GPU and CPU news and discussions

[weatheriscool]

Intel Is Allegedly Working on a Entirely New x86 Architecture Named Cobra Core
Intel is reportedly trying out several new technologies to drastically improve single-core performance in upcoming architectures.
By Josh Norem September 16, 2024

Intel has been working towards fulfilling its "five nodes in four years" strategy since 2021, culminating in 2025 with the Intel 18A process and the Panther Lake mobile architecture. Beyond that, things start getting murky, as nobody knows precisely what Intel has up its sleeve. However, a new round of rumors points to an entirely new x86 architecture being in development with radical new features aimed at boosting single-core performance while also improving efficiency.

Intel has reportedly been working on a project code-named Royal Core/Cobra Core for some time now, though the details of the architecture remain mostly unknown. It has been listed as a new x86 architecture via an Intel employee's LinkedIn profile, summarized by Videocardz. The original source notes Arrow Lake is just a bridge to Royal Core, which will first appear in 2026's Nova Lake platform. Cobra Core will then succeed it in 2027 or so.

https://www.extremetech.com/computing/i ... amed-cobra

[weatheriscool]

World's Fastest Memory Appears at 9,600MHz, 10,000MHz on the Way
These CUDIMMs could provide quite a performance boost to next-gen Intel systems.
By Josh Norem September 17, 2024
https://www.extremetech.com/computing/w ... on-the-way

DDR5 memory arrived on the consumer market in 2021 with Intel's 12th Generation Alder Lake CPUs. It was then given a sizable boost in 2022 with the arrival of AMD's Zen 4 platform. However, back then, DDR5 was expensive, hard to find, and not much faster than its DDR4 predecessor. That speed discrepancy is finally starting to change in 2024, though, as DDR5 pulls ahead. Now, we have a world's first on that front: DDR5 memory that runs at a blistering 9,600MHz.

The memory in question comes from a Chinese company named Asgard, which announced it on the WeChat app. In the announcement, it humorously states that its new Thor-branded DDR5-9600 memory is far superior to the existing DDR5-9200 modules from its competitors. The memory sticks are CUDIMM or Clocked Unbuffered DIMM, and Asgard says each stick contains a Clock Driver that regulates data transmission from CPU to DRAM.

[Tadasuke]

The first Lunar Lake laptops are expected to start shipping as soon as September 24th. They will offer up to 24 hours of battery life depending on what the user is making the laptop do. Lunar Lake is a large upgrade over the previous generation (Meteor Lake) and probably a bit better than the current competition from AMD, Apple and Qualcomm. This is to no surprise, as they will be using the most advanced lithography machines and the newest architectures. Lunar Lake will be produced in both Intel and TSMC foundries in multiple countries on multiple continents.

Lunar Lake is not a minor re-iteration, but a relatively major new product line, able to deliver noticeably more than Raptor Lake or Meteor Lake at a constant power envelope (for example 25 watts for the whole SoC). XeSS supersampling will work on the new chips. When XeSS is supported by a game and working as intended, you see more frames per second than you would without it, with little to no discernible loss in visual quality. The tech has been improved since Meteor Lake launched, and Intel claims the new Lunar Lake chips can deliver around 30% better gaming performance thanks to integrated GPUs built on a new Xe2 architecture, that is more efficient than the Xe architecture. All of these Lunar Lake laptops have an NPU capable of at least 40 TOPS, so as long as you buy one with at least 16GB of RAM (which is really the minimum in 2024) they all should support the AI features of Copilot+.

[weatheriscool]

First Intel Lunar Lake Laptop Test Shows Almost 24-Hour Battery Life
It offers more than twice the battery life of the older Corer i7 mobile CPU.
By Josh Norem September 24, 2024
https://www.extremetech.com/computing/f ... ttery-life

Intel first discussed the mobile architecture it calls Lunar Lake way back in June, and stated at the time it would arrive in Q3. The initial reveal included some lofty performance claims, so we've all been waiting to find out if those numbers were marketing BS or the real deal. Now it appears Lunar Lake laptops are finally in reviewers' hands, and our sister site PCMag has published a "first look" of sorts for a laptop from Asus that shows it offers almost 24 hours of battery life.

