A few weeks ago, we wrote how Eliyan’s NuLink PHY could do away with silicon interposers and integrate everything into an single, elegant package. How, essentially, the socket could become the motherboard.
At the recent 30th annual North America Technology Symposium, the Taiwan Semiconductor Manufacturing Company (TSMC) revealed plans to construct a version of its chip-on-wafer-on-substrate (CoWoS) packaging technology that could lead to system-in-packages (SiPs) over twice the size of the current largest ones.
“With System-on-Wafer, TSMC is providing a revolutionary new option to enable a large array of dies on a 300mm wafer, offering more compute power while occupying far less data center space and boosting performance per watt by orders of magnitude,” the company said.
An enormous amount of power
TSMC’s first SoW offering, a logic-only wafer based on Integrated Fan-Out (InFO) technology, is already in production.
A chip-on-wafer version using CoWoS technology is expected to arrive in 2027, and will enable the “integration of SoIC, HBM and other components to create a powerful wafer-level system with computing power comparable to a data center server rack, or even an entire server.“
Reporting on the move, Tom’s Hardware expands on this saying, “One of the designs that TSMC envisions relies on four stacked SoICs mated with 12 HBM4 memory stacks and additional I/O dies. Such a giant will certainly draw an enormous amount of power – we are talking about thousands of watts here and will need a very sophisticated cooling technology. TSMC also expects such solutions to use a 120x120mm substrate.”
TSMC’s ambitious pursuit to create gigantic chips, however, is dwarfed by Cerebras Systems’ newest Wafer Scale Engine 3 (WSE-3), termed the “fastest AI chip in the world”. The WSE-3 boasts four trillion transistors and is twice as powerful as its predecessor, the WSE-2, while maintaining the same energy consumption and price. This new chip created on a 5nm TSMC process, provides a staggering peak AI performance of 125 petaflops – which is equivalent to 62 Nvidia H100 GPUs.
Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!
South Korean memory giant SK Hynix, which recently announced plans for the construction of the world’s largest chip factory, has now announced a major partnership with top Taiwanese semiconductor foundry, TSMC.
The two firms aim to cement their positions in the fast-growing AI market by developing and producing the next-generation of High Bandwidth Memory, known as HBM4.
The production, scheduled for 2026, will make use of TSMC’s state-of-the-art packaging technology. The project’s initial focus will be on the improvement of the performance of the base die, the element of HBM that connects directly to the GPU. SK Hynix will reportedly adopt TSMC’s advanced logic process for HBM4’s base die, allowing for extra functionality to be packed into a minimal space. The two firms intend to optimize the integration SK Hynix’s HBM and TSMC’s CoWoS (Chip on Wafer on Substrate) technology.
A strategic move for TSMC
“We expect a strong partnership with TSMC to help accelerate our efforts for open collaboration with our customers and develop the industry’s best-performing HBM4,” said Justin Kim, President and the Head of AI Infra, at SK Hynix. “With this cooperation in place, we will strengthen our market leadership as the total AI memory provider further by beefing up competitiveness in the space of the custom memory platform.”
Dr. Kevin Zhang, Senior Vice President of TSMC’s Business Development and Overseas Operations Office, and Deputy Co-Chief Operating Officer, agreed, stating, “TSMC and SK Hynix have already established a strong partnership over the years. We’ve worked together in integrating the most advanced logic and state-of-the art HBM in providing the world’s leading AI solutions. Looking ahead to the next-generation HBM4, we’re confident that we will continue to work closely in delivering the best-integrated solutions to unlock new AI innovations for our common customers.”
This collaboration is a strategic move for TSMC as much as it is for SK Hynix – if not more so – and demonstrates the firm’s potential in a role beyond foundry service provider. It’s hard to guess at the future, but should TSMC wish to continue its growth trajectory, it may need to consider expanding its strategic horizons even further, a move which could see it competing with some of its partners such as AMD or even rivals like Intel.
Competition is fierce in the semiconductor industry and moving up the value chain – leveraging its advanced technology to drive higher margins – is a risky move, but one that could potentially be hugely rewarding for the Taiwanese chip giant.
Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!
Despite some stiff competition from the likes of Lenovo, Asus, MSI, and more, Valve’s Steam Deck has made its position as one of the best PC gaming handhelds in the market known time and again. And with the announcement that Nvidia’s GeForce NOW will introduce wider support for the Steam Deck, its dominance is even more assured.
There are two particularly excellent updates geared towards Steam Deck users. The first is a new beta installation method for GeForce NOW on the Steam Deck, which will automatically install Google Chrome to the device and add all necessary settings so users can immediately start playing after booting up for the first time.
And the second is a GeForce NOW update that lets users navigate on the browser with a gamepad, including on the Steam Deck. This makes finding and playing games on the handheld without worrying about system specs — including non-Steam games and titles with Nvidia DLSS support and ray-tracing.
Nvidia and Valve had been in talks since 2023 to add more support for the Steam Deck, so these updates were certainly a long time coming.
Steam Deck is charging full steam ahead in the gaming space
These latest updates from Nvidia help to smooth over some of the Steam Deck’s issues, mainly that its lack of power compared to other PC handhelds is becoming more apparent over time. The Steam Deck OLED fixes some of these problems, but certain titles cannot natively run without mods.
But being able to stream titles through GeForce NOW this easily is the best way to combat this issue (at least until the Steam Deck 2 finally launches, of course) since the hardware is completely irrelevant to how well the best PC games run. As long as you have a solid connection, framerate and general performance are as stable as it can be, even for titles that the system can’t otherwise play.
And with the added controller support coupled with the Steam Deck’s largest advantage over the competition — its excellent OS fully optimized for the handheld — it truly feels like it’s gotten meaningful enhancements that keep it competitive with its more powerful rivals.
Get the hottest deals available in your inbox plus news, reviews, opinion, analysis and more from the TechRadar team.
Nvidia continues to invest in AI initiatives and the most recent one, ChatRTX, is no exception thanks to its most recent update.
ChatRTX is, according to the tech giant, a “demo app that lets you personalize a GPT large language model (LLM) connected to your own content.” This content comprises your PC’s local documents, files, folders, etc., and essentially builds a custom AI chatbox from that information.
Because it doesn’t require an internet connection, it gives users speedy access to query answers that might be buried under all those computer files. With the latest update, it has access to even more data and LLMs including Google Gemma and ChatGLM3, an open, bilingual (English and Chinese) LLM. It also can locally search for photos, and has Whisper support, allowing users to converse with ChatRTX through an AI-automated speech recognition program.
Nvidia uses TensorRT-LLM software and RTX graphics cards to power ChatRTX’s AI. And because it’s local, it’s far more secure than online AI chatbots. You can download ChatRTX here to try it out for free.
Can AI escape its ethical dilemma?
The concept of an AI chatbot using local data off your PC, instead of training on (read: stealing) other people’s online works, is rather intriguing. It seems to solve the ethical dilemma of using copyrighted works without permission and hoarding it. It also seems to solve another long-term problem that’s plagued many a PC user — actually finding long-buried files in your file explorer, or at least the information trapped within it.
However, there’s the obvious question of how the extremely limited data pool could negatively impact the chatbot. Unless the user is particularly skilled at training AI, it could end up becoming a serious issue in the future. Of course, only using it to locate information on your PC is perfectly fine and most likely the proper use.
But the point of an AI chatbot is to have unique and meaningful conversations. Maybe there was a time in which we could have done that without the rampant theft, but corporations have powered their AI with stolen words from other sites and now it’s irrevocably tied.
Get the hottest deals available in your inbox plus news, reviews, opinion, analysis and more from the TechRadar team.
Given that it’s highly unethical that data theft is the vital part of the process that allows you to make chats well-rounded enough not to get trapped in feedback loops, it’s possible that Nvidia could be the middle ground for generative AI. If fully developed, it could prove that we don’t need the ethical transgression to power and shape them, so here’s to hoping Nvidia can get it right.
