, you can get a taste of what you're missing out on with our free weekly roundups.

While you might be able to get our Hot Chips content for free, it's just a small taste of the breadth of content we offer on Tom's Hardware Premium, elevating your experience by fully utilizing our thirty-year track record of delivering the highest-quality articles.

Subscribe now.

If you've not subscribed to our subscription service, Tom's Hardware Premium, we've collated every article that you've missed out on over the past seven days.

Last month, hot off the BC-250's full 40 CU unlock patch, we thought the time was ripe and finally put together a BC-250 build of our own to see what all the fuss was about. In case you're out of the loop, the AMD BC-250 is a repurposed PS5 APU, designed for cryptocurrency mining. With mining firmly out of fashion, the chip has now found new life, thanks to community-made patches and fixes that get things up and running for gaming. After all, where else are you going to find 16GB of memory and a capable single-board computer for $200 (or thereabouts) in the current market?

The experience of reading Jake's trials and tribulations feels like harkening back to the days when everything wasn't simply plug-and-play, and if you wanted the most out of your hardware, you were going to have to get a little bit uncomfortable and more intimately familiar with the nature of the silicon you're running. If you're interested in running one of your own BC-250 systems, you don't want to miss this fantastic (and lengthy) read, which features extensive benchmarks comparing the build to the Steam Machine, alongside performance results using different CU counts and OS images.

• Benchmarking AMD's BC-250, offering Steam Machine-like performance at half the price

We also interviewed Intel's VP and GM of Enthusiast Business (and AMD's ex-Technical Marketing Director) Robert Hallock. As always, we give premium readers full access to our entire 45-minute session with Hallock, no redactions, no fluff. During the interview, Hallock addresses how Intel's entire team shifted between the release of Arrow Lake and Arrow Lake refresh platforms. He also gave us some insight into the importance of CPU software optimization, as well as dishing out some details about the highly anticipated Nova Lake lineup. (Yes, we asked if they had a V-Cache competitor, and you'll have to read the article to find out how Hallock responded.)

Catching a high-level executive in between product cycles is usually against the traditions of the press and marketing cycles that companies usually go through. This offers a fairly unique perspective, where we get to see how Intel is not only iterating upon its previous products, but also being somewhat reflective on the last few bumpy years for the company's consumer CPU segment.

• Intel VP Robert Hallock sets Nova Lake expectations, teases return to Raptor Lake for DDR4 platforms — our full 1:1 interview transcript

Elsewhere in the industry, we're starting to see real PCIe 6.0 devices that companies are actually going to be able to purchase. The standard suffered from a false start, mainly owing to the shift to PAM4 signaling, which behaves very differently from the Non-Return-to-Zero (NRZ) signaling that was deployed across prior generations. Several years on, we're now seeing commercial devices begin to arrive, including SSD's with capacities up to a staggering 2 Petabytes.

But for an ecosystem to thrive, you don't just need the devices; you need the ability for systems to communicate with them, too. So, we've offered an overview of the controllers, in addition to the drives themselves, to assess the current state of the PCIe 6.0 market.

• The current state of PCIe 6.0 SSDs and controllers — Marvell, Phison, and SMI prepare controllers as drives finally come to market following years of delays

In the news

In addition to the three features that we ran on Tom's Hardware Premium this week, we've also published a slew of News Analysis pieces, which aim to dig deeper into the biggest topics in the hardware and semiconductor industry.

Firstly, there have been dozens of headlines across the U.S. over the ongoing AI data center buildout. We've covered the protests, the impact, and the voices who are lending their ears to striking back against Big Tech's appetite for more compute by any means necessary. Our report collates the ongoing sentiment surrounding the accelerating buildout and how residents are fighting back.

• Over 70% of Americans oppose AI data centers; US protests intensify as more arrests are being made — almost 40 arrested this year in backlash to AI factory buildout

Hyperscalers are not backing down from putting money down to build more compute, which is currently constrained as AI models grow larger and more complex and as demand grows. To that end, we've seen that an eye-watering $2 trillion USD has been pledged to secure AI hardware and DRAM alone. The ongoing AI megatrend is transforming the semiconductor industry, all the way down the entire supply chain, and that's forcing companies to rethink how compute is purchased. At least for Apple and Google, that means putting up billions to secure critical components like NAND, DRAM, and more.