The laptop PCMag received for testing is the Asus Zenbook S 14, which sports the Intel Core Ultra 258V, which is an upper-midrange version of the CPU. All Lunar Lake CPUs sport a 4+4 core configuration with four P-cores and four E-cores, with only clock speeds, power draw, and graphics changing slightly between them. The point being performance should be quite close between Lunar Lake CPUs, as they are all very similar in terms of specs. In the site's battery rundown test, it looped a 720p video at 50% brightness and found it was capable of a whopping 23 hours and 23 minutes of battery life.

[weatheriscool]

Intel Announces Next-Gen Xeon 6 CPUs With 128 Performance Cores
The company has removed the word 'Scalable' for its newest Xeons, which are now available in both P-Core and E-Core flavors.
By Josh Norem September 24, 2024

Today, Intel is pulling the wraps off its sixth-generation Xeon CPUs outfitted exclusively with performance cores, or P-cores in Intel parlance. These chips go by the codename Granite Rapids and will replace its fifth-generation "Scalable Xeon" Emerald Rapids CPUs in the company's lineup for high-performance computing (HPC). These CPUs will join forces with the previously announced efficiency core Xeon 6 Sierra Forest CPUs to round out the company's data center lineup. Both Sierra Forest E-core CPUs and Granite Rapids P-core chips are Xeon 6, with the letters E or P in the name to differentiate them.

https://www.extremetech.com/computing/i ... ance-cores

[Tadasuke]

Geometric mean of all test results (done in Linux OS) for some selected Intel and AMD server CPUs:



As you can see, Sapphire Rapids (2022) is on average 56.3% faster than Ice Lake (2021), Emerald Rapids (2023) is 18% faster than Sapphire Rapids (2022) and Granite Rapids (2024) is 38.3% faster than Emerald Rapids (2023). Granite Rapids is 155% faster than Ice Lake after more than 3 years while consuming around 50% more energy, which means about 70% better energy efficiency. Perhaps will get better to maybe 75-80% with time. My prediction of +20% energy efficiency and +20% performance/price per year on average seems to be quite accurate. This would mean doubling Ice Lake in 2025.

source: https://www.phoronix.com/review/intel-x ... 0p-power/7



And here are various benchmarks of AmpereOne A192-32X 192-core ARM Server CPU which is very competitive to x86 CPUs, AMD also has now 192-core CPU using their Zen 5 architecture: https://www.phoronix.com/review/ampereone-a192-32x/6

[Tadasuke]

I read these reviews of Intel Lunar Lake's performance and efficiency .... drudging, irksome, tiresome, mundane and totally unexciting.

https://www.notebookcheck.net/Intel-Lun ... 405.0.html

https://www.notebookcheck.net/Intel-Lun ... 167.0.html

Like I wrote in the Singularity thread, geomean year over year efficiency gains in computer hardware will be +19% and I asses that recent hardware reviews only confirm this. There's no point in further checking, reading, watching or listening. Nothing especially exciting is going to happen.

[Tadasuke]

Ultra 5 245K will have 14 cores 14 threads
Ultra 7 265K will have 20 cores 20 threads
Ultra 9 285K will have 24 cores 24 threads

They will launch on October 24th according to Intel. In GeekBench 6, Ultra 7 265K allegedly has 2.7% higher multi-core result and 0.8% lower single-core result than Ryzen 9 9950X which is already widely available. I expect around 40% better efficiency than Raptor Lake which came out in 2022. So the 285K might have 12% better performance and 20% lower power usage which is exactly +40% performance/power usage.

[weatheriscool]

Intel Is Now Working on Its Third-Generation 'Celestial' GPU Architecture
The company appears to be mostly finished with its second-generation 'Battlemage' GPUs, even though they've yet to launch in discrete form.
By Josh Norem October 11, 2024

https://www.extremetech.com/gaming/inte ... chitecture
Intel famously entered the GPU market in 2022 and presented a roadmap of four architectures it would develop over time. It has already launched the first-generation offering in this family, which was the Arc Alchemist lineup from 2022. It followed that up with its second-generation Battlemage GPUs in its new Lunar Lake mobile CPUs. Now, the company is apparently already working on its third-generation Celestial architecture.

