Graphics Cards - Page 67

Micron has announced that it is now sampling its next-generation GDDR7 graphics memory, with the industry's highest bit density.

Micron is using its new 1β (1-beta) DRAM technology and innovative architecture to build its new GDDR7 delivering 32Gbps of speed in a power-optimized design. We're looking at over a blistering 1.5TB/sec of memory bandwidth, an incredible 60% leap in bandwidth over current-gen GDDR6 memory, and four independent channels to optimize workloads.

This culminates in Micron's next-gen GDDR7 memory providing faster response times, smoother gameplay, and reduced processing times. On top of that, GDDR7 has a 50% power efficiency gain over GDDR6, enabling improved thermals and longer battery life inside of portable devices (laptops, etc) while the new sleep mode reduces standby power by up to an incredible 70%.

NVIDIA's next-gen Blackwell gaming GPUs are keenly anticipated, of course, and while Computex 2024 didn't give us any of the details on RTX 5000 that we wanted, a leaker has stepped into the breach with some fresh nuggets of rumors.

This comes from well-known leaker Kopite7kimi on X (formerly Twitter) who fired off a couple of posts, one of which called NVIDIA's Computex keynote boring, while imparting something more interesting - the mentioned purported details on GB202 and GB203 - in a further tweet (can we still call them tweets?).

Kopite7kimi tells us that we will only see GB202 and GB203, the GPUs for the RTX 5090 and 5080 respectively, this year - and all the other Blackwell graphics cards are coming next year. (At least that's what we expect the leaker means by outside the short term).

At Computex 2024, AMD's big Radeon announcement wasn't RDNA 4 or the next-gen desktop Radeon RX 8800 XT for PC gaming but a new dual-slot GPU for workstations and generative AI workflows. The AMD Radeon PRO W7900 Dual Slot (with blower fan) is a sleek and powerful unit with 48GB GDDR6 and up to 864 GB/s of bandwidth.

This is a smaller version of the triple-slot Radeon PRO W7900, making it a better fit for workstations with limited space and unlocking the full potential of RDNA 3 and the Navi 31 GPU. Other specs include 192 RDNA 3 AI Accelerators, 123 TFLOPS of FP16 peak half-precision performance, and a 295W Total Board Power rating. As a two-slot GPU, the length is 280mm, the same as the previous triple-slot edition.

Performance-wise, AMD is touting up to 52% better performance per dollar (this thing has an MSRP of $3,499 USD) than the NVIDIA RTX 6000 ADA GPU - partly because NVIDIA's AI monster costs around $6,800, almost double.

NVIDIA ACE technology first debuted at Computex last year with a demo showcasing an AI-powered digital human or NPC that you could interact with. The first demo took place in a futuristic bar. With your microphone, you could walk up to the AI NPC and interact with them using natural language. It is like talking to a character in an RPG, but without a dialogue wheel, you decide, choose, and ask what you want.

The demonstration was running on NVIDIA's powerful GPU hardware in the cloud at the time - the sort of data center rig created to run complex and sophisticated AI models. We've seen the technology evolve and improve in the past year. Still, this year's Computex 2024 NVIDIA ACE 'Covert Protocol' game tech demo (developed in collaboration with Inworld AI) will run locally on a GeForce RTX AI PC.

NVIDIA ACE technologies like Audio2Face and NVIDIA Riva automatic speech recognition can now run locally on devices with GeForce RTX hardware.

With AI capabilities coming to a range of new portable Copilot+ Windows 11 devices in the coming months, using hardware like the latest Arm-based Snapdragon X Elite processor - Microsoft has made it clear that to run its Copilot+ features locally, you'll need a processor with an NPU capable of 40 TOPS of AI performance, or Trillions of Operations Per Second.

Basically, a Windows PC with 40 TOPS of performance is AI ready. With AI features in Adobe, Blender, and other content creation tools, and even in PC games with tech like DLSS, a GeForce RTX graphics card offers a lot more AI power than 40 TOPS.

GeForce RTX owners have been using RTX AI-capable apps for a while now, which offer impressive tools and features. As part of its Computex 2024 presentation, NVIDIA has outlined the AI performance you can expect (in TOPS) for its entire GeForce RTX line-up across desktop and laptop GPUs. The current flagship GeForce RTX 4090 offers a whopping 'Up to 1321 TOPS' of AI performance - the equivalent of 33 Copilot+ AI PCs.

