GPU and CPU news and discussions

[Tadasuke]

Elsewhere in this forum I wrote about computers or processors possibly evolving in performance approximately 10x a decade. Here's a comparison in PassMark of the top Intel Xeon in 2013 vs now the top AMD Threadripper in 2023. These are the most performant CPUs of 2013 and 2023.

[wjfox]

IBM Begins Posting "PowerPC Future" Compiler Patches For What Is Likely Going To Be POWER11

19 October 2023

Just as IBM was posting "future" processor compiler patches in 2019 for what ended up being early POWER10 enablement, they are once again repeating their same compiler enablement technique with sending out "PowerPC future" patches for what is likely to be POWER11.

The "PowerPC future" patches sent out today are just like before -- complete with mentions like "This feature may or may not be present in any specific future PowerPC processor...Again, these are preliminary patches for a potential future machine. Things will likely change in terms of implementation and usage over time."

https://www.phoronix.com/news/GCC-Power ... re-POWER11


-----


So I guess we're looking at 2025 for POWER11.

Graph I made a few years back –

[wjfox]

NVIDIA announces H200 Tensor Core GPU

14th November 2023

The world's most valuable chip maker has announced a next-generation processor for AI and high-performance computing workloads, due for launch in mid-2024. A new exascale supercomputer, designed specifically for large AI models, is also planned.

Read more: https://www.futuretimeline.net/blog/202 ... re-gpu.htm

[weatheriscool]

AMD 96-Core Threadripper Sets Multiple World Records Using 1,000W of Power
And they haven't even gotten around to using LN2 yet.
By Josh Norem November 17, 2023

AMD officially launched Zen 4 Threadripper a few weeks ago, and now the company is flexing its silicon muscle by seeing how it overclocks on air and liquid. For the past two days, the company has let one of its own overclockers tinker with a 96-core Threadripper CPU in its testing lab to see how far they can push the chip. Using a custom cooling loop, they set multiple world records for Cinebench and GPUPI while pulling down between 800W and 1,000W each run.

The overclocker goes by the online handle SAMPSON and is listed as Bill Alverson on HWBot and LinkedIn. Their job title is listed as "AMD overclocking/IO performance optimization," which is pretty nice work if you can get it. Sampson began the experiment yesterday by running a gigantic air cooler on the Threadripper Pro 7995WX, which features 96 cores and 192 threads. It sports a 5.1GHz single-core boost clock along with a 2.5GHz base clock. Using a 3.5-pound air cooler with four 120mm fans, they wrestled the chip up to 4.8GHz on all 96 cores, which set some new performance records. However, that was just an appetizer for the next day's tests using a custom liquid cooling loop, which anyone can put together if they're inclined.

https://www.extremetech.com/computing/a ... w-of-power

[weatheriscool]

AMD Reportedly Taps Samsung Foundry for 4nm Zen 5c Designs
The company is diversifying its chip sourcing, according to a new report.
By Josh Norem November 17, 2023

AMD has traditionally used TSMC to craft all of its most advanced silicon designs, but now that it's working on its next-gen Zen 5 architecture, it looks to be switching things up a bit. A new report from Taiwan states AMD has enlisted Samsung Foundry to produce Zen 5c products for its next-generation platform, which marks a huge shift for the company. It will still use TSMC as well, but it appears to want to diversify its sourcing for Zen 5 instead of relying solely on TSMC as it has done with Zen 4 and the previous generations.

The new report on AMD's future plans comes from a paywalled article from DigiTimes. This Taiwanese news source reportedly cites a South Korean business site as its source, according to Tom's Hardware. The report states AMD will use both TSMC and Samsung to manufacture next-generation Zen 5c products, which are code-named Prometheus. It'll utilize Samsung's 4nm process for "basic" versions of its future chips while allowing TSMC to make more advanced variations on

its cutting-edge 3nm node.

https://www.extremetech.com/computing/a ... 5c-designs

[Tadasuke]

From what I recently checked, between 1996 and 2001, in the $200-$350 range, single-threaded CPU performance (often the most important kind of performance) improved by 15x, and that was probably the time of the largest techno-optimism in the spheres of computers and the internets (also connected to and resulting in the DotCom Bubble).

