GPU and CPU news and discussions

[Tadasuke]

Ice Lake (10th gen) has very good energy efficiency. It also added AVX-512 instructions (useful i.e. in AI workloads). Overall, Meteor Lake seems to be slightly over 2 times faster than Whiskey Lake from 2018 on average. The improvement is certainly noticeable, but not huge. Meteor Lake CPUs are unfortunately more expensive than 8565, 1065 or 1185, but might go down in price later this year. Meteor Lake (Ultra series) drivers seem to be unoptimized.

source: https://www.phoronix.com/review/intel-w ... meteorlake

[Tadasuke]

How has memory (GDDR) speed improved in Nvidia 80-class cards since 23rd of May 2013 GTX 780:



How has memory (RAM) bandwidth improved in Apple iPhones (2010, 2016 and 2022):



2023 iPhones have the same bandwidth as 2022 iPhones. Perhaps, 2024 iPhones will have 68.2 GB/s, like the competition already had in 2023. Improvements aren't great. Especially since 2018 in dGPUs (or more like since 2010 or 2013) and since 2016 in smartphone SoCs. I hope this accelerates soon. GDDR7 is coming later this year and faster LPDDR5 (9.6 GT vs 5.6 GT) is also coming later this year.

I would like to see 65% faster GDDR in RTX 5080 than in RTX 4080 (37 vs 22.4). And I hope that intervals between GPU architectures won't be getting longer. Turing --> Ampere was 24 months and Ampere --> Lovelace was 26 months. Next architecture is allegedly coming in February or March 2025 which would be 26 or 27 months since Lovelace.

[Tadasuke]

After I personally experienced i7-2600K 4 cores 8 threads all running @5GHz back in 2011 (and it wasn't crazy expensive), I was utterly disappointed by all the CPUs which followed. I hoped that they (companies like Intel) would be preparing and working for 5 years to release something revolutionary (in a very positive way) in 2016. Turned out that there wasn't anything interesting on that front in 2016 (both AMD and Intel were practically releasing the same thing in 2011-2016).

Zen 1 wasn't very interesting either. Games and programs couldn't fully utilise 16 threads and single-thread still was only on Haswell level, so it wasn't impressive at all. You could match 4 GHz Zen 1 with 5 GHz Sandy Bridge. 2x corecount was nice, but not very useful in 2017 for most people (more useful when going from an older i3 to 2017 Ryzen 3). Coffee Lake was better, because overclocked (still up to 5 GHz) single-thread performance exceeded 4 GHz Zen 1 as well as 5 GHz Sandy Bridge.

That illusionary hope kept me less depressed and anxious during those 5 years, but more depressed and anxious after I realised nothing much much better would be coming, just more of slightly better things we've already seen before. Some were not surprised. I was honestly very surprised.

[weatheriscool]

Intel Arrow Lake CPU Spotted With 24 Threads
AVX-512 functionality is once again MIA.
By Josh Norem February 2, 2024

Intel is expected to launch Arrow Lake in late 2024 as the next-gen replacement for Raptor Lake. It'll mark a monumental shift for the company from monolithic chips to a new tile-based design. Testing these new CPUs seems to be underway, as an Arrow Lake chip with interesting specs has been spotted online. The chip is reporting it has 24 threads, is running at 3GHz, and has no AVX-512 instruction set (once again).

The chip is labeled "Arrow Lake Client Platform," so no guesswork is required here. Details of the chip were posted on X by @InstLatX64, and the boot log can be found here. It shows a CPU with 24 threads, which sounds unimpressive since Raptor Lake i9 chips have 32 threads. However, the current rumor is that Intel will ditch Hyper-Threading for its next-gen processors, similar to Apple's Arm chips, which could mean it's a 24-core CPU. Intel offers the same number of cores for its current i9 chips, divided between P-and-E cores (16+8).

https://www.extremetech.com/computing/i ... 24-threads

[Tadasuke]

The chip is reporting it has 24 threads, is running at 3GHz, and has no AVX-512 instruction set

This doesn't seem even remotely impressive or exciting, unless there is some crazy high performance per clock going on, which I doubt. Even 40% performance per clock improvement compared to Golden Cove wouldn't change anything, unless Arrow Lake operates at 8+ GHz on air cooling and overclocks to 12+ GHz on liquid nitrogen. But again, always take "leaks" and rumours with a large grain of salt. AMX can replace AVX-512 in AI workloads btw.