Want to put a new GeForce RTX 4080 in a new small form factor PC case? The SFF-ready label is here to make it much easier. NVIDIA has announced a new Small Form Factor (SFF) guideline for its GeForce graphics cards, enabling a much easier time with case selection for end consumers.

The new SFF-Ready GeForce card guideline was announced at Computex 2024 and will be the stamp of approval consumers will be looking for when purchasing a GeForce graphics card. More specifically, the SFF stamp of approval will be on the packaging alongside other GeForce features, and will be a guarantee that specific graphics cards are compatible with a small form factor PC case.

Having the SFF-ready stamp will make picking a new GPU and a case much easier as all consumers will have to do is look for NVIDIA's seal of approval. NVIDIA explained it has already partnered with a bunch of AIBs for the SFF-ready stamp, with companies such as ASUS, CoolerMaster, Fractal, Galaxy, GIGABYTE, INNO3D, MSI, NZXT, Palit, ZOTAC and many more on board.

NVIDIA pranked the world back in 2017 with its GeForce GTX 1080-shaped "G-Assist" USB flash drive, saying it had technology inside to power deep learning algorithms and bring AI to gaming... well... fast-forward to today, and that's actually the reality on GeForce RTX series graphics cards with ray tracing and AI-powered upscaling through DLSS.

NVIDIA now has multiple versions of DLSS (Deep Learning Super Sampling) powered by AI algorithms, plenty of beautiful ray-traced games supported by GeForce RTX series GPUs, and even NPCs that are powered by large language models for a totally new world of gaming in the near future.

Well, now the official NVIDIA GeForce account on X has posted "the future is never far away..." while teasing its GeForce GTX G-Assist flash drive. We have no idea what NVIDIA is hinting at here, but if it has anything that the G-Assist USB drive was meant to have, we could expect "Ghost Play" that plays the games while you're doing something else, or BossBoost, which is meant to help gamers through tough fights. There's also NV Nurture, a fictional technology that reminds gamers to recharge their energy by connecting to their fridges.

NVIDIA's next-generation flagship GeForce RTX 5090 graphics card is back into the rumor mill, with updated specs through new rumors.

We're now expecting to see 28GB of ultra-fast GDDR7 memory on an (interesting) 448-bit memory bus, with 1.5TB/sec of memory bandwidth on the GeForce RTX 5090. There were previous rumors of a monster 512-bit memory bus for the RTX 5090, with other rumors suggesting a 384-bit memory bus, but new rumors are right in between: a 448-bit memory bus.

This means that NVIDIA's flagship GeForce RTX 5090 graphics card might not be using all of the 16 available GDDR7 memory modules, and only 14 of the memory modules instead. The 28GB of GDDR7 memory on a 448-bit memory bus will be spitting out an impressive 1568GB/sec (1.56TB/sec) which is a huge 50% upgrade in memory bandwidth over the RTX 4090, which should make for some really impressive upgrades in 4K gaming.

MSI announced the new MSI GeForce RTX 4090 24G SUPRIM FUZION Graphics Card in an email confirming that it and other products have won Computex 2024 awards ahead of the show. It's a 4.5-slot beast because it includes an "innovative all-in-one hybrid liquid-cooled" solution with a pump, radiator, and short pipe waterway in a single block.

Ahead of its Computex reveal, German outlet Allround-PC (via Videocardz) has gotten some exclusive pictures of the new MSI GeForce RTX 4090 SUPRIM FUZION GPU and its unique design, which includes a compact cooling block with pump. You can see the copper heat sinks around the block and the pump in these new shots.

Integrating the radiator with the GPU is definitely going to turn a few heads, and we'll be doing our best to go hands-on with it at Computex and potentially review it in the near future, so stay tuned. According to the Computex Best Choice Awards, the design can deliver a "15% increase in performance compared to traditional air cooling technology."

Microsoft has announced that DirectSR is now available via the Agility SDK 1.714.0-preview release for game developers working with DirectX 12. Microsoft partnered with AMD, Intel, and NVIDIA to create this new API, allowing developers to integrate all the various upscaling technologies into their games.

This includes AMD FidelityFX Super Resolution (FSR), Intel XeSS, and NVIDIA DLSS Super Resolution. "These cutting-edge solutions elevate the gaming experience by boosting frame rates while maintaining exceptional visual quality," Microsoft's Joshua Tucker writes in the announcement. "Yet as the number of SR variants continues to grow, developers are in search of a common SR API that will seamlessly scale across the hardware ecosystem."

