GPU and CPU news and discussions

[weatheriscool]

Intel Confirms Raptor Lake Refresh
The company will soon be launch both 14th generation and "1st" generation CPUs for desktop and mobile.
By Josh Norem June 23, 2023

https://www.extremetech.com/computing/i ... ke-refresh

Intel's operation in China has confirmed the company will be launching a Raptor Lake refresh later this year, which is the first time it's put its stamp of approval on all the recent rumors. The company will still be unveiling Meteor Lake as well, but as expected it will be mobile-only. However, there will also be upgraded Raptor Lake mobile parts as well, which will receive new "1st Generation" branding, which sounds like it will be kind of confusing. Intel isn't talking about launch dates or core counts yet, but at least we know there will be updates across the board later this year.

Intel China confirmed the company's plans via the usual methods; a post on the social media site Bilibili. The translated version, via Tom's Hardware, shows a complete list of revamped products spanning the company's product stacks for desktop and mobile, and confirms several previous rumors. The first is that the upgraded Raptor Lake for desktop will indeed be named 14th Generation, which is odd since it's likely just a refresh of the existing CPUs. However, Intel may throw us a bone and bump up core counts in addition to clocks for some of its i5 and i7 SKUs. The chart confirms it'll be the same lineup as what we have now; Raptor Lake desktop parts in i3, i5, i7, and i9 versions. There will also be upgraded high-end mobile parts with the HX moniker, and all of these will be labeled Intel 14th generation, both desktop and mobile.

[weatheriscool]

Micron Announces GDDR7 for GPUs Coming in First Half of 2024
Though it'll arrive just in time for mid-cycle refresh from AMD, Nvidia, and Intel, it's unclear if there will be any takers just yet.
By Josh Norem June 29, 2023

Next-gen GPUs just got a shot in the proverbial arm thanks to a new announcement from Micron about its GDDR7 plans. The company's fiscal Q3 earnings call shed new light on the much-hyped memory standard, which is set to replace GDDR6/X at some point in the future. The company says it is on track to deliver GDDR7 to the market in the first half of 2024, which is good news for those still holding out on a GPU upgrade.

The new memory standard promises to offer a massive increase in bandwidth, even for midrange GPUs with wee 128-bit memory buses. This technology would have been useful for the beleaguered RTX 4060 and Radeon RX 7600, as it would have practically doubled the available memory bandwidth depending on the memory clocks. Micron is targeting 36G/bs of bandwidth per pin, according to Wccftech. That compares with the current 22Gb/s GDDR6X used by Nvidia, with AMD settling for 20Gb/s modules for its 7900 series GPUs. On the call, the company stated, "We plan to introduce our next-generation G7 product on our industry-leading 1ß node in the first half of calendar year 2024."

https://www.extremetech.com/gaming/micr ... lf-of-2024

[weatheriscool]

Intel reports high-rendering graphics with low-power GPUs
https://techxplore.com/news/2023-07-int ... -gpus.html
by Peter Grad , Tech Xplore

Setting its sights on evolving graphics processing units in a growing universe of generative AI, Intel announced the release of several papers outlining efforts it is pursuing in what observers say is a multibillion-dollar opportunity in coming years for the semiconductor chip giant.

Intel is presenting seven papers over three conferences covering advances in computer graphics.

The first papers were formally presented at last month's joint conference conducted by the High Performance Graphics (HPG) forum and the Eurographics Symposium on Rendering at the the Delft University of Technology in the Netherlands. The remaining papers will be discussed at a conference to be held by SIGGRAPH (Special Interest Group on Computer Graphics and Interactive Techniques) in August.

A key focus is on how to improve historically heavy graphics-rendering processes.

The papers discuss two processes in particular, ray tracing and path tracing. Both are used for recreating realistic images, especially in gaming, where accurate representation of the physics of light is critical for natural-looking imagery.

Ray tracing applies algorithms to track the trajectory of light waves and calculate color values, reflections and shadows. The enormous processing power required for real-time rendering is so great that frame rates often take a noticeable hit.

Path tracing can require even heavier processing. It follows multiple rays of light, tracking paths as they reflect off surfaces and interact with lighting among other elements. A process known as Monte Carlo integration helps determine accurate color and shading values.