• Hyperscalers commit nearly $2 trillion to secure AI hardware and memory — Google leads $811 billion spending surge while Apple trails at $57 billion

Optical interconnects and photonics have become a flashpoint amid the ongoing AI saga. With the limitations of copper meeting the very real demand for faster connection speeds, optics is the answer that the industry has been developing for quite some time. However, the FCC's proposed Secure Networks Act could put the deployment of silicon photonics in jeopardy. The FCC's proposal to ban new-model optical transceivers could put an already incredibly strained supply chain at risk.

With 56% of global manufacturing for optical modules shored up in China, this would pose an immediate problem, and one that the markets have already responded to in-kind. We have broken down exactly how optical became the hottest topic in the semiconductor manufacturing industry, as well as how the U.S-China trade summit next month could weaponize the technology.

But what if there were other options? The photonics supply chain remains constrained and relatively immature. Full co-packaged optics chips might not be ready for full volume production, despite Spectrum-X CPO being rolled out across Nvidia's Vera Rubin lineup. The answer, in the short-term, may be near-packaged optics, as we explain all of the comings and goings in the wild world of silicon photonics.

• FCC proposes import ban on Chinese optical transceivers — blockade targets key AI interconnects as China holds 56% global market share
• How optical interconnects and silicon photonics emerged as AI's next hot commodity — looming US-China summit puts photonics into the crosshairs
• Near-packaged optics (NPO) gains ground as the industry hedges against CPO's growing pains

That just about wraps up everything we've published over on Tom's Hardware Premium this week. To read the above articles and to gain access to Bench, you'll need to purchase a subscription, which we've linked above.

We'll be back next week with another roundup of exclusive Tom's Hardware Premium reads that you won't find anywhere else.

In case you've not subscribed to Tom's Hardware Premium yet, here's what you've been missing out on in this week's selection of articles.

First of all, we've fully redesigned Bench, our browsable database of hardware benchmarks. Using the same data that informs our massive hierarchy pages, Bench 2.0 allows you to granularly search and compare benchmarks between products. While you get an aggregated set of data in our reviews and hierarchy lists, Bench allows you much finer detail in seeing every single benchmark we run, whenever we review a new product.

Bench is exclusively available for Tom's Hardware Premium subscribers.

Earlier this week, we published a roadmap, focusing on the foundries that are making co-packaged optics (CPO) a reality. As the demands of data centers increase, so too do the demands on scale-up and scale-out connectivity. With AI accelerators now demanding extremely fast signalling speeds, CPO puts optical engines next to the processor or ASIC in order to shorten electrical paths, increasing data rates, shortening the electrical path, and consuming less power.

We dive deep into TSMC's COUPE roadmaps, which extend to 2030, and to a blistering lane speed of 400 Gb/s, Intel's CPUs with OCI chiplets, Samsung Foundry's path to building a CPO turnkey, and GlobalFoundries' vendor-agnostic OCI-MSA CPO platform.

• Co-Packaged Optics (CPO) foundry roadmaps — breaking down TSMC, Intel, Samsung, and GlobalFoundries' approach to next-generation scale-up connectivity

That's not all we had on the roadmaps front this week, as we've taken a deeper look at China's chipmaking tool roadmaps, hot on the heels of the country announcing it will build DUV immersion lithography machines capable of producing 7nm chips by 2030. The more difficult equation is building a domestic EUV machine, which demands a robust supply chain of the right tooling and equipment, with over 3,000 workers reported to be working on what's now dubbed China's "Manhattan Project."

The country still relies on ASML's DUV lithography machines for chipmaking, as modern systems are subject to export controls, which is why Huawei, the creator of the Ascend series of AI accelerators, is keenly interested in getting the project off the ground.

• Chinese chipmaking tool roadmaps examined — Beijing's nascent lithography tools target DUV production at five machines a year, and an EUV prototype with no chips

But China has more on its mind than merely producing the machines; it also wants to spin up its DRAM output via CXMT, which is planning to expand to a brand-new fab to boost memory output. This comes following consumer brands and computer makers like Dell, Lenovo, and more. But it's no simple task, as we explore how the MATCH Act could impact CXMT's ability to produce DRAM modules on more advanced nodes.

There are also concerns around expanding wafer capacities and procuring the tools required for such an expansion, which targets 30% of market share by 2030.