If you’re a frequent gamer, by now you’ll have heard of ray tracing and NVIDIA’s DLSS (Deep Learning Super Sampling), two powerful tools that combine to boost performance in games and provide stunning visual fidelity by replicating realistic lighting and reflections. NVIDIA pioneered these technologies in gaming with the introduction of its RTX 20-Series graphics cards, and in the two hardware generations since then, the techniques and AI hardware behind them have improved considerably. With DLSS 3.5 and a new Ray Reconstruction system working hand-in-hand, ray-traced games can look all the more lifelike and run significantly smoother at the same time. So how do these technologies work?
DLSS has been an evolving technology. At its start, its focus was on rendering games at lower, easier-to-process resolution and then increasing the output resolution by filling in the gaps between pixels , giving gamers the benefit of sharper visuals with increased frame rates in lower resolution. NVIDIA managed this by training its AI model on high-quality game visuals so that it could understand what they should look like and know how to fill in the gaps when stretching a game’s lower resolution frames to higher resolutions. This process could also flipped somewhat with DLAA (Deep Learning Anti-aliasing), which can render at a display’s native resolution but use the same AI logic to figure out what an even higher-resolution frame would look like and then down-sample that to output an effectively anti-aliased image.
With DLSS 3, it introduced Frame Generation, which fills in pixel data between frames. While DLSS has always relied on the special Tensor Cores inside of RTX GPUs, which handle AI operations, DLSS 3 Frame Generation leverages even more hardware in RTX 40-Series GPUs, like Optical Flow Accelerators, to understand how objects in games are moving and intelligently blend between frames for extra smooth visuals. Combining the super sampling and Frame Generation tech in DLSS, can dramatically increase frame rates in games without a hit to visual quality.
Alongside these developments in DLSS, NVIDIA has continued to push forward its RTX ray tracing technology. Ray tracing remains a computationally intensive process, especially to get the most realistic results. It requires simulating a staggering amount of light rays and all reflections they make between a light source and the viewer, making it unreasonable to perform this for every pixel in every frame of a fast-running game. But taking a reasonable number of light ray samples produced a noisy image and entailed specially trained denoisers to get a usable image. Even then, the results could lack detail or have unusual ghosting artifacts, and the more denoisers necessary to handle numerous ray-traced effects could further bog down performance.
NVIDIA’s DLSS 3.5 introduced Ray Reconstruction to solve this. With Ray Reconstruction, all of the denoisers (and their corresponding computational needs) get replaced. Much like DLSS and Frame Generation use AI to figure out how to intelligently fill in pixels, Ray Reconstruction uses AI to fill in the gaps between simulated light rays. Ray Reconstruction has a deeper understanding of the games it runs on and their ray-traced effects, allowing it to know when to use different techniques to fill in missing data, helping it provide sharp visuals without artifacts.
For gamers, higher resolutions and higher graphical settings used to mean making big sacrifices to framerate. But the technologies wrapped up in DLSS 3.5, like Frame Generation and Ray Reconstruction, flip the script.
Now, gamers on any RTX GPU can peer into lifelike worlds with ray-traced lighting, shadows, reflections while enjoying higher resolutions. And those gaming on RTX 40-series graphics processors can tap into Frame Generation for yet another boost to frame rate. Taking advantage of these technologies is as simple as running an NVIDIA RTX graphics processor in your desktop or laptop (or even tapping into one through NVIDIA’s GeForce NOW Ultimate or Priority service), and making sure to toggle these features on inside of your games. You can find all the games and apps that support NVIDIA’s DLSS technologies here.
The way NVIDIA has leveraged AI to upgrade the gaming experience is just one way the company is putting AI to work. With NVIDIA’s RTX hardware, you can do a lot more than gaming, and NVIDIA’s AI Decoded blog series highlights many of the different ways you can take advantage of your RTX hardware running all sorts of AI-powered tools.
In a darkened room at Nvidia’s ‘Future of Gaming’ showcase at a fancy London hotel earlier this week, I was ushered before a bank of computer screens showing live gameplay from a very exciting game: Valve’s Half-Life 2.
Alright, HL2 isn’t a new game – in fact, it turns 20 this year! – but it’s still a timeless classic that has topped many a list of the best PC games. And while Valve will probably never release Half-Life 3, the mega-hit second game still has plenty of life left in it, thanks to Nvidia’s incredibly clever RTX Remix tool.
