GPU and CPU news and discussions

[weatheriscool]

Intel Core i9-14900KF Breaks Single-Core Speed Record in PassMark
Intel has yet to talk much about its upcoming desktop CPU lineup, but it's already putting up impressive numbers.
By Josh Norem September 22, 2023
https://www.extremetech.com/computing/i ... n-passmark

This week, Intel had its Innovation event, where the company pulled the wraps off a dizzying array of next-gen technology. Notably absent was any discussion of its Raptor Lake refresh CPUs, which was unexpected. Perhaps a "refresh" isn't something it wanted to highlight at the event, which focused on new manufacturing nodes, engineering breakthroughs, and the future of computing. Regardless, it seems the company's upcoming 14th Generation CPUs will be pretty potent, as the flagship CPU is now the fastest single-core CPU ever, at least in the PassMark benchmark.

The Core i9-14900KF CPU was recently tested in PassMark and ended up eclipsing its predecessor—the Core i9-13900K—by a tidy 3.5%. Obviously, it's not a huge gap, but since the core configuration is unchanged at 24 cores and 32 threads, we can chalk it up to its massaged clock speeds. Whereas the former CPU topped out at 5.8GHz, the revamped 14th Gen version goes up to 11 6GHz. That lofty frequency was previously reserved for the spendy 13900KS, which was made just for one-off benchmarks and to show that Intel is still the king of overclocking.

[weatheriscool]

Meteor Lake Is Coming to Desktops, Intel Confirms
The mobile version will arrive in December, with a desktop version to follow in 2024.
By Josh Norem September 25, 2023
https://www.extremetech.com/computing/m ... l-confirms

Last week, Intel held its annual Innovation event, showing off upcoming hardware and titillating the crowd with naked wafers. At the show, Intel finally unveiled the long-awaited Meteor Lake architecture and gave it a launch date—Dec. 14. Given that so far, all of the leaked information about its various Core Ultra CPUs have fewer cores than a Raptor Lake desktop chip, it was assumed this was to be a mobile-only platform. However, according to Intel, that is not the case, and desktop versions will arrive in 2024.

[weatheriscool]

IBM Research’s New Prototype AI Chip With 14 Times Energy Efficiency
September 25, 2023 by Brian Wang

IBM showed it’s possible to build analog AI chips that can handle natural-language AI tasks with an estimated 14 times more energy efficiency. Researchers from IBM labs around the world presented their prototype analog AI chip for energy-efficient speech recognition and transcription. Their design was utilized in two AI inference experiments and the analog chips performed these tasks just as reliably as comparable all-digital devices — but finished the tasks faster and used less energy.

https://www.nextbigfuture.com/2023/09/i ... iency.html

[wjfox]

NVIDIA B100 “Blackwell” GPUs To Be Made On TSMC 3nm Process, Launching In Q4 2024

Sep 26, 2023

NVIDIA has reportedly finalized a deal with TSMC to utilize its 3nm process node for the creation of its next-gen Blackwell B100 GPUs for AI.

The rapid rise in genAI developments has indeed bolstered the pace of innovation within the industry. It is now reported that NVIDIA has contacted TSMC for its next-generation "Blackwell" B100 AI GPUs, based around the Taiwan giant's 3nm process.

Before diving into the report, it is important to go down the memory lane, to get to know about Blackwell GB100 GPUs. They are set to be Team Green's first batch of AI GPUs to feature a chiplet design, which is expected to bring decisive uplifts in terms of performance per watt. It has been reported in the past that GB100s will feature a much more complex "structural design", which may setback production figures in the beginning.

https://wccftech.com/nvidia-b100-blackw ... n-q4-2024/

[weatheriscool]

Samsung Officially Launches the 990 Pro SSD in 4TB Capacity
The company has caved to the demands of storage aficionados who demanded a 4TB drive.
By Josh Norem October 3, 2023
https://www.extremetech.com/computing/s ... b-capacity

In what can only be referred to as a "storage encore," Samsung released a 4TB version of its flagship NVME SSD almost a year after it launched in smaller capacities. The new version of the drive comes in direct response to outcry from its customers who demanded a 4TB version of Samsung's best SSD. The arrival of the new drive marks the first time in recent memory a hardware company has listened to its fans and provided exactly what they were asking for, and it's done so at a pretty reasonable price, too.