DirectSR breaks upscaling down to a common set of inputs and outputs so that a "single code path" can then plug straight into DLSS, FSR, and XeSS. With DirectSR, more PC game releases will support all three technologies at launch. It's a standalone solution that doesn't require "vendor-specific SDKs" or additional tools.

Since the release of the RTX 3090/Ti and RTX 4090 graphics cards have seemingly been getting bigger and bigger, with both of the aforementioned GPUs taking up three slots on a motherboard.

The idea behind making graphics cards larger is simply for cooling, as the GPUs become more sophisticated and provide more compute performance, more power is required. And when more power is required more cooling is required to thermal regulation. With the release of the RTX 3090 and RTX 4090 we were seemingly taking on a trajectory of bigger and bigger graphics cards with each new generation. But that might not be the case.

Renowned GPU leaker @Kopite7kimi recently took to Twitter and gave some information regarding NVIDIA's Blackwell gaming GPUs expected to launch sometime later this year. According to the leaker NVIDIA will be moving backwards in terms of size, with the company purportedly opting for a dual-slot, dual-fan solution over a three-slot solution. Notably, NVIDIA hasn't released a dual-slot GPu since the RTX 2080 Ti in 2019.

FromSoftware and Bandai Namco are prepping to launch the highly anticipated expansion for its smash-hit action RPG Elden Ring on June 21 called 'Shadow of the Erdtree.' As one of the most critically acclaimed and popular game releases this decade, fans are excited to explore the new Land of Shadow region. And thanks to Gunnir, PC gamers can do so in style with a pair of new Shadow of the Erdtree graphics cards.

However, there is a catch: These GPUs are limited to Intel Arc A750 and Intel Arc A770 models sporting custom Gunnir 'Photon' designs inspired by Elden Ring and the new DLC expansion. So yeah, it's not precisely high-end performance. Still, the minimal black and gold look is impressively stylish, with Elden Ring: Shadow of the Erdtree art found on the backplate.

Definite collector's items, these custom GPUs could be limited to the Asian market. There's no word on whether they will ship with in-game bonuses or Elden Ring-related goodies.

NVIDIA's next-generation flagship GeForce RTX 5090 graphics card will sport the "physically monolithic" GB202 "Blackwell" GPU die, according to the latest leaks.

We should expect the GB202 GPU to power the GeForce RTX 5090 with up to 192 SMs that will see a monster 24,567 CUDA cores (if the GB202 chip retains its 128 cores per SM design from the AD102 "Ada Lovelace" GPU inside of the RTX 4090). In a new reply on X, leaker "kopite7kimi" said that the GB202 GPU is a "physically monolithic" chip.

The reason this is important, is that NVIDIA used a chiplet design for its HPC and AI GPUs including the Blackwell B100 and B200 AI GPUs, but looks to use a physically monolithic GPU design for its gaming-focused GeForce RTX 50 series GPUs. We should also see NVIDIA rolling out next-generation GDDR7 memory on its RTX 50 series GPUs, with the RTX 5090 rumored to feature a hefty 32GB of ultra-fast GDDR7 memory.

NVIDIA is feeling the squeeze by the US government sanctions aimed at stifling China's progression in the AI race, which the chip manufacturer has responded to by slashing prices to beat the local Chinese competition.

The GPU manufacturer is attempting to make high-end products for the lucrative Chinese market that accounts for approximately 17% to its revenue for fiscal 2024., but it must now work within the strict US restrictions that ban the exportation of powerful chips that will likely be used to train AI models. NVIDIA plans to still release new GPUs in China, but these new models will have to be cutdown versions of new hardware that's limited to a specific performance point.

Reports indicate that NVIDIA found some of its Chinese customers weren't to thrilled about the new performance capped hardware, and would instead opt for domestically manufactured chips that are exponentially growing in presence. Reuters now reports NVIDIA has been forced to slash prices of some of its products to compete with new kits from Chinese technology giant Huawei.

The CEO of AMD, Lisa Su, has taken center stage at imec's ITF World 2024 conference to accept an award and outline AMD's goal of reaching 100x power efficiency by 2027.

Su accepted the prestigious Imec Innovation Award for innovation and industry leadership, joining prominent technology figures such as Bill Gates. Following the award's acceptance, Su began illustrating AMD's quest to achieve 100x performance improvement by 2027 and how AMD is going to overcome big problems such as the compute/energy required to train exponentially growing advanced AI models.