Announced way back in September 2022 at Nvidia’s GTC showcase, RTX Remix is an AI-powered tool for remastering old 3D games with updated graphics and fancy modern features like ray tracing. Remix entered open beta earlier this year, and is primarily targeted at modders looking to visually upgrade their favorite games; it’s built on Nvidia’s AI-focused ‘Omniverse’ platform, offering a comprehensive box of tricks for making the remaster process faster and easier than ever before.
Portal with RTX looks fine, but Portal already looked good – Half-Life 2 is a much better showcase for RTX Remix. (Image credit: Valve, Nvidia)
Nvidia worked with official developers to create a showcase for Remix in Portal with RTX, a shiny new remaster of the legendary 2007 puzzle game, but Half-Life 2 RTX is a passion project being produced by modders with Nvidia providing some background support. And can I be honest? It looks f*cking awesome.
Giving new life to Half-Life
The live demo I was shown by Nvidia showcased one of the game’s most iconic locations: the run-down, zombie-infested town of Ravenholm, where protagonist Gordon Freeman ends up after a less-than-fortuitous series of events. Two displays were set up side-by-side to show both the original game and the Remixed work in progress – and the difference was phenomenal. For all its accolades, classic Half-Life 2 does very much look like a game made two decades ago, while the work-in-progress remaster looks great.
RTX Remix allows for modding in real-time – in other words, you can have the game running while also having the dev environment open in a different tab, and changes you make are reflected (after a short delay) in the live game. An Nvidia staffer gave me a brief walkthrough on using the tool, and it’s a remarkably streamlined process. You can generate an asset library by simply walking around in-game and letting Remix capture your surroundings, and use generative AI tools to rapidly produce improved textures and 3D models.
While the ray-traced lighting and reflections are the most immediately noticeable difference, the improved texture detail and remodelled assets are good too. (Image credit: Future / Valve)
One scene displayed during the demo saw me walking through a dimly lit courtyard beset with corpses – including a dismembered pair of legs hanging from a tree. Despite the grisly setting, I couldn’t help but be impressed: the legs swing from a rope and cast realistic dynamic shadows in the Remixed version, while the original features no shadows at all. It also highlighted the improved ground textures, and Nvidia pointed out in a different scene that Remix’s generative AI capabilities can extrapolate the environment to add extra details – in this case, climbing vines and weeds partially covering a ruined wall.
Get the hottest deals available in your inbox plus news, reviews, opinion, analysis and more from the TechRadar team.
CONTENT WARNING: You can see the clip below, but please be aware that it contains some 2004 gore as described above, now with added ray tracing! I do wish Nvidia had given me a slightly less gruesome example.
Another standout example (albeit a rather mundane and less gory one) was a simple 3D object asset of a broken engine. I lined up both versions of the game to view the item’s original and Remixed models, and the new model is massively more detailed. And I do mean massively: not only does the ray tracing work wonders to more realistically illuminate the object, but it actually looks like a 3D asset from a 2024 game, not a 2004 one.
The same 3D game asset looks incredibly different in the Remixed version of Half-Life 2. (Image credit: Future / Valve)
A change of heart – and art
I’ll admit, when RTX Remix was first announced, I was pretty cold about it. I liked the look of the (still in development) Elder Scrolls III: Morrowind remaster, but I had fears that Remix would become a blunt implement for major publishers to churn out low-effort remasters that threw the important art direction of classic games in the trash. I wasn’t the only one, either – my colleague Allisa James had some major reservations about Remix too.
But I’m relieved that two years on, Nvidia is putting power in the hands of modders rather than shareholders. When I asked Nvidia GeForce ‘Evangelist’ Jacob Freeman (amusingly appropriate name, I know) about RTX Remix, he explained that the tool would be available for official game dev projects, but had been created with modders – not profits – in mind. It’s certainly a refreshing change of pace, and it helps that some of these Remix projects look pretty fantastic.