The Samsung 990 Pro arrived in November 2022, and expectations were high since Samsung is known for producing the fastest NVME SSDs in the industry. Despite this reputation for "raising the bar," Samsung toed the line a bit with the 990 Pro by making it a PCIe 4.0 drive instead of Gen 5 like the entire world expected. In Samsung's defense, those drives are not quite ready for prime time and won't be for another year or two, at least. Intel and AMD's latest chipsets only support PCIe Gen 4 storage, though the X670E chipset does support a single Gen 5 storage device. Therefore, there's little tangible benefit in the real world in buying a Gen 5 drive versus a Gen 4 device at this time, so Samsung's decision was pragmatic.
990 Pro

[wjfox]

Nvidia May Move to Yearly GPU Architecture Releases

By Anton Shilov
published about 16 hours ago

In a bid to maintain its leadership in artificial intelligence (AI) and high-performance computing (HPC) hardware, Nvidia plans to speed up development of new GPU architectures and essentially get back to its one-year cadence for product introductions, according to its roadmap published for investors and further explained by SemiAnalysis. As a result, Nvidia's Blackwell will come in 2024 and will be succeeded by a new architecture in 2025.

But before Blackwell arrives next year (presumably in the second half of next year), Nvidia is set to roll out multiple new products based on its Hopper architecture. This includes the H200 product, which might be a re-spin of the H100 made for enhanced yields, or just higher performance, as well as GH200NVL, which will address training and inference on large language models with an Arm-based CPU and Hopper-based GPU. These are set to come rather sooner than later.

https://www.tomshardware.com/news/nvidi ... e-releases

[weatheriscool]

Samsung Lands 3nm Server Chip Contract as Battle With TSMC Heats Up
The purchaser of the next-generation chips is an unknown US company.
By Josh Norem October 12, 2023

Even though Samsung was the first silicon foundry in the world to announce its transition to 3nm technology and gate-all-around (GAA) transistors, it's been unclear if it was selling this cutting-edge technology to any customers. It's one thing to beat TSMC to the 3nm punch, but would it have any customers that wanted to be first in line for it? According to a new press release from Korea, the company has found at least one customer, and in a bit of a plot twist, it's an American company. Sadly, the identity of this company is unknown at this time.

The US firm tapped a fabless Korean company named ADTechnology to design and produce a 3nm server chip using high-bandwidth memory (HBM). ADTechnology announced it would be using Samsung's 3nm process for the order, making it the first public announcement we can recall of a 3nm chip manufactured by Samsung. Both companies describe the arrangement in glowing terms, saying it would be one of the biggest projects in the industry, but without going into too many details. The PR states it'll use a 2.5D design with logic and HBM memory placed atop an interposer.

https://www.extremetech.com/computing/s ... c-heats-up

[weatheriscool]

Intel Officially Announces Raptor Lake 14th Gen CPUs, Arriving Oct. 17
The company is also adding 'AI-Guided overclocking' to the menu this time.
By Josh Norem October 16, 2023

Intel has officially pulled the wraps off its refreshed lineup of Raptor Lake CPUs today. Despite most of them only offering a bump to their maximum turbo clock, the company has branded them 14th Gen CPUs. The company will launch three refreshed chips (with six SKUs) on Oct. 17, almost exactly one year after it first launched Raptor Lake on Oct. 20, 2022. This will effectively give Intel a second bite at the apple, as these chips will be taking on AMD's 7000-series CPUs once again, but this time, they're aimed at its V-Cache chips, which didn't exist when AMD launched Zen 4 in September 2022.

