GPU and CPU news and discussions

[Tadasuke]

[Tadasuke]

Apple M2 was announced:

[Tadasuke]

M2 has 35% faster iGPU than M1 after 18 months. Assuming +20% iGPU perf increase per year (+100% every 4 years), standard Apple MacBooks will have 40 teraflops iGPUs. Assuming that 20 teraflops is needed for 4K gaming, MacBooks will be capable of that in 2031. All with the same power consumption and price. Battery life will get longer, because of higher capacity batteries, so expect 40 hours battery life in 2032. I imagine that neural engine's performance will improve at a slightly faster rate and CPU's performance at a slightly lower rate. Interesting things are awaiting us. 2032 MacBooks will probably also feature 4K HDR webcams capable of recording in low-light conditions. Screens may also be 4Kish. It's remarkable how long it took for webcams to move from 720p to 1080p, I expected that to happen much sooner.

[Tadasuke]

According to this SemiAnalysis report, "Intel 4" process is going to bring substantially lower density improvements than 2x compared to "Intel 7".

[Tadasuke]

The newer benchmark you use in comparing CPUs or GPUs, the bigger difference you will see between older and newer ones.

For example between i3-10100 and i3-12100 there is 20% difference in Cinebench R15, 40% in Cinebench R20 and 50% in Cinebench R23.

Newer programs are better suited for newer architectures and instructions. But not every game and program is new or modern.

[wjfox]

Japan targeting 2025-7 for 2nm process

15th June 2022

Japanese and US companies are to work together to establish a Japanese capability to manufacture 2nm ICs by 2025-7, reports the Nikkei.

https://www.electronicsweekly.com/news/ ... s-2022-06/

[wjfox]

The First High-Yield, Sub-Penny Plastic Processor

It took a major redesign for cheap flexible chips to reach their promise

14 Jun 2022

For decades, hopeful techies have been promising a world where absolutely every object you encounter—bandages, bottles, bananas—will have some kind of smarts thanks to supercheap programmable plastic processors. If you’ve been wondering why that hasn’t happened yet, it’s that nobody has built working processors that can be made in the billions for less than a penny each.

It hasn’t been for want of trying; in 2021 Arm reproduced its simplest 32-bit microcontroller, the M0, in plastic, but even this couldn’t hope to meet the mark. The problem, according to engineers at the University of Illinois Urbana-Champaign and at British flexible-electronics manufacture PragmatIC Semiconductor, is that even the simplest industry-standard microcontrollers are too complex to make on plastic in bulk.

In research to be presented at the International Symposium on Computer Architecture later this month, the transatlantic team presents a simple yet fully functional plastic processor that could be made at sub-penny prices. The Illinois team designed 4-bit and 8-bit processors specifically to minimize size and maximize the percentage of working integrated circuits produced. Eighty-one percent of the 4-bit version worked, and that’s a good enough yield, says team leader Rakesh Kumar, to breach the one-penny barrier.

“Flexible electronics has been niche for decades,” says Kumar. He adds that this yield study shows “that they may be ready for the mainstream.”

https://spectrum.ieee.org/plastic-microprocessor



Credit: University of Illinois/PragmatIC Semiconductor

[Tadasuke]

Turns out that M1 Ultra, despite all these transistors (the most on the market), is exceptionally good only in video editing and code compilation. In everything else it's worse than 12900K + 3090.

[Tadasuke]

Arm unveiled new mobile CPUs and GPUs. Changes are significant and you can read about them in detail in the linked news articles. In short, the new CPU is 25% faster per core, 20% more efficient and the new GPU is 15% faster per core, 100% faster in machine learning and 300% faster in ray-tracing. The new Mali G715 cores support Variable Rate Shading and are currently the most efficient (I guess until new iPhones come out later this year). First products with next-gen APUs are expected early 2023 (for example Samsung Galaxy S23).




new CPUs: https://community.arm.com/arm-community ... -unleashed
new GPUs: https://community.arm.com/arm-community ... -unleashed

[weatheriscool]

Intel’s Raptor Lake to Support Both DDR4-3200 And DDR5-5600

By Josh Norem on July 6, 2022 at 11:23 am
https://www.extremetech.com/computing/3 ... -ddr5-5600

Intel’s Alder Lake spearheaded the company’s efforts to turn the tide against AMD. It was successful too, earning praise similar to “Intel is back” from many reviewers. Now that Alder Lake is in the history books, the company now turns to its successor, Raptor Lake. It will be an enhanced version of Alder Lake, as Intel is using its famous tick-tock strategy again. Despite the similarities between the two architectures, one lingering question was whether or not Intel would continue to support DDR4 memory. It did so with Alder Lake, but that launched last year when DDR5 memory was almost non-existent. It has now been confirmed that Intel will indeed support both memory types on its upcoming platform. The news comes from a leaked slide Intel showed to sales teams in China.

The slide was posted by a Twitter uses named @9550pro and it lays out in tidy detail what to expect from Intel’s upcoming Raptor Lake-S platform. Support for DDR4-32oo is unchanged from Alder Lake. Still, it wasn’t clear if Intel would drop it in order to get people to adopt DDR5 instead. The fact that it’s keeping it around for one more generation is good news for upgraders on a budget. It will also help Raptor Lake appeal to more people than if it removed DDR4 support. On the DDR5 side of the equation, the news here is it’s upping the frequencies a bit. Alder Lake maxed out at DDR5-4800, but it will supports speeds up to 5600 on Raptor Lake. It also notes there are “additional PCIe 4.0 lanes for the chipset” but it doesn’t say how many.