Pentium 4 1.8 GHz in 2001 used 66.1 watts, compared to 11.2 watts of Pentium 133 MHz. So the power draw for one core increased by as much as 5.9x in just 5 years. Therefore, Pentium 4 was 2.54x more efficient than Pentium 1, after 5 years. Because of impressive speed gains, new chips required much more energy.

Core 2 Duo in 2006 used 65 watts for 2 cores, so 32.5 watts for one core and each core was 3x faster than Pentium 4 1.8 GHz (single thread CPU). Therefore, Core 2 Duo E6600 was 6.1x more efficient than Pentium 4, after 5 years. Very substantial improvement in efficiency.

Sandy Bridge in 2011 used 25 watts for one core, that was 2x faster than Core 2 Duo single core from 2006. Therefore, Sandy Bridge cores were 2.6x more efficient than Core 2 Duo, after 5 years.

Alder Lake in 2021 used 20 watts for one core, that was 2x faster than Sandy Bridge single core from 2011. Therefore, Alder Lake cores were 2.5x more efficient than Sandy Bridge, after 10 years. In 2024, same single core performance will be achievable using 10-15 watts for x86.

[Tadasuke]

AWS Unveils Next Generation AWS-Designed Graviton4 and Trainium2 Chips

Graviton4 provides up to 30% better compute performance, 50% more cores, and 75% more memory bandwidth than current generation Graviton3 processors, delivering the best price performance and energy efficiency for a broad range of workloads running on Amazon EC2.

Trainium2 is designed to deliver up to 4x faster training than first generation Trainium chips and will be able to be deployed in EC2 UltraClusters of up to 100,000 chips, making it possible to train foundation models (FMs) and large language models (LLMs) in a fraction of the time, while improving energy efficiency up to 2x.

More than 10,000 organizations worldwide—including Comcast, Condé Nast, and over 50% of the Fortune 500—rely on Databricks to unify their data, analytics, and AI. "Thousands of customers have implemented Databricks on AWS, giving them the ability to use MosaicML to pre-train, finetune, and serve FMs for a variety of use cases," said Naveen Rao, vice president of Generative AI at Databricks. "AWS Trainium gives us the scale and high performance needed to train our Mosaic MPT models, and at a low cost. As we train our next generation Mosaic MPT models, Trainium2 will make it possible to build models even faster, allowing us to provide our customers unprecedented scale and performance so they can bring their own generative AI applications to market more rapidly."

https://www.techpowerup.com/316175/aws- ... ium2-chips

[Tadasuke]

[Tadasuke]

Intel's own evaluation of their deep learning inference performance:



Intel Xeon Emerald Rapids table with some specs and prices:

[weatheriscool]

Intel VP Says the Company Is Ready to Begin 2nm Production in 2024
The company claims that in 2024, it will once again regain the lead in the semiconductor race.
By Josh Norem December 15, 2023
https://www.extremetech.com/computing/i ... on-in-2024

Intel has lagged behind its main competitor, TSMC, for many years now when it comes to progress on advanced nodes. To remedy this situation, CEO Pat Gelsinger set in motion an aggressive comeback strategy when he returned to the company in 2021. Its IDM 2.0 strategy involves moving through five nodes in four years, which will theoretically allow the company to leapfrog all of its competitors and land at the apex of semiconductor manufacturing once again. Now, it appears the company might just be on track to deliver on that promise.

In a recent interview, one of its VPs stated the company is set to begin production on its "2nm" node next year, which it calls Intel 20A. If it pulls this off—always uncertain in the semiconductor business—it would put Intel years ahead of TSMC in the race to 2nm. Sanjay Natarajan, the company's chief vice president of technology development at Intel, commented in a paywalled article flagged by Wccftech. Natarajan stated, "We will enter mass production [of its 2nm process] in 2024, and Intel will lead miniaturization again." It's a bold claim from the company but also means it's still adhering to its roadmap.