The revelation that work on Celestial is already happening comes from Phoronix, which is a site dedicated to the Linux operating system. The site reports that Intel engineers who work on Linux drivers have posted documentation that features references to Xe3, or third-generation Intel Xe graphics. There are also several mentions of PTL, which is Intel code for its next-generation Panther Lake mobile platform. Just like Intel's Battlemage first appeared in Lunar Lake, Celestial is expected to debut in Panther Lake in late 2025.

[weatheriscool]

AMD Announces 5th Gen Epyc CPUs and MI325X AI Accelerators
The company is now unleashing Zen 5 on the data center.
By Josh Norem October 10, 2024
https://www.extremetech.com/computing/a ... celerators

AMD has officially unveiled its next-generation data center CPUs and GPUs at an event in San Francisco dubbed Advancing AI. The 5th generation Epyc CPUs will come in two distinct configurations, which are now part of the same 9005 family instead of being given different code names, such as Genoa and Bergamo for its 4th generation CPUs. The company's CEO also showed off its second-generation MI325X AI accelerator while noting that the MI350X and MI400X are in development.

Its Zen 5 Epyc CPUs come in two flavors, using either regular Zen 5 cores or efficiency-focused Zen 5c cores. The CPUs with regular Zen 5 cores can scale up to 128 cores and 256 threads and are made with TSMC's 4nm process with 16 Zen 5 CCDs. The Zen 5c version will offer up to 192 cores and 384 threads via 12 CCDs and is made on TSMC's 3nm process. Both versions also feature a TSMC 6nm I/O die, and AMD says 5th gen Epyc features 150 billion transistors.

[Tadasuke]

EPYC 9965 (192 Zen 5C cores) has an average power consumption of 275 W
EPYC 9755 (128 Zen 5 cores) has an average power consumption of 324 W
Xeon 6980P (128 Redwood Cove cores) has an average power consumption of 322 W (332 W for MRDIMM-8800)
Xeon 8592+ (64 Raptor Cove cores) has an average power consumption of 260 W
EPYC 9654 (96 Zen 4 cores) has an average power consumption of 245 W

[weatheriscool]

[firestar464]

More on that by MoneyControl:

https://www.moneycontrol.com/technology ... 42005.html

[weatheriscool]

First Intel Core Ultra 9 285 Non-K Benchmark Shows Surprisingly Good Performance
At just 65W, it could be the new bang-for-the-buck performance champ.
By Josh Norem October 21, 2024

Intel announced its new "Arrow Lake" Core Ultra 200 series CPUs almost two weeks ago. Per tradition, it will launch with the K-series chips first, which feature an unlocked multiplier for overclocking. Those CPUs will eventually be followed by non-K SKUs, which cannot be overclocked at all, so they are mainstream CPUs for the unwashed masses. Now, we have the first benchmark score for the flagship Core Ultra 285 non-K CPU, and its performance looks impressive, especially for a 65W chip.

The Core Ultra 285, which has not been announced by Intel at this time, has appeared in the Geekbench database, and its scores are surprisingly good. According to the listing, this chip has the same 24 cores and threads as the K part, but it features a 100MHz reduction in maximum boost clock at 5.6GHz compared with 5.7GHz. Its base clock is also drastically lower at 2.5GHz versus 3.7GHz for the 285K. And it's a 65W CPU instead of 125W for the 285K, so it offers a major reduction in power (and clocks) for what is allegedly an already efficient CPU.

https://www.extremetech.com/computing/f ... erformance

[Tadasuke]

There is some improvement in AI and science performance:













The 19% per year efficiency or perf/price gain I wrote about seems to be true.