[Tadasuke]

[weatheriscool]

Raptor Lake to Offer ‘Unlimited Power’ Mode for Those Who Don’t Care About Heat, Electric Bills
https://www.extremetech.com/computing/3 ... er-setting
By Josh Norem on August 11, 2022 at 10:45 am

Intel’s 13th generation Raptor Lake architecture is getting closer, as recent leaks indicate. As each new leak or screenshot appears online, we discover new capabilities Intel has planned for it. The latest is an “unlimited power” feature that lets you crank things up to 11, assuming you have sufficient cooling. On the Core i9-13900K, this setting allows for extreme power consumption to the tune of about 350W.

The news comes from Raichu on Twitter via Wccftech. They are a reliable provider of CPU-related information and seem to have a retail version of Raptor Lake. That is not confirmed, though, and they partially blurred out the description. This time around, they posted screenshots and Cinebench scores for the flagship Raptor Lake CPU. The Core i9-13900K sports eight hyper-threaded performance cores and 16 efficiency cores. It’s noted the CPU offers a dual-core boost clock of 5.8GHz, which is the same as previous leaks. Its all-core boost clocks are 5.5GHz, which is a big step up from the Core i9-12900KS’s single-core boost at that speed. This will be Intel’s first 24-core, 32-thread consumer CPU. Raichu posted screenshots of the CPU running Cinebench. R23. The benchmark was run with both a “power limited” setting and an “unlimited” setting.

[weatheriscool]

Intel May Drop TSMC’s 3nm tGPU From Meteor Lake
https://phys.org/news/2022-08-antarctic ... ought.html
By Josh Norem on August 9, 2022 at 9:25 am

Over the past few weeks, there’s been a flurry of rumors about the imperiled GPU tile Intel needs for its 14th generation CPUs, code-named Meteor Lake. Intel planned for TSMC to make the GPU tile on its 3nm process originally. It would then combine it with three other tiles, two made by Intel, the other one by TSMC, to complete the CPU’s design. However, recent reports stated the 3nm tile might be delayed. It wasn’t clear if TSMC couldn’t deliver it on time, or if Intel’s own tiles wouldn’t be ready. Now a new rumor has surfaced saying Intel is cancelling it altogether. Instead, it will just use TSMC’s current 5nm process.

News of Intel revising its plans follows a report last week from Trendforce about Meteor Lake being delayed. It stated TSMC was planning on pushing its 3nm production schedule back quite a bit. This would have had a volatile ripple effect throughout its fab operations. It said Intel was responsible, but it wasn’t clear which company’s tiles were the culprit. Also, TSMC has granted Intel special status as a customer and devoted a sizable amount of capacity to it as well. If Intel were to say “not right now” to TSMC at this stage, it could cause problems, to put it lightly. TSMC is expected to enter high-volume manufacturing of its 3nm process very soon.

[Tadasuke]

Next generation AMD desktop CPUs (7950X, $699) will be up to 74% faster at 65 watts, compared to Ryzen 5950X ($749 in 2020). This means that with a constant energy envelope, you could do significantly more than previously. It's especially important in mini-PCs, All-in-Ones (AiO) and laptops/ultrabooks. My own Ryzen 5700X uses 65 watts, which is half as much as my previous overclocked i7-2600K, that was 4 times slower and costed slightly more. Zen 4 CPUs support AVX-512 Vector Neural Network Instructions (VNNI) INT-8 calculations for inferencing, designed to accelerate convolutional neural networks.

[Tadasuke]

Zen 1 4-core CCX was 1.4B transistors
Zen 2 8-core CCD is 3.9B transistors
Zen 3 8-core CCD is 4.15B transistors
Zen 4 8-core CCD is 6.57B transistors
CCX = Core Complex
CCD = Core Chiplet Die
Zen 4 has 135% more transistors per 8 cores than Zen 1
between Zen 1 and Zen 4, AMD has done multiple architectural improvements over the last 5.5 years

[Tadasuke]

Ok , thanks to ECSM_Official and OneRaichu, we have now real data on the performance improvements of the upcoming 24-core 32-thread i9-13900K on a Z790 motherboard, compared to the current 12900K on a Z690 motherboard. It is 41.78% faster, but unfortunately uses more energy. I'm most interested in Cinebench R15 (50,82%) and FP32 performance (44.73%). Its Cinebench R15 scores (for comparison) are 331 single and 6253 multi-core. The 13900K is the first Intel 1 teraflops mainstream consumer CPU . The 7950X should be of a similar performance. The 13900K will be over twice as fast as the 10900K from 2020 and around three times as fast as the 9900K from 2018. I wonder if Intel will keep that 3x every 4 years cadence... if you have the 9900K, it may be a good idea to upgrade to the 13900K (depending on your workloads).



And also, soon we will have affordable 10-core CPUs (13400F and 13400).

[wjfox]

[Yuli Ban]

Prices ought to be cratering sooner or later. Even the 4090 ought to be $899 at most.

[Yuli Ban]

[Yuli Ban]