Morrowind is another good example of a game that definitely could benefit from an AI-powered remaster using RTX Remix. (Image credit: Nvidia, Bethesda)
I still maintain that some games might suffer from AI-infused remastering. Mirror’s Edge is a good example. For starters, it still looks pretty good today, but more importantly, it’s a game where color and lighting are used very carefully to inform the gameplay – meaning that a Remixed version might impact the delicate intersection between art direction and game mechanics. But for games like Morrowind and Half-Life 2, which have decidedly not aged well from a graphical standpoint, it’s perfect. I can’t wait to see what talented modders create with it.
If you’re interested in RTX Remix, you can download the beta version from Nvidia’s website. Bear in mind that right now, Remix only works with DirectX 9 and some DirectX 8 games – so in other words, you’re currently limited to a pool of games released between 2000 and 2006. Of course, the tool is still in development, and Nvidia has said that it would like Remix to work with a much broader spectrum of PC games, so in other words: watch this space!
Nvidia’s next generation of graphics cards, reportedly known as Nvidia Blackwell, has been the subject of plenty of rumors since at least 2023.
The most recent one seems to confirm that the RTX 5080 and RTX 5090 will be launching in Q4 2024, according to Twitter leaker Kopite7kimi and reported on by PC Gamer. While finally getting a hold of what could easily be the best graphics cards sounds like great news to some, finally getting a sign of the cards’ imminent release, they’re honestly coming out way too soon.
There have been reports that the Nvidia RTX 5080 would have better ray tracing capabilities and pricing than the 4080 and that the RTX 5090 would be up to 70% faster than the 4090, which sounds great on paper. But when you consider how already powerful the 4000 series of cards are, it seems a bit pointless. It’s wonderful that ray tracing, speed, and performance might be improved, but to what benefit?
We aren’t even taking full advantage of current gen cards, so what’s the value of buying a Nvidia 5000-series GPU later this year?
PC games haven’t even reached the point that they truly challenge the mid-range and high-end 3000-series of graphics cards (hell, an argument can be made that they can’t fully challenge the RTX 2070 Super in ways that truly matter), let alone the current gen. One could argue that the only reason why the current gen was so necessary was thanks to the crypto mining craze completely ruining the 3000-series market and making it nigh impossible to get your hands on one for a remotely affordable price.
And right now, the 4000-series is quite excellent performance-wise, as it’s able to handle ray tracing and other advanced effects and tools like no other. The RTX 4090 in particular is nearly perfect in what it can offer, and the lower cards are still absolute performance beasts. This isn’t even mentioning the Super series refreshes, which added even more oomph to your favorite graphics cards while keeping the prices either the same or lowering them.
There’s also the fact that some cards, like the RTX 4070, didn’t sell nearly as well as Nvidia wanted and in fact were rumored to be a “disaster.” While that doesn’t reflect the sales numbers for the rest of the graphics cards, it’s not a good look to see the base versions of your mid-range cards doing poorly. And while the RTX 4080 Super seems to be out of stock in many retailers, that could just as well be due to low stock in the first place.
With all these issues cropping up, releasing RTX 5080 and 5090 doesn’t seem to be such a smart move on Nvidia’s part. Though the Q4 2024 launch date is still just a rumor, it would be wise for the tech giant to release these cards in 2025 at the earliest to give its 4000-series some proper breathing room.
In a move to cut its dependency on Nvidia‘s high-cost AI chips, Naver, the South Korean equivalent of Google, has signed a 1 trillion won ($750 million) agreement with Samsung.
The deal will see the tech giant supply its more affordable Mach-1 chips to Naver, by the end of 2024.
The Mach-1 chip, currently under development, is an AI accelerator in the form of a SoC that combines Samsung’s proprietary processors and low-power DRAM chips to reduce the bottleneck between the GPU and HBM.
Just the start
The announcement of the Mach-1 was made during Samsung’s 55th regular shareholders’ meeting. Kye Hyun Kyung, CEO of Samsung Semiconductor, said the chip design had passed technological validation on FPGAs and that finalization of SoC was in progress.