Su explained AI models typically doubled in size every two years, but now its growing at 20x per year, completely outpacing computing and memory advancements. To combat these problems AMD will be implementing a multi-facet strategy that will introduce new silicon architectures, AI-specific architectures, software, system and data tuning, and hardware co-designs.

According to a new report at PC Mag, which the outlet has confirmed with the US Trade Representative (USTR), the Biden administration is set to resume tariffs on PC hardware from the Trump era. These tariffs would place a 25% duty or import tax on goods assembled in China, including graphics cards, motherboards, and PC cases.

There's no word on when the tariffs will resume, nor has there been an official notice with confirmation that they're right around the corner. A spokesperson from the USTR told PC Mag, "We are maintaining tariffs on products currently subject to the action, including the two tariffs you've inquired about." Translation: Yes, the GPU tax is coming back.

What does this mean for those putting together PC rigs for gaming, recreation, or even work at home? Higher prices at a time when the last thing you'd want to see is a notable rise in the cost of buying a new GPU or motherboard. Hardware like graphics cards will fall under the new 25% duty, as China is the primary manufacturing and assembly base for PC hardware.

NVIDIA announced its financial results for the first quarter, and you can probably guess how it went - data center revenue from AI and GPU sales reached record highs, pushing the company's stock price to over $1,000 per share. With everyone all in on AI and NVIDIA leading the charge technology and hardware-wise, you might be wondering how the company's gaming side and GeForce business is going.

Unlike AMD's struggling Radeon and console gaming revenue, NVIDIA's Gaming sector is doing well - with $2.6 billion in revenue, up 18% from a year ago. Granted, it's nowhere near the $22.6 billion in data center revenue or the incredible 427% year-over-year growth seen in that sector - but it's good news for the traditionally quiet quarter for gaming revenue.

Gaming revenue was down 8% from the previous quarter, but the year-over-year growth is being hailed as a win for GeForce hardware. However, as AI continues to rule the day, NVIDIA's financial results cite 'AI gaming technologies' and 'AI performance optimizations' as critical milestones for NVIDIA's gaming business during the last quarter.

AMD Radeon Anti-Lag 2 has been formally announced and is now available in 'technical preview' form in Counter-Strike 2. Like NVIDIA's Reflex latency reduction technology, the new Radeon Anti-Lag 2 requires game integration - unlike Anti-Lag or last year's short-lived Anti-Lag+.

The choice of Counter-Strike 2 to debut Anti-Lag 2 is an excellent move from AMD because of the whole Anti-Lag+ controversy. Here's a quick rundown: Anti-Lag+ was launched as an in-driver solution for latency reduction in September 2023. Fast forward to October 2023, and Counter-Strike 2 players with Anti-Lag+ enabled were banned for cheating. It turns out Anti-Lag+ worked by accessing and even modifying game code and data.

From there, AMD disabled Anti-Lag+ and then removed it entirely. And now, several months later, we've got the new and very different Anti-Lag 2 that requires game integration.

NVIDIA will be rolling out next-gen GDDR7 memory on its next-generation GeForce RTX 50 series GPUs, with the GeForce RTX 5090 Founders Edition graphics card reportedly featuring 16 GDDR7 memory modules in a denser design.

It seems NVIDIA is changing up the memory layout for its new GeForce RTX 5090 over the current-gen RTX 5090, with the GB202 "Blackwell" GPU inside, next-gen GDDR7 memory on a huge 512-bit memory bus. If that's true, it would explain why NVIDIA would need to use additional GDDR7 memory modules around the GPU.

NVIDIA's new GeForce RTX 5090 will use GDDR7 memory arranged in a denser configuration, with the layout shifting from 3-4-1-4 to 4-5-2-5 (clockwise), which allows the board to support up to 16 memory modules.

NVIDIA's next-generation GeForce RTX 5090 Founders Edition graphics card is being teased, with a sophisticated PCB design.

In a new post on X, leaker "kopite7kimi" said that the RTX 50 FE graphics cards have 3 separate PCBs: the motherboard, I/O rigid board, and a separate PCIe slot component, that the leaker adds "perhaps it should not be considered as the third PCB".

The new GeForce RTX 50 series GPUs will have a bunch of new tech: a next-generation Blackwell GPU architecture, and next-generation GDDR7 memory (and some more surprises, I'm sure). We should expect NVIDIA's next-gen GeForce RTX 5090 Founders Edition graphics card to feature up to 16 x GDD7 memory modules surrounding the GPU, something we've learned in new leaks, too.