Intel says these tracing methods can be accomplished more efficiently. One of its papers, "Sampling Visible GGX Normals with Spherical Caps," describes an innovative approach to calculating hemispheric items that accomplished "systematic speed-ups in our benchmarks."

Another paper reveals a 500% speed improvement in renderings of "glittery" objects such as speckled car paints, snow, molded plastics and running water. "Real-Time Rendering of Glinty Appearances using Distributed Binomial Laws on Anisotropic Grids" explains that current approaches achieve stunning realism but "come at a very high cost" in terms of processing power and speed.

In a paper to be discussed at the August SIGGRAPH conference, Intel will review advances in neural graphics, an approach the company says "is revolutionizing the graphics field." It is used to quickly scale high-quality graphics across games and movies.

[weatheriscool]

Intel's Meteor Lake Architecture to Offer Much Higher iGPU Clocks
Due to the platform's mobile-only focus, improved iGPU performance is seemingly a key objective.
By Josh Norem July 12, 2023

Intel is putting the final touches on its upcoming Meteor Lake laptop architecture and has sent qualification samples to some folks. Now a hardware tipster is reporting on the clock speeds of these early samples, and it seems like iGPU performance could be vastly improved over previous offerings thanks to much higher clock speeds. Clocks for the CPU tile are also high, but as a mobile part, not quite as high as current desktop chips. Still, it seems Intel is focusing on mobile GPU performance, which makes sense given the ubiquity of its CPUs in the laptop market and stiff competition from AMD's Phoenix APUs.

The informant du jour is a Chinese hardware journalist who goes by the handle Golden Pig Upgrade. This person has been known to offer pre-release information about new hardware, and it now appears they have had some hands-on time with a Meteor Lake CPU or know someone who has one. According to their post on the social media site Bilibili, they had access to a Meteor Lake qualification sample, which featured a configurable TDP that ranged from 20W to 65W. Additionally, the CPU tile featured a maximum boost clock of 4.8GHz, while the iGPU maxed out at 2.2GHz.

https://www.extremetech.com/computing/i ... gpu-clocks

[weatheriscool]

Samsung Beats Micron, Announces GDDR7 Memory for Future GPUs
The company has seemingly beaten Micron to the punch in delivering the next memory standard for HPC, AI, and of course, next-gen GPUs.
By Josh Norem July 19, 2023

Samsung has announced it has completed developing the world's first GDDR7 memory module for high-performance computing (HPC). This paves the way for its adoption by both AMD and Nvidia in future GPUs, which will be arriving in 2024 and beyond. It also marks a turning point for the industry, which has used GDDR6 since 2018, when it first appeared in Nvidia's Turing line of graphics cards. It promises a major leap in performance and efficiency over GDDR6 and is exciting news for fans of pure PC power.

https://www.extremetech.com/gaming/sams ... uture-gpus

[Tadasuke]

Bergamo is on average 2x faster than Sapphire Rapids in server workloads, which is quite a lot.



source: https://www.phoronix.com/review/amd-epyc-9754-bergamo

There are now 128-core 256-thread x86 CPUs in the market. That's 2^7 cores. Seven doublings since the Athlon XP 20 years ago. 512 threads with 2 sockets. Probably 1024 in 2 years.

[Tadasuke]

Bergamo is visibly (by 50%) faster than Genoa in V-Ray and massively (2.37x increase) faster than Sapphire Rapids in V-Ray. One 96-core Genoa chip is enough to match TWO 60-core Sapphire Rapids chips. No wonder Intel is losing money... seems like Intel is defeated. Sapphire Rapids was delayed by 2 years, it was supposed to come out at the beginning of 2021. Sapphire Rapids refresh won't help. AMD is already preparing their next-gen Turin CPUs.





source: https://www.igorslab.de/en/amd-epyc-975 ... ids-chips/

[Tadasuke]

Newest Ryzen 9 7945HX3D is a 16-core Zen 4 laptop CPU with a base clock of 2.3 GHz and boost up to 5.4 GHz. This means that the base clock is 200 MHz lower than the existing Ryzen 9 7945HX. The CPU has a TDP up to about 75 watts and is equipped with 1 MB L1, 16 MB L2 and 128 MB (2x more than the non-V-cache 7945) of L3 cache. Supports 64 GB of DDR5 with 83.2 GB/s bandwidth. Integrated iGPU Radeon 610M with 128 SPs, 8 TMUs, 4 ROPs and 563.2 gigaflops. AMD claims that their new SKU will be on average 15% faster than the existing 7945HX SKU.