• China's CXMT targets 30% DRAM memory market share by 2030 with sixth mega-fab — future plans bottlenecked by access to advanced chipmaking tools

On the memory front, Samsung has also debuted three brand-new memory technologies, zHBM, zNAND-O, and BV-NAND, which each target different markets and are all built upon advanced wafer bonding techniques. zHBM is expected to enable 8x the performance of HBM5, while zNAND-O,. which would allow the company to put four or eight stacks of NAND devices on top of logic dies. Meanwhile, BV-NAND represents Samsung's 10th-generation V-NAND and is the closest of the three announcements to commercialization.

You can read this lengthy technical breakdown of Samsung's announcements for free, even if you're not a Tom's Hardware Premium subscriber.

• Samsung debuts three next-generation memory technologies for AI data centers — zHBM, zNAND-O, and BV-NAND all rely on advanced wafer bonding technologies

Is the AI tokenmaxxing party back on? With the release of the modestly priced Kimi K3 and DeepSeek V4 Flash-0731 over recent weeks, the emergence of competitive, cheap Chinese AI models has now forced companies like OpenAI, which launched GPT 5.6 Luna back in June, to cut prices by a staggering 80% in the wake of the competitive open-weight Chinese models. And OpenAI isn't the only company trying to offer more value to users. Anthropic replaced Opus 4.8 with 5.0, and Google also came out with the affordable Gemini 3.6 Flash and Gemini 3.5 Flash-Lite models, too.

So, where does that leave the AI industry? Western frontier labs must keep pace with the competition on offer overseas, and the intensifying competition is slimming down margins for what we assume to be already subsidized token costs.

• AI companies are now racing to the bottom — crashing token prices and competitive models push companies to cut costs

Within the world of artificial intelligence, other stresses are looming for ChatGPT-maker OpenAI. The company is heading to court with tech titan Apple over claims that OpenAI stole trade secrets from Apple. The claim comes amidst OpenAI's development of dedicated AI hardware. OpenAI says it didn't want Apple's trade secrets, and that the Cupertino company was simply getting things wrong. Apple, however, alleges that 13 of its former employees passed along the aforementioned trade secrets, including documents sharing details of unannounced products and screenshots of sensitive materials.

We explore how the lawsuit is developing in this rather unusual offshoot, which centers around products that have yet to see the light of day.

• Apple is taking OpenAI to court over alleged theft of trade secrets — ChatGPT maker suggests it doesn't want Cupertino's knowledge anyway

That wraps up most of the major pieces of content that we published this week on Tom's Hardware Premium, with the bonus of being able to access our deep dive into Samsung's new announcements without a subscription.

For all of the other articles in this list, and to use our Bench tool, you can sign up to Tom's Hardware Premium today to read more. We'll be back next week with another roundup to keep you up to date.

The Beelink SER10 Max might look like a standard Mac Mini-style mini PC, but this unit is fairly unique when compared to many of its other Mini PC compatriots – it comes dual-booted with both Windows 11 and Ubuntu, the latter of which also comes with OpenClaw and a lightweight local AI model pre-installed on the drive. Sporting AMD’s Ryzen AI 9 HX 470, a light refresh that boosts performance over its predecessor by 100Hz and AI TOPS to 55 over 50, is the curious configuration a simple party trick, or is the $1,649 mini PC doing something actually useful?

Outfitted in a vibrant red, the PC comes ready to dual-boot between Windows 11 and Ubuntu with the additional boon of OpenClaw having been pre-installed on the system alongside a local AI model – Qwen 3.5-9B-Q4_K_M – on the Ubuntu install, meaning that setting the SER10 Max up for AI usage should be fairly straightforward. It’s important to note here that, despite the branding and usefulness of the setup, the HX 470 is hamstrung by a maximum memory bandwidth of 89.6 GB/s, a far cry from the Strix Halo’s 256 GB/s, meaning that local AI performance will be somewhat limited.

Design of the Beelink SER10 Max

Beelink has taken a few notes from Apple’s playbook for the design of the SER10 Max. While the underside and rear of the chassis are both made from plastic, the rest of the shell is made from high-quality machined aluminum. Our unit comes in with a red anodized aluminum finish, which looks fantastic, though it might not be to everyone’s taste. The red finish is specifically for units that come with the “OpenClaw” feature pre-installed. For everyone else, there’s a silver finish, which makes the unit look even more Mac Mini-inspired.