Intel's Core i9, i7, and i5 CPUs will be the tip of the spear in the upcoming silicon skirmish, with each CPU launching in both K and KF variants. The K version is the chip with integrated UHD 770 graphics, while the KF version eschews the iGPU and usually costs around $20 less because of it. The specs of the final chips match what was previously reported, with the i9 and i5 chips only receiving a turbo clock bump, while the i7 gets an additional four E-cores and more cache for better multi-core performance. As expected, the Core i9-14900K will boost to 6GHz out of the box, making it the first Intel CPU to do so without special branding.

https://www.extremetech.com/computing/i ... ing-oct-17

[weatheriscool]

First 96-Core AMD Zen 4 Threadripper Tests Show Utter Domination Over Intel
As expected, AMD's newest Ryzen Threadripper Pro 7995WX should break every record there is for CPU performance.
By Josh Norem October 20, 2023
https://www.extremetech.com/computing/f ... over-intel

AMD officially unveiled its long-anticipated Threadripper lineup for Zen 4, bringing 96 cores and 192 threads to the workstation market for the first time. To celebrate the much-hyped launch, our sister site PCMag put the flagship CPU through its paces (remotely) in several popular benchmarks. As expected, the top-dog Threadripper Pro CPU shredded the competition, indicating it would fulfill the prophecy of being the fastest workstation CPU ever, just like its 64-core, 128-thread predecessor.

Dell provided a remote link for our colleagues at PCMag to test the flagship CPU in the Threadripper Pro line; the 96-core 7995WX. Though this isn't as fun as testing it first-hand, Dell said the remote link only drops performance by 1-3%. This is a 350W CPU—just like every Threadripper Pro Zen 4 part—installed in a Dell Precision 7875 workstation with two RTX 6000 Nvidia workstation cards. For comparison, it was tested against two workstations with the fastest current-generation parts. Those include an HP Z8 Fury workstation with a 56-core Xeon w9-3495X CPU, Intel's fastest current workstation SKU. The previous generation 64-core Threadripper Pro 5995WX Zen 3 chip was also added for comparison via a Lenovo PC. One big disparity between the three machines was the Zen 4 machine was outfitted with 512GB of memory, while its competition had just 128GB each.

[wjfox]

IBM's NorthPole chip runs AI-based image recognition 22 times faster than current chips

October 20, 2023

A large team of computer scientists and engineers at IBM Research has developed a dedicated computer chip that is able to run AI-based image recognition apps 22 times as fast as chips that are currently on the market.

In their paper published in the journal Science, the group describes the ideas that went into developing the chip, how it works and how well it performed when tested. Subramanian Iyer and Vwani Roychowdhury, both at the University of California, Los Angeles, have published a Perspective piece in the same journal issue, giving an in-depth analysis of the work by the team in California.

As AI-powered applications become mainstream tools used by professionals and amateurs alike, scientists continue work to make them better. One way to do that, Iyer and Roychowdhury note, is to move toward an "edge" computer system in which the data is physically closer to the AI applications that are using them.

Commercial applications such as ChatGPT, for example, rely on data accessible through the internet, which introduces time delays. In this new effort, the research team at IBM has developed and built a computer chip that combines the processing module and the data it uses—they named it NorthPole. The team reports that the design of their chip was inspired by the way the human brain works.

The chip uses a two-dimensional array of memory blocks and interconnected CPUs to accomplish its tasks—its all-digital architecture allows the computing cores to communicate directly with far-away blocks as easily as with those that are nearby, a design that allows the chip to process data and return answers quickly.

https://techxplore.com/news/2023-10-ibm ... image.html



Credit: Science (2023). DOI: 10.1126/science.adh1174

[Tadasuke]

AMD Ryzen Threadripper PRO 7995WX offers 12.17 TFLOPs of FP32 and 6.07 TFLOPs of FP64, measured by AIDA64 v6.92.6616. For comparison, Intel Core i9-13900K offers about 2.5 TFLOPs of FP32. 12.17 FP32 TFLOPs is more than PlayStation 5, AMD Radeon 5700 XT or Nvidia GTX 1080 Ti. The processor costs $9,999.



source: link to @9550pro's tweet

[Tadasuke]

14900K often wins by a slight margin over 7950X or 7950X3D, but AMD Zen 4 uses less energy for the same final effect. Because "5nm" TSMC is better than "Intel 7", which is more equivalent to TSMC's "7nm".