[weatheriscool]

Intel: 3nm Node Has Met Performance Targets for 2024 Launches
The company is prepping two server chips for 2024 using Intel 3.
By Josh Norem July 31, 2023
https://www.extremetech.com/computing/i ... 4-launches

Intel made headlines last week when it unveiled its earnings for the second quarter of 2023. The company announced it had finally returned to profitability after back-to-back quarters of losses, marking a surprising turnaround for the former 800-pound gorilla of chip making. Buried inside the lengthy earnings report was an update on its future nodes, including its Intel 3 process, expected to debut in 2024. The company announced it has already hit its yield and performance targets for its newest node, which will theoretically allow it to launch both Granite Rapids and Sierra Forest server chips on schedule next year.

To backtrack briefly, Intel is currently using Intel 7 (formerly 10nm) for Raptor Lake, with Intel 4 (formerly 7nm) ramping now for Meteor Lake. After that, there’s Intel 3 (formerly 5nm), which is seen as an evolution of Intel 4, and it's the company’s second generation of extreme ultraviolet lithography (EUV) CPUs. CEO Pat Gelsinger stated in the earning call that the company’s first EUV node, Intel 4, is “essentially complete” as it’s currently ramping production. Moving on to Intel 3, he noted it’s already hit the company’s defect density (yield) and performance targets and is on track to hit the final targets for both metrics in 2024.

[wjfox]

[weatheriscool]

Apple Bought All of TSMC's 3nm Capacity for an Entire Year
A new report sheds some light on Apple's quest to be the only tech company in the world offering 3nm products to its customers.
By Josh Norem August 8, 2023
https://www.extremetech.com/computing/a ... ntire-year

In the final days of 2022, TSMC announced it had started production on its highly-anticipated 3nm process, which it calls FinFlex. Several months later, it was reported that Apple was the first TSMC customer to open up its checkbook for the fledging node, as it had reportedly purchased the entirety of TSMC's first run of 3nm wafers. That wasn't a surprise, as Apple has always been the first company to utilize the most advanced node TSMC offers. However, a new report shows that Apple has gone even further than just buying that first batch, as it has secured all of TSMC's 3nm production for the first year of its production.

Details on the arrangement between the two companies regarding 3nm wafers are included in a new report from The Information, which was flagged by ArsTechnica. It states that Apple was eager to invest in TSMC's most advanced node to offer its iPhone and MacBook customers A17 Bionic and M3 chips. However, Apple needed assurances that it wouldn't be stuck paying for defective dies on the new wafers, so TSMC agreed to let it off the hook and only charge it for perfect chips. Notably, the new report says TSMC's yields for N3 are already at 70%, which is higher than previous reports of it being around 55%.

[weatheriscool]

Screenshots Confirm AMD’s First Hybrid Processor as ‘Strix Point’ APU
The mobile chip will feature a 4+8 P-core and E-Core configuration.
By Josh Norem August 10, 2023
https://www.extremetech.com/computing/s ... -point-apu

AMD has long been toying with the idea of going head-to-head with Intel in the hybrid processor game. With Zen 4, it stuck to its standard P-core-only design for both mobile and desktop chips, but now it appears to be testing a true hybrid chip with both P-and-E cores for the next iteration of its Zen platform. This was rumored previously but is seemingly confirmed by new screenshots showing an AMD engineering sample with a hybrid design.

ITHome has posted two screenshots from performancedatabases.com that show CPU-Z and HWINFO. They purport to show AMD's first hybrid processor in action. When this chip popped up with 12 cores, we incorrectly guessed it sported 8 P-cores and 4 E-cores, but it’s the opposite. According to the screenshots, it’s a 4+8 design for 12 cores and 24 threads, and it runs at 45W while being manufactured on TSMC’s 4nm process. It will be branded as a Ryzen 8000 APU code-named Strix (Point) for AMD’s Zen 5 platform, seemingly due in 2024. AMD started providing developers with documentation about its hybrid architecture in March.

