Graphics Cards - Page 67

Real-time ray-tracing is not only hardware intensive to the point where it needs dedicated hardware to run complex calculations for advanced lighting effects but also heavy on VRAM usage. This has led to situations where 8GB video cards like the GeForce RTX 4060 Ti run into bottlenecks in heavy RT workloads.

According to a recently published patent filed by Microsoft, it is all about using level-of-detail (LOD) to improve ray-tracing performance and quality as needed. Different LOD settings and rendering exist in games, where objects in the distance are rendered with less geometry or detail to boost performance. This also applies to things like texture quality and foliage, which look their best the closer your in-game character is to a particular spot.

Ray-tracing involves a new "residency map for a sub-tree corresponding to a bounding volume hierarchy of objects," where ray-tracing detail is calculated based on an LOD system. Distant ray tracing will be of lower quality, but greater detail will be available up close, freeing up resources like VRAM.

Intel's next-generation Battlemage "Xe2-HPG" GPUs have been spotted, running with 20-core and 24-core GPU variants inside the SiSoft Sandra database. Check them out:

Intel is expected to roll out its next-gen Battlemage GPU later this year, filling out the Xe2-HPG and Xe2-LPG variants: Xe2-HPG will be discrete graphic cards, while Xe2-LPG will be integrated GPUs inside upcoming processors like Intel's next-gen Lunar Lake CPUs.

The new Battlemage Xe2-HPG GPUs were spotted by "miktdt" with the 20-core Xe2-HPG GPU featuring 160 Vector Engines and 2560 ALUs, while the 24-core Xe2-HPG GPU packed 192 Vector Engines and 3072 ALUs. Both of the Xe2-HPG GPUs featured 8MB of L2 cache, which is half of the cache Intel uses on Alchemist Xe-LPG GPUs with 32 cores. Both of the GPUs also packed 12GB of VRAM on what seems to be a 192-bit memory interface.

NVIDIA's set to reduce its production of Lovelace graphics cards from now on in preparation for the launch of Blackwell GPUs later this year.

That's the latest from the rumor mill, specifically Korean tech site Quasarzone as noticed by Wccftech.

Industry experts in South Korea are cited as sources who are saying that NVIDIA has told its manufacturing partners that it'll be dropping shipments in terms of supply for existing RTX 4000 graphics cards.

NVIDIA's RTX 3050 received a new spin on the graphics card not so long ago, a considerably cut-down version with 6GB of VRAM, and apparently it's selling well in China.

Indeed, this GPU - intended as a fresh option in place of the old GTX GPUs which have been chucked on the scrapheap - is under 'huge' demand in China, according to a report highlighted by Wccftech.

So much so, that apparently graphics card manufacturers over in the country are increasing the asking price of the RTX 3050 6GB, or that's the word on the grapevine, specifically from the Chinese Board Channel forums.

NVIDIA announced its next-generation Blackwell GPU architecture and the world's fastest AI GPU, the Blackwell B200 AI GPU, last week at GTC 2024 (GPU Technology Conference). CEO Jensen Huang said it'll be around 10 years before something like DLSS 10 renders every frame using AI.

NVIDIA CEO Jensen Huang predicted the future of AI in gaming, saying that in under 10 years we will see AI generating every single pixel in every game, all in real-time. Jensen said: "I think with almost, almost everything in technology, the S curve is not longer than a decade once it becomes true, once it becomes practical and better. And, of course, ChatGPT is not only practical; in most cases, it's better. I think it's less than ten years away".

Jensen continued: "In ten year's time, you're at the other end of that S curve. In five years from now, you're probably right in the middle where everything is changing in real-time, and everybody's going, 'Oh, look at that, this is happening.' And so you just got to decide, are we two years into it, into that ten years? Probably. We're probably already two years into it. And so I would say that within the next five to ten years, somewhere in between, it's largely the case".