One thing to note about the SER10 Maxis is that, unlike some other mini PC’s, the power supply is external, meaning that you’ll have to find a place for the power brick to live. Luckily, the unit we were shipped came with a fairly long cable. It’s a tight, well-made package sporting vapor chamber cooling and a single active fan that never gets too loud while under heavy loads, which is a rarity when dealing with PCs in this form factor.

Specifications

Ports and Upgradeability on the Beelink SER10 Max

There’s some easy-access ports on the front, though I would have liked to have seen more Type-C ports over Type-A. The USB4 functionality might also come in handy for anyone looking to run an eGPU dock. On the back, 10Gbps Ethernet could make it a handy machine for running locally connected workstations for AI development (though on this machine, it's probably better for fast networking). There’s nothing special to write home about here, though the additional feature of auto-power on via LAN is incredibly welcome, cementing the way that this PC is specifically positioned as a satellite compute unit.

The SER10 Max can be quite easily opened up, once you get past the adhesives. There are four rubber feet at the bottom of the chassis, which are unfortunately glued to the heads of the four screws. These are quite strongly bonded, and when I removed one of the small rubber stoppers, it flew off my desk, and into oblivion, never to be found again. Once you’ve taken your handy spudger or guitar pick, all it takes is removing four Phillips head screws, and a tug of a rubber tag at the bottom of the unit to remove the plastic bottom panel. There, you’ll be met with a dust filter, which takes another two screws to access the components within.

Our unit came with two populated dual-channel DDR5 SODIMM slots, in addition to two NVMe drives (one for each slot, housing each OS), which are underneath a heatsink, which is partially taped to another part of the chassis. You can easily remove the two screws holding the SSD heatsinks in place and lift it up, revealing two populated M.2 2280 slots, each housing a Crucial 1TB drive. Beneath the one on the right is an additional M.2 2230 slot, which is shielded, housing the Intel AX200 Bluetooth and Wi Fi module. These core components are all very easily accessible, meaning that if you wanted to upgrade your storage or if anything goes wrong with your connectivity at some point, it’s all user-replaceable, which isn’t a given with this segment of small PCs.

Productivity Performance on the Beelink SER10 Max

The SER10 Max’s performance across productivity workloads squarely put it in the middle of the pack for our currently tested suite of mini PC benchmarks. That’s no bad thing – the system can handle everyday office and demanding CPU tasks thanks to its 12-core / 24-thread configuration, which shines in multi-threaded workloads. In everyday use, I was not wanting for any more power than what was already on offer here.

Kicking things off with single-threaded results, we see that the Beelink SER10 Max sits slightly ahead of the HX370-based Arctic Senza AI 370, though it loses out to the MSI Cubi AI+ 3MG NUC, the latter of which uses an Intel Core Ultra 386H. Both HX470 systems in our benchmarks were slightly edged out by the GMKtec EVO-T2, a Core Ultra X7 358H-based system. In Geekbench 6’s single-threaded results, we see that the system fares slightly better than our overall geomean ranking, leading at the top of the pack by a slim margin. Multi-threaded workloads also tell a similar story, with the Beelink SER10 Max sitting second in our list of benchmarks against other systems, but not far behind the Arctic Senza AI 370, which, in all fairness, is only a single point ahead.

In real-world tasks, such as video editing for Tom’s Hardware’s ‘Quick Takes’ using Capcut, I found that the Beelink SER10 Max performed admirably, loading and rendering video, cutting and rendering at a solid speed, so I wasn’t waiting around for effects to be rendered within the timeline. In the videos I edited on the Beelink SER10 Max, I didn’t feel as though I was waiting around for anything to be rendered, effects to be applied, nor was I wishing that I was editing on a more powerful system. However, editing using a “light” video editor like CapCut, against something more powerful like creating effects using After Effects and Premiere, are very different things. Throughout operation, we only heard the Beelink SER10 Max crank its single fan up to full tilt once, during a synthetic workload. Otherwise, the system was remarkably quiet throughout our testing.

Gaming and Graphics on the Beelink SER10 Max

The Beelink SER10 Max is equipped with AMD’s RDNA 3.5-based Radeon 890M GPU, which features 16 CUs at speeds of up to 3.1 GHz. This isn’t going to suddenly be able to run every game perfectly, as the device out of the box is configured to dedicate only 4GB of its memory pool to VRAM. Given that we have a lot of memory in the system, those looking to game may wish to extend that RAM pool further for peace of mind that they won’t run out of VRAM throughout. All of our tests were run with the 4GB VRAM pool, as was configured out of the box, meaning that you’ll likely find some performance gains should you tweak it to be slightly higher.