[weatheriscool]

MediaTek Announces First TSMC 3nm SoC With 60% Boost in Logic Density
The company is the first to announce it's using TSMC's most advanced process for an upcoming SoC.
By Josh Norem September 8, 2023

Ever since TSMC announced it was beginning production of its next-generation 3nm process at the very end of 2022, we've been waiting to see which company would be first out of the gate with a consumer product utilizing its most advanced technology. It was always assumed it would be Apple since it reportedly bought pretty much all of TSMC's 3nm capacity for 2023, but now MediaTek has beaten it to the punch, on paper, by announcing its own 3nm SoC.

https://www.extremetech.com/mobile/medi ... ic-density

[wjfox]

Nvidia GeForce RTX 5080: Everything We Know

Updated: Sep 8, 2023 11:09 pm

The Nvidia GeForce RTX 5080 looks to be a couple of years away, as the RTX 4090 is still going strong. Nvidia hasn’t officially announced the card or the next-generation architecture it will be built around, but there are a few key details we’ve gleaned over the last several months.

Whenever it appears, it will have big boots to fill. Nvidia’s GeForce RTX 4080 is one of the best graphics cards with its superb performance, ray tracing chops, and vigorous frame rate-boosting DLSS 3. Its hefty $1,199 price tag raised a few eyebrows, but if the RTX 5080 launches with a similar slate of improvements, we could be in for another pricey top-tier card.

Since 2016, Nvidia has launched a new generation of RTX cards every two years, with the latest appearing in 2022. While that would put the 50 series on track for a 2024 release, things look to be shaping up a little differently this time around.

Nvidia’s next-generation GPU architecture is on track to release in 2025, according to a development roadmap that was shared during a press presentation and captured by German site Hardareluxx. That would put it a tad later than expected, but only by a year.

Whenever the 50 series does eventually arrive, however, we can expect the 5080 to be in its starting line-up. Nvidia tends to lead each generation with its higher-tier cards. The RTX 4080 was the first of the 40 series to hit shelves, for instance, followed soon after by the RTX 4090. Similarly, the initial launch of the 30 series consisted of the RTX 3070, RTX 3080, and RTX 3090. There seems little reason this coming generation won’t follow suit.

https://www.ign.com/articles/nvidia-gef ... ng-we-know

[weatheriscool]

AMD's First Hybrid Processor Has Been Photographed
The long-rumored Phoenix 2 APU has finally been revealed.
By Josh Norem September 11, 2023
https://www.extremetech.com/computing/a ... otographed

AMD has been talking about following Intel's lead in the hybrid processor game for about two years, but so far, there has been nothing to show for it. The company's current Phoenix mobile APUs use standard Zen 4 dies, while its newest server chip uses smaller Zen 4c chips, yet the two have yet to be combined. A new photograph of a hybrid AMD processor has appeared, combining both Zen 4 and Zen 4c dies and a tiny GPU, making it the company's next-generation Phoenix 2 APU.

The naked die appears on Twitter via a prolific hardware leaker named @9550pro/HXL. It shows both Zen 4 and Zen 4c cores side-by-side for the first time, indicating AMD is set to launch this new piece of hardware soon, focusing on efficiency over outright horsepower. According to a further breakdown provided by HXL, the Zen 4c cores are roughly 35% smaller than their big brothers. Yet, they offer the same instructions-per-clock (IPC) along with the same ISA as Zen 4, a different approach compared with Intel. So far, the Zen 4c dies have only appeared in its newest server chip, the 128-core Epyc Bergamo CPU.

[weatheriscool]

AMD Unveils Final 4th Gen Epyc CPU Code-Named Siena
The company is focusing on efficiency with this 64-core, Zen 4c-based CPU for the 'intelligent edge.'
By Josh Norem September 18, 2023
https://www.extremetech.com/computing/a ... amed-siena

AMD has announced the final CPU in its 4th generation lineup of Zen 4 processors for the data center with its new Siena line of 8004 CPUs. This lineup will occupy the bottom rung on the performance ladder in exchange for maximum efficiency, as it's designed for smaller deployments where cost and square footage are paramount concerns. The addition of Siena to the Epyc lineup plugs a big hole in the lineup, as all the previous processors were designed for either maximum performance or maximum thread counts. Siena tops out at just 64 cores and exclusively uses the company's efficiency-focused Zen 4c cores.