Samsung wasn't the only one showing off next-gen GDDR7 memory chips at NVIDIA's GPU Technology Conference (GTC) 2024 event, with SK hynix also having some GDDR7 on the show floor. Check it out:

Our friends over at HardwareLuxx were on the show floor at NVIDIA GTC 2024, spotting SK hynix's new GDDR7 memory modules that offer up to 40Gbps per pin for up to 160GB/sec memory bandwidth per module. We've heard and reported that GDDR7 memory will start its life at 28Gbps on NVIDIA's next-gen GeForce RTX 50 series GPUs, with refreshes along the way packing higher-speed GDDR7 modules.

Samsung has claimed it is working on 37Gbps GDDR7 memory while also showing off next-gen 32Gbps and 28Gbps GDDR7 memory modules at NVIDIA GTC 2024. We should expect NVIDIA to roll out GeForce RTX 50 series graphics cards with 28Gbps GDDR7 memory, but I would lean more towards 32Gbps GDDR7 to give them a huge performance leap over the GDDR6X-based flagship GeForce RTX 4090.

NVIDIA may have new spins on its RTX 4070, 4060 Ti and 4060 graphics cards inbound, but if so, let's make it clear from the outset that these won't be anything interesting.

A tweet has surfaced from regular leaker MEGAsizeGPU claiming some new versions are planned by NVIDIA, but before you begin imagining some kind of 4060 Super (or indeed 4060 Ti Super), these graphics cards will keep the exact same specifications as the existing versions (if they're anything at all - add salt as ever).

So, what's the point of that? Well, they will have a different chip, with the changes being as follows:

With Intel's Arc graphics now covering discrete GPUs like the Arc A750 and recent Arc A580 and integrated graphics in the new Intel Core Ultra series of mobile processors, the diligent team over at Team Blue continues to improve driver support for new games and older titles.

Over the past year, we've seen significant overhauls to DirectX 12 and DirectX 11 performance, with per-game optimizations bringing some game-changing improvements. The latest Intel Game On Graphics Driver release for Arc Graphics is no different, with new game support for Dragon's Dogma 2, Horizon Forbidden West, and Diablo IV's upcoming ray-tracing update; a bunch of older titles are also getting "up to 174% average FPS uplift."

Those massive performance increases are limited to a handful of titles, like Detroit: Become Human (Vulkan) and Granblue Fantasy: Relink (DX11). Still, a 36% performance bump to God of War running on a desktop Arc GPU and a 17% performance bump running on an Intel Core Ultra with built-in Arc are great to see.

AMD has released AMD Software: Adrenalin Edition 24.3.1 for Radeon GPU owners, and the latest driver update from Team Red is a big one - not only in terms of new game support but also in the number of 'Fixed Issues.'

First, the Adrenalin Edition 24.3.1 adds day-one support to the two biggest PC game releases of the week (and month), with Capcom's Dragon's Dogma 2 and Sony's Horizon Forbidden West Complete Edition getting some Radeon love. Outpost: Infinity Siege, the third game listed in the release notes, is more of an unknown. It's described as an "FPS, Base Building, Tower Defense" game, launching next week on Steam.

Onto the fixed issues, high-end Radeon RX 7900 Series GPU owners playing the smash-hit Helldivers 2 will be pleased to know that Adrenalin Edition 24.3.1 resolves and improves driver timeout issues and crashes when playing the game. Another win for all Radeon owners is AMD fixing micro stutter issues with AMD Fluid Motion Frames, improving the driver-based Frame Generation solution that's part of HYPR-RX.

NVIDIA only just announced and detailed its new Blackwell GPU architecture and new B200 AI GPU at GTC 2024 this week, but now we're hearing fresh rumors on the next-gen Vera Rubin GPU, which could be released in the first half of 2025.

The news is coming from sources close to Moore's Law is Dead, with the leaker saying one of his sources said that AMD is doing everything possible to get its next-gen Instinct MI400X out in the wild, which would arrive in Q1 2025 "at the earliest". NVIDIA doesn't want to stop the AI GPU domination, so don't think Blackwell is where it stops... as Vera Rubin dropping in 2025 would be absolutely insane.

Blackwell is said to be the GPU architecture for both the AI and data center GPUs from NVIDIA and its upcoming next-gen GeForce RTX 50 series graphics cards. Vera Rubin will succeed Blackwell, but will only be for the AI GPU market like Hopper and Volta are. Vera Rubin won't be for gamers, so we could expect NVIDIA to reveal its next-gen Vera Rubin GPU architecture at its GTC 2025 event.