[wjfox]

[raklian]

So... with all those obvious advantages of using a refined glass substrate, why didn't we use it in the first place?

[wjfox]

So... with all those obvious advantages of using a refined glass substrate, why didn't we use it in the first place?

My thoughts exactly!

[wjfox]

Qualcomm brings receipts: Snapdragon X Elite gets benchmarked, completely dunks on Apple’s M2 processor

By Daniel Rubino
published about 4 hours ago

- Qualcomm recently announced its Snapdragon X Elite platform built on the Oryon processor with 12 cores, peaking up to 4.3GHz, while using 1/3 the power of Intel’s current best laptop CPU.

- During the Snapdragon Summit, a live benchmarking session on two differently configured reference design laptops demonstrated how powerful the chips are.

[...]

The results are astounding and the real deal.

On PCMark 10, the results are literally off the chart compared to anything we’ve benchmarked during our laptop reviews. Both configurations hit over 13,000 on the score, whereas the ASUS ROG Strix SCAR with the brand-new AMD R9 7945HX3D, the most powerful Windows laptop we’ve tested, could only muster 9,000. Even the MSI Titan GT77, with Intel’s previous best processor, the Core i9-12900HX, could only muster 8,555.

[...] The critical thing to remember during all these benchmarks is that Qualcomm matches or beats the competition (as of today) at all these CPU and GPU tests, but at less power than the others, sometimes up to 70% less power than Apple or Intel.

https://www.windowscentral.com/hardware ... -available

[weatheriscool]

Samsung Provides Early Details of 1.4nm Chip Process
https://www.extremetech.com/computing/s ... ip-process
Samsung plans to increase its lead with GAA technology to make its 1.4nm chips shine.
By Ryan Whitwam October 31, 2023

Today's most advanced microprocessors have components just a few nanometers across, and precious few companies can produce these chips. Taiwan's TSMC owns the lion's share of the market, but Samsung is running in a distant second. The runner-up's next big chip evolution might help it gain some ground, though. Early details of Samsung's upcoming 1.4nm process node suggest future Samsung chips will offer a significant boost to power and efficiency.

Jeong Gi-Tae, vice president of Samsung Foundry, has offered up a general roadmap that could bring the company to parity with TSMC. Currently, Samsung's foundry offers 5nm (also known as SF5) manufacturing for a variety of chips. It can also manufacture 4nm and 3nm chips, but it appears Samsung only offers this technology for low-power embedded systems. Next year, Samsung will spin up a proper SF3 platform, offering customers 3nm chips with various options. However, we might not see many 3nm Samsung chips right away. Google's rumored 2024 Tensor G4 is rumored to be based on Samsung's 4nm (SF4) tech.

[weatheriscool]

Intel's Dec. 14 Meteor Lake Launch Dubbed 'AI Everywhere'
It's time to prepare for the onslaught of AI-based marketing.
By Josh Norem November 1, 2023

In September, Intel officially unveiled its first disaggregated PC architecture, codenamed Meteor Lake. The company stated it would be arriving on Dec. 14, and now Intel has added a name to the event: AI everywhere. We previously knew the company would be leaning heavily on this feature in its pitch to consumers, but now it appears to be the core subject matter of the launch. Intel is not alone on this bandwagon, as Lenovo just held its annual event with the tagline, "AI for all."

https://www.extremetech.com/computing/i ... everywhere

[Tadasuke]

Top Intel Sandy Bridge HEDT i7 in 2011-2012 ($999, about $1350 now) vs the newest 96-core AMD Zen 4 HEDT Threadripper ($9999):



PassMark comparison can be useful for some workloads and users. Obviously not for gaming or other non-professional tasks.

[Tadasuke]

processors using glass substrate later this decade

I guess this might improve performance by 50% at the same power requirements on the same fabrication node (transistor size)? Possibly even more? Pat Gelsinger already talked on stage about using glass.



Here is Intel's Newsroom article about this:
https://www.intel.com/content/www/us/en ... rates.html