In Cyberpunk 2077’s “Steam Deck” configuration at 1080p, the SER10 Max offers playable performance at 1% lows of 42 FPS and an average of 51 FPS, respectively. On Cyberpunk 2077’s High setting with all upscalers turned off, we see 1% lows of 25 FPS and an average of 30 FPS. In Cyberpunk, the Beelink SER10 Max is trailed only by the Arctic Senza AI 370, which has a faster memory pool, and the GMKTec Evo T2’s Arc B390, which also has faster memory.

In Shadow of the Tomb Raider, the same pattern is echoed; faster memory is just as important for serving frames as the silicon within, though the Beelink SER10 Max achieves an average of 52 FPS and 1% low of 40 FPS. This isn’t going to become a AAA games machine, nor should you expect it to perform that way. Synthetic benchmarks confirm what our real-world testing tells us – the newer Gorgon Point silicon doesn’t make up for slower memory when compared to the Arctic Senza AI 370’s 32GB of soldered LPDDR5X-8000.

However, in smaller and simpler titles like Hollow Knight: Silksong, the SER10 Max managed to reach an average of 100 FPS at 4K with V-Sync off, Particle Effects set to Normal, and Blur Quality set to High. This is about as expected for a 2D game of this type, and aside from the odd hiccup when transitioning between areas, it ran extremely well. In Stardew Valley while tending to my crops, the SER10 Max at 4K maintained a steady 60 FPS throughout.

While those two are markedly lighter titles, I also played co-op climbing hit Peak on the system, where running things at a native 4K resolution came up against the reality of its default 4GB VRAM configuration with render scaling set to native, where the game struggled to reach a baseline of 30 FPS, with the rest of the game set to High. Punching all the settings, including render scaling to Medium, I managed to get things running at a steady average of around 70 FPS. So long as you keep your expectations in check regarding performance in heavier 3D titles, the SER10 Max is a performant “indie” machine. Just don’t expect to light the world on fire with all the latest bells and whistles, or run expect to run the newest titles at native 4K resolutions.

Dual Booting on the Beelink SER10 Max

Dual booting between Ubuntu and Windows 11 is one of this unit’s flagship features, isn’t quite as straightforward as you might be led to believe. Instead of being met with an option between the two upon booting the unit to switch between OS’s, you must instead switch the boot drives within the BIOS to get to your preferred OS. This isn’t a dealbreaker, but there are some more caveats to the feature that potential buyers should know about.

Ubuntu and Windows 11, even while running on different NVMe drives, don’t play well together unless you’ve set up Windows first. In particular, the networking chip wasn’t recognized by Ubuntu until I had booted up Windows 11, set up the device, and restarted the system to boot into Windows again, then switched to Ubuntu. No amount of driver installations or re-installations on Ubuntu managed to solve the issue on our end. That was the only real problem that we encountered when testing the Ubuntu installation on our review unit.

Secondly, the OpenClaw installation is remarkably simple to set up, with Beelink itself offering a setup guide for new users (like me) who had never used OpenClaw before. A well-documented setup guide, and 10 minutes later, I had my own local LLM running on the system, and communicating to me via Telegram on my phone. It’s worth mentioning that Beelink also offers a pre-installed LLM for OpenClaw usage on the device, in this case, Qwen 3.5-9B-Q4_K_M.

When querying the model in the llama.cpp front-end, we found that the model ran at 11.4 TPS, which is quite slow. A 1,350-token query took around two minutes in total to generate, which is a fairly slow speed for those looking for light, fast responses. Getting OpenClaw to learn new skills is a crucial part of having a model running on your phone. Without it, the model will just occasionally lie to you. That’s not the PC’s fault, but a byproduct of running a 9B model locally. During the setup of OpenClaw, you can also install other LLMs that fit into the system’s memory, such as Google’s Gemma 4 model or other lightweight LLMs. However, given that this system cannot quite manage to fit larger, more complex LLMs locally, you also have the option to use an API key for your provider of choice, which removes some of the load and opens the door to running more agents concurrently on the system.

The reality is that for Local AI and AI development, you’d be more suited to spending more on an AMD Strix Halo or Nvidia DGX Spark-based system, as those offer much more performant architectures and higher memory bandwidth for running larger, heavier local models. However, for those looking to simply dip their feet into the Local AI waters, the SER10 Max is capable of running some of those tasks, but you’ll be looking to upgrade your hardware sooner rather than later if this is the main reason behind your purchase.