AMD shipped its Radeon RX 5600 XT graphics card with 6GB of memory, but now YouTuber and modder Paulo Gnomes has more than doubled the memory capacity from 6GB to an incredible 16GB.

Not only that, but the modder also increased the stock 192-bit memory bus up to a higher 256-bit memory bus, which, when overclocked, the PowerColor Radeon RX 5600 XT Red Devil graphics card that was modded was up to 29% faster than its stock 192-bit memory bus and stock 6GB of GDDR6.

The modder had to modify the BIOS on the Radeon RX 5600 XT Red Devil graphics card, adjusting the right hexadecimal code to tell the GPU to unlock the additional memory controllers inside of the Navi 10 GPU die.

Intel's Arc Graphics platform, which covers discrete desktop GPUs like the Intel Arc A750 and mobile chips like the new Intel Core Ultra 155H with Intel Arc built-in, is designed for modern gamers and creators - leveraging the latest PC tech. For creators, Intel Arc offers hardware video encoding and decoding covering AVC, HEVC, and AV1.

This is all handled by Arc's Intel Xe Media Engine. In a new blog post (and video), Intel explains precisely what encoding and decoding video is about while presenting benchmarks for its Arc hardware - comparing performance to competing products. The post doubles as an excellent primer for learning about video compression and decompression while showcasing why a GPU-based hardware solution is essential in 2024.

In its head-to-head transcoding comparisons, we see the Intel Arc A750 versus the GeForce RTX 4060 and the Intel Core Ultra 155H versus the AMD Ryzen 7 PRO 7840U. And in a lot of workloads, Arc pulls ahead.

Horizon Forbidden West Complete Edition is out this week, bringing the highly anticipated PlayStation exclusive to the PC platform. Horizon Zero Dawn was one of the first major PS4 to PC ports, and this sequel, built to take advantage of the more powerful PS5 hardware, will present a significant visual upgrade to the original. A massive open-world action adventure, the Complete Edition includes the PS5-only Burning Shores expansion.

In addition to supporting unlocked frame rates, ultrawide displays, and advanced graphics settings to take advantage of modern GPUs, it will arrive on day one with NVIDIA DLSS 3, DLAA, and Reflex support. And with that, NVIDIA's latest GeForce Game Ready Driver 551.86 - WHQL with optimizations for Horizon Forbidden West Complete Edition. Download it here.

Usually, NVIDIA provides benchmarks for high-profile releases like this. Instead, it reiterates Sony's performance info, where GeForce RTX 4080 gamers can expect solid 4K 60 FPS performance using the game's 'Very High' quality setting without DLSS. OF course, DLSS will present a notable free performance bump for all GeForce RTX owners.

A new report from TrendForce reveals that "significant capital investments" have been made in the memory sector due to HBM's high ASP and profitability.

According to TrendForce Senior Vice President Avril Wu, by the end of 2024, the DRAM industry is expected to allocate around 14% of its total capacity to making HBM, with an estimated annual supply bit growth of around 260%. Also, HBM's revenue share is expected to increase to around 20.1% by the end of the year, up from the 8.4% that HBM had in 2023.

Wu continued, saying that in terms of production differences between HBM and DDR5, the die size of HBM is around 35-45% larger than DDR5 even on the same process and capacity (comparing 24Gb to 24Gb). Yield rates including TSV packaging (Through-Silicon Via) for HBM is around 20-30% lower than DDR5, with the production cycle taking 1.5 to 2 months longer than DDR5.

SK hynix has just announced that it has begun volume production of HBM3E memory, the newest AI memory product for ultra-high performance that it supplied to a "customer" (NVIDIA) in late March.

SK hynix is the first provider of next-gen HBM3E memory after enjoying great success with its HBM3 memory on various AI GPUs on the market. The company issued a press release to announce that it is the first in the industry to mass provide HBM3E memory and that supply to a customer after 7 months of development offers best-in-class performance to lead level of AI capabilities to the world's best.