[Tadasuke]

In most video games, when Ryzen 1600 came out (11th of April 2017, with a launch price of 219 USD), it was only 33% faster than previous AMD "six-core" - the FX-6300 (release date 23rd of October 2012 for $132). Intel i5-7500 (MSRP $192) was 25% faster. 1600 was usually as fast in most video games as i5-2500K at default clockspeeds released back in January 2011. By the way, Intel i5-2500K could overclock to 5 GHz which was impossible even for Zen 2, let alone Zen 1 (hard to do even for Zen 3 with all cores fully enabled).

In some rare more multi-threaded games, Ryzen 1600 was 57% faster than the FX-6300 and as fast as the i5-4690K. i5-7500 was 12.5% faster.

In the most rare, the most multi-threaded games, it was as fast as i7-6700K, but the i7-6900K was still 34% faster. 1600 was 161% faster than i3-4170 and 149% faster than FX-6300 and those were the extremely rare games, where the first gen 6-core Ryzen was shining and used to be a relatively decent deal for the money.

For example, in Watch Dogs 2 from November 2016, developed by Ubisoft Montreal, Ryzen 1600 @3900 MHz achieved as many frames per second as i7-4790K @4550 MHz, i7-6700K @4220 MHz or i7-6800K @3740 MHz. i5-7600K (came out on the 3rd of January 2017) @5000 MHz was 5% slower while being slightly more expensive and was only 12.5% faster than the six years older i5-2500K @5000 MHz.

In AIDA64 Engineer Edition, Ryzen 1600 @3200 MHz was only 3.2% slower than the 2x more expensive i7-6800K @3400 MHz and 66.1% faster than the new then i5-7600K @3800 MHz. 160.5% faster than AMD FX-6300 @3500 MHz. So it could be a significant upgrade when using the most multi-threaded applications which made full use of it.

[Tadasuke]

8-core 8-thread Skylake [Coffee Lake-R] vs 4-core 8-thread Skylake [Kaby Lake] vs 4-core 8-thread Sandy Bridge (Q1 2011) in 3D particle movement simulation, which can utilize various AVX instructions. Sandy Bridge has only AVX1, while Haswell and all subsequent Intel architectures have also AVX2, which is potentially up to two times faster with floating-point arithmetic. As you can see, the Q4 2018 i7-9700K is about ~3x faster than i7-2600K or i7-2700K (they are the same processors with slightly different names). i7-9700K performance is usually very similar to i7-8700K from Q4 2017.



Here's a new comparison with much newer CPUs from both Intel and AMD:



Turns out that Intel is behind now.

[wjfox]

Um...

This seems more confusing than before.

[weatheriscool]

Intel CEO Confirms Arrow Lake Will Feature TSMC 3nm GPU tiles
Its upcoming Lunar Lake mobile platform will be using these tiles as well.
By Josh Norem February 23, 2024
https://www.extremetech.com/computing/i ... -gpu-tiles

At this week's Intel Foundry event, the company made news by talking about new nodes and partnerships. Though the conference was essentially about how Intel plans to leapfrog its rival TSMC regarding advanced nodes, Intel CEO Pat Gelsinger also confirmed its plans to use TSMC's cutting-edge 3nm process for its upcoming platforms.

Gelsinger's comments about its partnership with TSMC were delivered off-stage in a Q&A with the press and analysts, according to Wccftech. In the scrum, Gelsinger noted the two companies have been partners for a long time and that Intel will continue to use the Taiwanese fab for its next-generation Arrow Lake and Lunar Lake platforms. Intel currently uses TSMC's 6nm process for its discrete Arc graphics cards and TSMC's 5nm for the GPU tiles on its Meteor Lake mobile platform. Gelsinger said the company will jump to TSMC's 3nm process for its next-generation products.

[Tadasuke]

*compared with an unspecified Intel 2021 server platform



This could be ~approximately in line with my "usually 2x every 4 years at the same price" prediction. I still don't know Sierra Forest CPUs prices.

[weatheriscool]

Intel Adds New 1nm Node to Its Foundry Roadmap for 2027
It also says AI 'Cobots,' or collaborative robots, will work alongside humans in future foundries.
By Josh Norem February 28, 2024
https://www.extremetech.com/computing/i ... p-for-2027

Intel held its first big foundry event for media and analysts last week, and the headline was adding Intel 14A to its foundry roadmap. That's essentially the company's 1.4nm process, though it didn't announce when it would arrive. Now that the show is over, it's modifying that announcement by giving it a launch date in 2026 and adding one more node just beyond it by stating that 1nm, or Intel 10A, will come right after in 2027.