To quickly recap, the 4th gen Epyc family includes four CPU lineups. There's Genoa, which is the "standard" server CPU with 96 Zen 4 cores. Then there's the Genoa-X, which is Genoa but with V-Cache for 1.1GB of L3 cache. For workloads where maximum core count is required, there's Bergamo, which has 128 Zen 4c cores for 256 threads. Then there's the new kid, Siena, which offers "balanced performance" along with superior performance-per-watt compared to its competitors. AMD is targeting cloud service, intelligent edge, and telcos with Siena.

[weatheriscool]

China Makes Huge Chip Breakthrough – 7 Nanometers Without EUV Lithography Machines
September 18, 2023 by Brian Wang

Huawei’s Kirin 9000S system-on-chip powers Huawei’s new Mate 60 Pro smartphone reportedly is using 2nd generation 7nm-class fabrication process and stacking made by China-based SMIC.

Huawei was known to have been stockpiling chips from its HiSilicon unit before TSMC cut ties to comply with US sanctions. TSMC started making 7 nanometer chips back in 2017.

https://www.nextbigfuture.com/2023/09/c ... hines.html

[weatheriscool]

Intel Unveils Meteor Lake Architecture for AI, Efficiency, and Better Graphics
The company's 1st generation tile-based design is practically brand new from the ground up.
By Josh Norem September 19, 2023

At today’s Innovation event, Intel is finally pulling the tarp off its Meteor Lake architecture. The company says it represents the most significant change to a CPU architecture it’s made in 40 years, and this design will be the foundation of its CPUs for the next decade, at least. The list of changes is long, and plenty of ink has already been spilled about Meteor Lake, so this article will just look at the highlights.
Intel Goes Tile-Based

Meteor Lake uses four tiles to comprise the CPU instead of a monolithic design, a first for Intel. The four tiles—CPU, GPU, SoC, and I/O—use three nodes from two different foundries. Intel made the CPU tile on its own Intel 4 process, formerly known as 7nm, while TSMC made the other tiles. The SoC and I/O tiles are built on the N6 process, while the Arc GPU tile is made using TSMC’s 5nm node.

https://www.extremetech.com/computing/i ... r-graphics

[weatheriscool]

Report: TSMC Could Push 2nm Node Back to 2026
Intel's plan is beginning to come together.
By Josh Norem September 21, 2023

TSMC has just begun volume production of its 3nm manufacturing node, as it's been busy cranking out A17 Pro SoCs for the iPhone 15 Pro and Pro Max. Meanwhile, it's also been hard at work on that node's successor, which will ditch FinFET transistors for a 2nm gate-all-around (GAA) nanosheet design. In April, the company stated that it expected to begin production of 2nm chips in 2025, but a new report from Taiwan says that process might be pushed into 2026 due to various factors. If that occurs, it would leave an opening for Intel and Samsung to leapfrog their biggest rival, assuming they don't also experience delays.

https://www.extremetech.com/computing/r ... ck-to-2026

[weatheriscool]

[img]Intel to Add Vertically-Stacked Cache in Future CPUs, Just Like AMD
Intel says it'll be using on-die cache, but its approach will differ from AMD's in some way.
By Josh Norem September 20, 2023
https://www.extremetech.com/computing/i ... t-like-amd

We all know imitation is the sincerest form of flattery, and as such, Intel's CEO just paid its rival AMD the biggest compliment possible. At the company's innovation event this week, Pat Gelsinger made news by stating Intel will indeed be adopting on-die cache similar to AMD's 3D V-Cache in future products. That juicy slice of additional L3 cache didn't make it into the company's newest architecture, code-named Meteor Lake, but will instead arrive in the future, according to Gelsinger. This revelation marks the first time Intel has acknowledged such a change to its roadmap, but it remains unclear when it will arrive and what form it will take.