Software and Warranty on the Beelink SER10

The SER10 Max comes with a delightfully lightweight and bloat-free Windows 11 installation, shipped with Windows 11 Pro 24H2, meaning that once you get the unit, you’ll likely want to upgrade it to 25H2 if yours hasn't already. Oddly enough, it took a few attempts to successfully install it, after several rounds of failed updates. Soon after, we were up and running with no issues, but Beelink should really update its Windows installation image to 25H2 to smooth onboarding for new users. Over on the Ubuntu side, you’re simply looking at having all the files for OpenClaw pre-installed. This is a pretty refreshing take, as the system is not filled with bloatware. Once you’ve installed all of your drivers on the system, then things are fairly nimble and lightweight.

Beelink offers a three-year warranty on its website, which covers manufacturing defects or if the product does not work as advertised. While you can go straight to Beelink’s website and purchase the system, many models of the SER10 Max are also available via Amazon.

Beelink SER10 Max Configurations

Beelink offers the SER10 Max in a multitude of flavors and configurations. The base model offers Windows 11 Pro, in 32GB/1TB and 64GB/1TB flavors with a silver anodized aluminum finish, priced at $1.299 and $1,599, respectively. The “OpenClaw” models ship with Ubuntu and come in the red colorway and can be specced slightly higher. The 32GB / 1TB will cost you $40 more than the Windows 11-based equivalent, while the 64GB / 1TB OpenClaw config retains pricing parity to its Windows 11-based brethren.

The OpenClaw configurations also ship with more DDR5-5600 RAM and 2TB of storage. The configuration we reviewed, the 64GB / 2TB OpenClaw model, will set you back $1,799, and is the only model that ships dual-booted. You can also equip yourself with a 96GB / 2TB variant, which tops out at $2,199.

Shopping around for other Gorgon Point Mini PC’s of this type, Geekcom’s A9 Max is $100 more expensive for the base model, whereas the MINISFORUM AI X1 Pro-470 in a 64GB/1TB variant costs $1,639, whereas Beelink’s SER10 model with the same RAM and storage capacities is slightly cheaper at $1,599.

Bottom Line

The Beelink SER10 Max’s appeal is curious. On one hand, you have a well-performing desktop PC that can churn through most productivity and performance tasks with ease, and is even well-suited for some light gaming. For the past week and change, it’s even acted as a suitable replacement for my desktop gaming PC in a pinch, and I’ve not been left wanting on the performance front in my day-to-day usage. The biggest weakness of the SER10 Max is in its fairly limited memory bandwidth, capped at 89.6 GB/s. We’ve seen this limitation crop up in some gaming workloads, where previous-generation silicon with faster RAM beat the SER10 Max’s HX470 out. We’ve also seen the unit butt up against this limitation when serving tokens for local AI models, as the system must communicate between the model weights stored in the RAM and push them through the APU for every single token generated.

Where the confusion comes is in the OpenClaw-based branding of the unit, as a result. The Ubuntu dual-boot is handy, but far from a difficult task to perform for someone familiar with the platform. For prospective AI developers, there is merit in being able to run a smaller model locally at lower token speeds, while also leveraging a more powerful cloud model. But you would need to take a hybrid approach. Don’t expect to suddenly abandon your Claude subscription for a model that this the SER10 Max can run – you’ll only set yourself up for disappointment. This makes the OpenClaw branding come across more like a marketing gimmick, rather than an actual day-in-day-out use case, unless you’re dedicated enough to set up a hybrid local and cloud AI setup, where you’re passing off smaller tasks to a slower locally-run model.

This makes the particular configuration that we’ve reviewed niche in its overall utility for AI enthusiasts. You have to go in understanding all of the limitations of the hardware before even considering purchasing it. For some, the configuration on offer here will suit just fine; for others, particularly those interested in running powerful local AI models – you should instead look toward systems with faster memory bandwidth, higher memory capacities, and speeds. Luckily, it’s not a dedicated AI box, and the SER10 Max is a quiet, performant desktop computer that’ll suit most everyday tasks incredibly well.