Intel gave a briefing to the media about Intel 10A at the show, but due to some confusion about when the embargo over the news would lift, it hasn't come to light until now. Tom's Hardware has posted about the presentation, which shows Intel will continue to pursue a very aggressive node progression strategy even after its "five nodes in four years" campaign theoretically ends in 2025 when Intel 18A arrives. Assuming 1.8nm launches next year, the company will follow it up with 1.4nm in 2026, then leap to 1nm in 2027. If this transpires, the company will likely be ahead of its rival TSMC, which is expected to be on 2nm by 2025 or 2026, with 1.4nm following after that.

[Tadasuke]

Intel will continue to pursue a very aggressive node progression strategy even after its "five nodes in four years" campaign theoretically ends in 2025 when Intel 18A arrives. Assuming 1.8nm launches next year, the company will follow it up with 1.4nm in 2026, then leap to 1nm in 2027.

From what read yesterday, it seems, that Intel just arbitrary sets a "new node", when it's between 14 and 15% better than the previous node. "Intel 4" is supposedly "~15% better" than "Intel 7". Whatever that means. I find those names rather illusory and deceptive. Wouldn't treat them too seriously.

If 10A is 14.5% better than 14A, which is supposedly 14.5% better than 18A, which is 14.5% better (?) than 20A, which is 14.5% (?) better than "Intel 3", which is 14.5% (?) better than "Intel 7" (the legacy name was "10nm") <-- that is being used in current products, like Intel "Raptor Lake" and "Emerald Rapids" ("Xeon Scalable 5th Gen"), then Intel 10A (1nm) would possibly be 96.8% "better" overall, than current "Intel 7" (or "10nm"). Perhaps in 2027, 2028 or 2029, Intel products could be ... about 2 times better I guess? 2x would be nicer than what they offer now...

i3-12100 (Golden Cove cores) at 3.9 GHz probably uses about half the energy of i7-7700K (January 2017) at 5.0 GHz, with very similar performance in most tasks. And PlayStation 5 "Slim" on "TSMC 6nm" uses ~180 watts, while having somewhere between 5 and 7x performance of ~120 watts "28nm" PlayStation 4 (and being significantly less noisy).

Sheer performance is not enough. More concerning are problems like Windows backup still being difficult to set up, cloud storages for some unknow reason deleting your files permanently (in tens of GBs), search engines not getting any better or bots prowling the web and spamming.

[weatheriscool]

Snapdragon X Elite Dethrones Intel Meteor Lake in Pre-Launch Tests
Hopes are high for the first really powerful Arm-based processor for Windows.
By Josh Norem March 13, 2024

Intel announced its upcoming Snapdragon X Elite SoC in late 2023, and it's promising to shake up the Windows processor landscape in 2024 by being a serious competitor to mobile chips from both AMD and Intel. The Qualcomm SoC will be targeting the ultra-thin laptop market when it debuts later this year, and new pre-launch benchmarks show it can already beat Intel's latest Meteor Lake CPUs in some tests.

A few weeks ago at Mobile World Congress in Barcelona, a YouTuber ran some benchmarks at a Qualcomm event, but only now uploaded a video showing the company's newest SoC going head-to-head with Intel's latest processor. The chips involved are a 12-core Snapdragon X Elite with the bizarre name of X1E80100, and an Intel Core Ultra 7 155H, which has 16 cores (6+8+2). Both chips are rated as having a 28W TDP, and though the Qualcomm laptop is shown as having 64GB of memory, it's assumed the Windows machine has a similar configuration.

https://www.extremetech.com/computing/s ... unch-tests

[weatheriscool]

Intel Announces 6.2GHz Core i9-14900KS 'Special Edition' CPU
Intel ups the ante once again in the clock speed wars.
By Josh Norem March 14, 2024

https://www.extremetech.com/gaming/inte ... dition-cpu

Today, Intel is officially unveiling the flagship CPU for its 14th Generation chips—the Core i9-14900KS Special Edition. This is the final CPU to arrive for the 14th Generation family, aka Raptor Lake Refresh. Like the rest of the chips in the family, it ups the single-core boost clock by 200MHz over its predecessor, so as expected it can reach 6.2GHz right out of the box. It's the highest-clocked CPU ever made and will cost around $699 when it goes on sale today.

The specs for the 14900KS are almost exactly the same as its predecessor, the Core i9-14900K, but with a 6.2GHz single-core boost clock instead of 6GHz. It's still a 24-core, 32-thread CPU on the LGA 1700 socket and is probably the final chip on that platform before Intel moves to LGA 1851 later this year. The E-core Turbo Max clock has also been nudged upwards by 100MHz to 4.5GHz. All this CPU juicing has raised the PL1 power rating from 125W in the previous chip to 150W for the KS version. The PL2 rating, which is maximum turbo power, is unchanged at 253W, but as we've seen already, that number can get up to 400W or so.