The exact volume of Mach-1 chips to be supplied and prices are still under discussion, but The Korea Economic Daily reports that Samsung intends to price the Mach-1 AI chip at around $3,756 each. The order is expected to be for somewhere between 150,000 and 200,000 units.
Naver plans to use Samsung’s Mach-1 chips to power servers for its AI map service, Naver Place. According to The Korea Economic Daily, Naver will order further Mach-1 chips if the initial batch performs as well as hoped.
Samsung sees this deal with Naver as just the start. The tech giant is reportedly in supply talks with Microsoft and Meta Platforms who, like Naver, are actively seeking to reduce their reliance on Nvidia’s AI hardware.
Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!
With the Naver deal, Samsung is also looking to better compete with its South Korean rival SK Hynix, which is the dominant player in the advanced HBM segment. Samsung has been heavily investing in HBM recently and at the start of March announced the industry’s first 12-stack HBM3E 12H DRAM. This reportedly outperforms Micron’s 24GB 8H HBM3E in terms of capacity and bandwidth and is expected to begin shipping in Q2 this year.
In a recent interview with CNBC’s Jim Cramer, Nvidia CEO Jensen Huang shared details about the company’s upcoming Blackwell chip which cost $10 billion in research and development to create.
The new GPU, which is built on a custom 4NP TSMC process and packs a total of 208 billion transistors (104 billion per die), with 192GB of HMB3e memory and 8TB/s of memory bandwidth, involved the creation of new technology because what the company was trying to achieve “went beyond the limits of physics,” Huang said.
During the chat, Huang also revealed that the fist-sized Blackwell chip will sell for “between $30,000 and $40,000”. That’s similar in price to the H100 which analysts say cost between $25,000 and $40,000 per chip when demand was at its peak.
A big markup
According to estimates by investment services firm Raymond James (via @firstadopter), Nvidia B200s will cost Nvidia in excess of $6,000 to make, compared with the estimated $3320 production costs of the H100.
The actual final selling price of the GPU will vary depending on whether it’s bought directly from Nvidia or through a third party seller, but customers aren’t likely to be purchasing just the chips.
Nvidia has already unveiled three variations of its Blackwell AI accelerator with different memory configurations — B100, B200, and the GB200 which brings together two Nvidia B200 Tensor Core GPUs and a Grace CPU. Nvidia’s strategy, however, is geared towards selling million dollar AI supercomputers like the multi-node, liquid-cooled NVIDIA GB200 NVL72 rack-scale system, DGX B200 servers with eight Blackwell GPUs, or DGX B200 SuperPODs.
More from TechRadar Pro
Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!
Samsung is reportedly planning to launch its own AI accelerator chip, the ‘Mach-1’, in a bid to challenge Nvidia‘s dominance in the AI semiconductor market.
The new chip, which will likely target edge applications with low power consumption requirements, will go into production by the end of this year and make its debut in early 2025, according to the Seoul Economic Daily.
The announcement was made during the company’s 55th regular shareholders’ meeting. Kye Hyun Kyung, CEO of Samsung Semiconductor, said the chip design had passed technological validation on FPGAs and that finalization of SoC was in progress.
Entirely new type of AGI semiconductor
The Mach-1 accelerator is designed to tackle AI inference tasks and will reportedly overcome the bottleneck issues that arise in existing AI accelerators when transferring data between the GPU and memory. This often results in slower data transmission speeds and reduced power efficiency.
The Mach-1 is reportedly a ‘lightweight’ AI chip, utilizing low-power (LP) memory instead of the costly HBM typically used in AI semiconductors.
The move is widely seen as Samsung’s attempt to regain its position as the world’s largest semiconductor company, fighting back against Nvidia which completely dominates the AI chip market and has seen its stock soar in recent months, making it the third most valuable company in the world behind Microsoft and Apple.
While the South Korean tech behemoth currently has no plans to challenge Nvidia’s H100, B100, and B200 AI powerhouses, Seoul Economic Daily reports that Samsung has established an AGI computing lab in Silicon Valley to expedite the development of AI semiconductors. Kyung stated that the specialized lab is “working to create an entirely new type of semiconductor designed to meet the processing requirements of future AGI systems.’
Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!