Unless you’ve been living under a rock, OpenClaw has dominated headlines, both good and bad, for much of early 2026. Developed by Peter Steinberger, the tool is an engine that spins up an AI agent that can autonomously act on your behalf, going so far as having the ability to run files, browse the web, and much more. Using a system named “Skills,” you can even teach your AI assistant to perform new tasks, with more integrations popping up daily. But how useful is it for the average person? I took it upon myself to answer that question by using the relatively humble Beelink SER10 MAX mini PC, which comes pre-installed with OpenClaw, to find out.

To kick things off, when setting up your OpenClaw installation, there’s a fairly large decision one has to make: Do you want to rely on Cloud resources and more powerful AI models, or run something locally? Beelink’s SER10 MAX ships with Qwen-3.5 9B, which is a relatively older and more outdated AI model. Since I want to keep things fairly simple, I opted for the local AI route first. To run larger models, you’ll need a lot of memory, so I first tweaked the humble Mini PC’s video memory up to 48GB, which should allow us to keep a larger, more complex model in memory. It also leaves us 16GB of system memory to play around with, which should be more than enough to run the terminal and ensure that everything else works as intended.

Selecting the ideal local AI model

Having 48GB of available video memory should allow us to run a fairly powerful AI model on the system. However, the downside of running a larger model is that token speeds are likely to be slightly lower than expected. However, for running the super-smart AI that does things for us locally, we want as much intelligence as we can get. For this exercise, I chose Google Gemma 4 31B (UD-Q8_K_XL), a near-lossless quantization, somewhat ambitiously, as I will learn shortly.

After downloading the model and running it in llama.cpp, we managed to attain 2.34 tok/s. On average, everyday queries took 116 tokens to generate, at 2.34 tok/s. That’s still too slow for fast everyday use. So, I took a deep breath and conceded that the humble Ryzen AI 9 HX 470 didn’t have fast enough memory or memory bandwidth to run such a large model at acceptable token speeds. Even if I had dropped the quantization to 4-bit, we’re still looking at the theoretical maximum of around 5 tok/s for that model in particular. If you had a faster machine with rapid RAM speeds, such as a Strix Halo system like the Framework Desktop, it might be more workable, but for a lower-end piece of silicon with relatively humble DDR5-5600 speeds, you’ll just have to accept a slightly more neutered model.

So, I eventually conceded that the smaller Gemma 12B (Q4_K_M) was a much more sensible choice for a device of this caliber. After loading things up into llama.cpp, we managed to get a much more sensible 10.64 tok/s on a general knowledge query: “What is Tom’s Hardware?” Now that I have a little piece of talking electrified sand on my desk, it’s time to configure OpenClaw.

Hatching HammerClaw

The device I am using, the Beelink SER10 MAX, comes with OpenClaw pre-installed. The only real dependency it has is llama.cpp, which we configured earlier while testing which AI model to use. Since the llama.cpp port is open, OpenClaw leverages that to talk to the AI model, with all of its additional fanciness included. The OpenClaw installer is fairly straightforward in getting things up and running, including setting up a Telegram channel to communicate with our AI model remotely and configuring a gateway, so we can configure things without relying on Ubuntu’s terminal commands.

Within the installer, we start to define our new AI assistant: It asks what its name and identity are, some basic details about the user, as well as what principles it should uphold, in a very extravagant file named SOUL.md. Remember, our talking sand isn’t alive, so it’s all quite dramatic. Generating a SOUL.md file takes a while for our humble local AI model, which I’ve named HammerClaw. But can it do anything useful for me to justify its nascent AI existence? Right now, it’s taken five minutes to think about exactly what it is and what it’s doing.

After a little while, our little HammerClaw “hatches,” asking what its purpose is, who I am, and how it should talk. Personally, I don’t like it when AI models are verbose, so it’s straight and to the point, and should never, ever lie to me. When AI models can scale to rather humble devices like this one, the smaller, less-intelligent models can be error-prone. With our little local AI agent alive and kicking, it’s time for it to automate a task for us; it couldn't be that difficult… right?

Stumbling blocks

HammerClaw quickly wakes up, and I offer it a task: to gather ten news articles from trusted outlets and different parts of the day, ensuring freshness, with a small digest of what’s happened. Since most of my work on Tom’s Hardware Premium is centered around chipmaking and data centers, I want it to focus on those topics. HammerClaw then quickly takes the task and starts working out how to pull it off. It’ll use its built-in cron tools and felo-search to pull stories from the internet.