[weatheriscool]

Cerebras Unveils CS-3 Wafer-Scale AI Chip With 900,000 Cores and 4 Trillion Transistors
The company takes an entire TSMC wafer and sells it as one huge chip instead of cutting it into individual dies.
By Josh Norem March 14, 2024
https://www.extremetech.com/computing/c ... 4-trillion

If you thought a chip like AMD's MI300A was big at 146 billion transistors, you ain't seen nothing yet. AI company Cerebras announced its third-generation AI chip, CS-3, a "wafer-scale" silicon monstrosity designed for AI training. It's basically an entire TSMC 5nm wafer sold as a single chip. The CS-3 is 56 times the size of Nvidia'a H100 and features over 4 trillion transistors, making it the biggest "chip" in the world by a wide margin. The company says it is the world's fastest AI chip, breaking the record set by its predecessor, the CS-2.

[weatheriscool]

Intel Core i9-14900KS Is Already Breaking Records, Overclocked to 9.1GHz on Liquid Helium
As expected, Intel's highest-clocked chip ever is already setting world records even though it's just one day old.
By Josh Norem March 15, 2024

Intel launched the swan song for its LGA 1700 platform on Thursday with the Core i9-14900KS. This "Special Edition" CPU is a binned version of the 14900K, plucked from obscurity due to its ability to overclock to 6.2GHz right out of the box. Now, an overclocking team has taken Intel's newest flagship CPU past 9.1GHz and set a new world record. The team had to use liquid helium to accomplish this task, which is slightly more effective than liquid nitrogen and represents the pinnacle of cooling technology.

An Asus overclocking team led by the esteemed clock-tickler Elmor is responsible for the new CPU world record at an Asus Republic of Gamers overclocking event. Elmor and his crew were able to push the Core i9-14900KS to 9117.75 MHz on an Asus ROG Maximus Z790 Apex motherboard, which costs around $650. To achieve this clock frequency, the CPU had to be chilled to sub-zero temperatures with liquid helium, and the video shows the elaborate setup required to get it to work correctly. The chip ran at a frosty -230C for the final run at 9.117GHz.

https://www.extremetech.com/computing/i ... d-to-91ghz

[Tadasuke]

9.1 doesn't seem like something significantly different than what we have already seen so far (seems about the same). Upcoming Intel architectures might achieve lower clock by the way, not higher, according to some people. I guess by 2030 overclockers will get to 10 or maybe even 11 GHz.

Here's what PS5 Pro with its newer 4nm SoC will be capable of, according to multiple leaks which seem to verify each other:
• larger GPU with faster memory, providing 45% improved performance in rasterized rendering
• massively improved ray tracing architecture that should deliver 2-3x speedup over the regular PS5 with peaks of 4x
• custom machine learning architecture, that supports 300 TOPS (Trillions of Operations Per Second) at 8-bit precision
• the ML architecture was built to enable the so-called PlayStation Spectral Super Resolution (PSSR) upscaling technique, the leaked document describes it as Sony's version of Multi Frame Super Resolution based on the PlayStation Machine Learning (PSML) algorithm, according to the leaked info PSSR currently supports up to 4K, but there are plans to add 8K support
• PSSR is an ML-enhanced version of Temporal Anti-aliasing Upscaling (TAAU) that requires similar inputs to NVIDIA DLSS or AMD FSR and fully supports High Dynamic Range (HDR) pipelines and no per-game training is required just like with the latest version of DLSS
• PSSR requires just 250MB of the PS5 Pro's memory, with a current rendering cost of 2 milliseconds to upscale from Full HD to 4K, although optimization is still ongoing
• the document also includes an image comparison with TAAU and AMD FSR 2, reportedly showing that PSSR offers superior image quality

[weatheriscool]

[Tadasuke]

Here's what our IT professor told us at university many years ago:
• AI will never really work, it will never really understand human language or the world, it might be useful in faking understanding or in some special cases using expert systems
• CPUs are the driving force of computing and software
• depending on the use case, somewhere between 4 and 64 cores might be the useful practical limit, multi-threading won't be useful beyond 2 threads per core
• CPUs might get to 1 THz sometime in the 2030s and they might get to 1 PHz by around 2100, but they will never get past that, as laws of physics won't allow for higher clockspeeds
• future CPUs might contain a few gigabytes of integrated fast cache as that is very possible
• C++ and Perl will be around and used for decades to come
• single board computers will be much more popular in the future
• GPUs are for video games, pre-rendering graphics and might be useful for some simulations
• quantum computers will never work
• open-source Linux is the future, also for mobile
• Apple company is a joke and Microsoft is evil
• Dennis Ritchie >> Steve Jobs

Seems like he wasn't 100% correct.

[erowind]

.