The company recently invested another $1 billion into future HBM memory while filling NVIDIA's new Blackwell B200 AI GPU with ultra-fast HBM3E memory sees SK hynix at the top of its game. Sungsoo Ryu, Head of HBM Business at SK hynix said that mass production of HBM3E has completed the company's lineup of industry-leading AI memory products, adding: "With the success story of the HBM business and the strong partnership with customers that it has built for years, SK hynix will cement its position as the total AI memory provider".

NVIDIA CEO Jensen Huang told CNBC this week that the new Blackwell B200 AI GPUs will cost $30,000 to $40,000 each.

In the CNBC interview, Jensen also said that NVIDIA spent a whopping $10 billion on R&D efforts into Blackwell, which is a monumental amount of money for what will be the best AI GPU technology the world has ever seen. At a price of $30,000 to $40,000 each, that's not bad considering that it was expected that NVIDIA would charge $50,000 to $60,000 each.

We don't know if Jensen meant the Blackwell B100 AI GPU or the flagship B200 AI GPU... but considering B200 was the one that NVIDIA showed off and detailed some of its specifications for at GTC 2024 this week, you'd expect Jensen meant B200. NVIDIA will be making some mighty profits if B200 costs $6000 to make, with a selling price of 5x that... profits for later this year and into 2025 are going to make the recent record gains seem like a tiny wave compared to the AI GPU financials with a B200-powered tsunami.

The R&D budget of NVIDIA's next-generation Blackwell GPU architecture was "something like" $10 billion, according to NVIDIA CEO Jensen Huang.

Jensen was on CNBC this week, where he said NVIDIA's next-gen Blackwell AI GPU "will cost $30-$40K USD. The very first one, the R&D budget of this generation is probably something like $10 billion USD". Rumor has it that NVIDIA's new Blackwell B200 AI GPU will cost over $6000 to make, so they can expect some truly healthy profits from Blackwell GPU sales.

Blackwell costs $30,000 to $40,000 per AI GPU, which means at a cost of, let's say, $8000, NVIDIA is in the green, pun intended. If we thought that record-breaking share prices pushing the company over the $1 trillion market cap was good, Blackwell will hit another few home runs for NVIDIA.

Samsung has been showing off its next-generation GDDR7 memory at NVIDIA's GTC 2024 event, showing off 28Gbps and 32Gbps at the show.

The new picture of Samsung's next-gen GDDR7 memory comes from HardwareLuxx, which spotted the GDDR7 memory at GTC 2024. NVIDIA unveiled its next-gen Blackwell GPUs at GTC 2024, but they were using ultra-fast HBM3E memory, while GDDR7 will be baked into the next-gen Blackwell-based GeForce RTX 50 series that will debut later this year.

Samsung says that its new GDDR7 memory has 20% better power efficiency at just 1.1V -- compared to the industry standard 1.2V -- much lower than the power-hungry 1.35V that GDDR6X memory modules consume.

AMD's RX 7900 XTX graphics card has dropped back down to $899 for one model, the lowest price seen in a while (although not the cheapest price ever, by any means).

VideoCardz noticed that the MSI Gaming Radeon RX 7900 XTX currently has a discount of $150 at Newegg, with the flagship at $899 as mentioned. Going by the reviews, this is a well-made board that runs cool and quiet, which is always a bonus.

So, is it worth forking out for the RX 7900 XTX at this price compared to what NVIDIA has to offer in this price bracket?

Following the unveiling of NVIDIA's B100 (Blackwell GPU for heavyweight lifting and AI usage) some fresh chatter on the rumor mill is causing concern regarding the incoming RTX 5090.

Kopite7kimi sparked off these worries on X (formerly Twitter), as well as another respected hardware leaker on the platform, Kepler, as you can see in the above and below tweets.

With performance gains of 50% or 60%, or even 70% being floated previously for the RTX 5090 (using GB202), those estimates have been called into question. As Kopite7kimi makes clear, GB202 will use the same process as GB100, based not on 4nm, but on 4NP, a custom performance-oriented spin on 4N (an upgraded 5nm process).