[weatheriscool]

AMD Ryzen 7 9800X3D to Offer Up to 5.3GHz Boost Clock
It's much higher than those of its Zen 4 predecessor.
By Josh Norem October 25, 2024
https://www.extremetech.com/gaming/amd- ... oost-clock

AMD sent out a teaser earlier this week saying a Zen 5 CPU with V-Cache was coming on Nov. 7. The company has big shoes to fill, as its current Ryzen 7 7800X3D is currently the world's fastest gaming CPU. However, we have confirmation that AMD is nudging clock speeds upwards on the first Zen 5 X3D chip, with new Geekbench scores now showing its boost clock at a relatively high 5.3GHz.

It's been assumed that AMD would launch the eight-core version of Zen 5 X3D first, as that configuration always performs the best in its lineup. Plus, it needs to put its best foot forward right now, given how the Zen 5 launch went for the company. The existence of an eight-core X3D CPU is now confirmed by new Geekbench scores, which show a Ryzen 7 9800X3D CPU was tested by Asus and Biostar recently. According to Videocardz, this is some kind of internal testing, as it says it's not aware of any press having these chips yet.

[weatheriscool]

AMD Ryzen 7 9800X3D Overclocks to 6.9GHz In CS2
It just takes liquid nitrogen and a little know-how.
By Josh Gulick November 8, 2024

https://www.extremetech.com/computing/a ... ghz-in-cs2
We’re getting a preview of the AMD Ryzen 7 9800X3D’s tolerance for extreme overclocking, thanks to Asus China’s charismatic general manager, Tony Yu. Relying on his extensive experience (and plenty of liquid nitrogen), Yu pushed AMD’s new heavyweight processor to a startling 6.9 GHz. It looks like the Ryzen 7 9800X3D is going to be a high priority for gamers—especially those who dabble with overclocking. Even without liquid nitrogen, users are going to be dialing up the new CPU’s frequency.

We’ve seen Yu’s informative (and entertaining) videos before. Last year, Yu posted a clip to video sharing service Bilibili in which he ran an Nvidia RTX 4090 off the M.2 slot of a RTX 4060 Ti GPU. For the Ryzen 7 9800X3D, he broke out the liquid nitrogen and started putting up eye-catching numbers in Counter-Strike 2.

[wjfox]

TSMC to stop supplying advanced AI processors for all of its China customers: Report

Chips at 7nm and below are affected.

By Anton Shilov
published 23 hours ago

After unintentionally producing an AI chiplet for Huawei via a proxy, TSMC is set to stop supplying sophisticated AI processors for all of its Chinese clients from Monday, November 11, reports Financial Times, citing Ijiwei.com, which in turn cited emails sent by the foundry to its customers. This change concerns advanced process technologies (such as 7nm and below) and will have a significant impact on China-based developers of AI processors. This does not, however, mean that TSMC will cease to serve these customers completely.

The new restriction is limited to AI processors and AI GPUs made on 7nm-class and lower nodes. Smartphone processors, chips for automotive applications, and other devices that cannot be used for military or dual-use applications are not affected, the Ijiwei report stated. Sources with knowledge of the matter reportedly told Ijiwei that future supplies of advanced AI processors to China-based entities would require some kind of approval process, which likely involves specialists from the U.S. Department of Commerce.

It is unclear what TSMC plans to do with wafers containing advanced processors already in production, as well as with already processed wafers sitting in its stock. Perhaps, they will be shipped to actual customers if they get appropriate export licenses.

https://www.tomshardware.com/tech-indus ... ers-report



Credit: Baidu

[Tadasuke]

AMD Ryzen 7 7700X, 9700X, 7800X3D and 9800X3D geomean video game framerate comparison:



9700X (2024) is 10% faster than 7700X (2022)
7800X3D is 36.7% faster than 7700X
9800X3D is 56.9% faster than 7700X, 42.5% faster than 9700X and 14.8% faster than 7800X3D

source: https://www.computerbase.de/artikel/pro ... eich.90208

9800X3D power consumption when gaming:


9800X3D application power consumption:


And here's TechPowerUp's 9800X3D review summary:


I think it's rather disappointing to be honest. (ʘ ʖ̯ ʘ)

[Tadasuke]

Intel i7 from 2010 to today and AMD Ryzen 7 CPU from 2017 to today single-thread PassMark score:



Of course PassMark is not the only benchmark out there. But it's something. Probably quite representative of most use-case scenarios.