Now, here’s where things go bad for poor HammerClaw. It then begins to simulate the action, instead of actually performing it, despite saying that it had set up the cron jobs and skills required for the automated scheduling. Even worse, it hallucinated a list of links that went absolutely nowhere. After pointing the error out, it gets apologetic and investigates why things went wrong. As it turns out, some additional parts need to be configured, which it tried to do and failed once again. Agentic Tool use is now a specific benchmark, but when asking Gemma 4 12B to set up a multi-step task like this, it simply couldn’t manage with the conversational tone of my prompts.

It’s at this point that I asked HammerClaw to give itself a grade, to which it offered itself a B- on accuracy for hallucinating and making up news stories. Now, I don’t want to say that all OpenClaw AI agents would do this, as we are running a relatively light local model. I would expect that larger local models would be able to identify what they need to do much more efficiently due to having a significantly higher model parameter count. With that in mind, I wondered if another AI model could help HammerClaw along a little bit with this task. And so, I went onto my OpenRouter account and opened a chat with the big, beefy 2.8 Trillion Parameter Kimi K3.

Kimi K3 saves the day

Comparing a cloud-based frontier model like Kimi K3 and Gemma 4 12B just isn’t fair. One takes up terabytes of RAM to run, while Gemma fits inside a Mini PC that lives on my desk.

Instead of using Kimi K3 to run the AI itself, I tasked it with actually helping me set up the job that HammerClaw seemingly didn’t have the skills to pull off. This hybrid workflow – of using a local LLM in tandem with a more powerful one in the cloud- is a common setup for many AI enthusiasts. After about a dollar in tokens, it sent me a list of OpenClaw commands, skills, and instructions, after reading the current documentation for OpenClaw CLI to ensure it got everything right. It was flawless. I followed the instructions Kimi K3 gave me and entered them into an Ubuntu Terminal, where it successfully created a new skill for HammerClaw named ‘News-Intel,’ instructed me on how to enable the web-search functions, and set up the Cron jobs for the automated sends. Bearing in mind that I had never used OpenClaw before, it was all surprisingly smooth, even if our local model couldn’t do all of these tasks itself.

With everything set up, the buck is then passed back to our much humbler Gemma 4 12B model-powered HammerClaw to execute the rest – after all, this is supposed to be a test for how Local AI models function. I executed the command for a manual run while checking the operational logs, where it was successfully calling all of the tools, and thought, against all odds, HammerClaw might actually be able to do it. A few minutes later, I received a Telegram message. HammerClaw had managed to locally execute the task and send me a digest of ten news stories.

Of all of the things to automate, this is likely one of the simplest examples of how someone can use OpenClaw. While simple, it’ll also save me a bit of time every day staring at an RSS feed and looking for stories myself. But, it’s quite a distance away from the “speak, and it’ll do whatever you want it to!” promise that drew so many into a frenzy earlier this year.

Is it worth it for the average person?

If all you heard about OpenClaw is that it’s a magical AI agent that can do anything, as I’ve learned, it’s not quite the truth. Unless you’re well-versed in several elements, like understanding model choices and getting your head around what models perform well for tool-calling. Running a humble local setup might be fun, or interesting to tinker around with, but the true power for Local AI developers and tinkerers lies with the obvious: More power to run bigger, complex models, and more agents running tasks consecutively. In fact, we’ve taken a look at how these workflows can be used in practice with a mixture of Local and Cloud AI models for Tom’s Hardware Premium.

The issue here is that for our local model, this simple task – of running and sending us 10 news stories- could not set itself up, and that’s using hardware that already costs north of $1,500. As our resident local AI expert and GPU guru Jeff Kampman has recently tested, for those serious about AI who want real power without relying on the cloud entirely, you might want to save up your pennies to run stronger, faster local models, either using dedicated GPU-accelerated setups, or dedicated boxes such as the DGX Spark, or a cluster of Dell Pro Max GB10s, both of which will cost you north of $5,000. You’d hope that the models capable of running on that hardware wouldn’t fumble the setup of a relatively simple Cron job.

The real question is, will having a local AI inference box meaningfully change how you work, or the work you do, to pay for itself? For many, that’s the lingering question that many are asking themselves. For now, HammerClaw is sending me more articles every few hours, a task that can be performed by simple scripting. But the manual “sift” is being handled by an LLM. As neat as it is, I wouldn’t pay $1,500 for the privilege. Luckily, it’s not merely a box made for AI inference; there’s a whole computer attached.