GPU and CPU news and discussions

[Yuli Ban]

Nvidia is launching its $1,999 GeForce RTX 3090 Ti today. After a mysterious delay following a January announcement, the “fastest GPU on the planet” will be available from Asus, Colorful, EVGA, Gainward, Galax, Gigabyte, INNO3D, MSI, Palit, PNY and Zotac in stores today. Nvidia has also created a limited edition Founders Edition board that will be available from Best Buy online.
The RTX 3090 Ti is the familiar-looking triple-slot design we’ve seen on the RTX 3090, and both cards look identical at first glance. Nvidia is really maximizing what its existing 8nm GA102 chip is capable of inside the RTX 3090 Ti, coupling it with 24GB of GDDR6X running at 21GB/s. That’s the same amount of VRAM as the RTX 3090 but with a nearly 7.7 percent faster memory clock resulting in a total memory bandwidth of 1008GB/s.
Nvidia says the RTX 3090 Ti will also include 40 teraflops of GPU performance, with a base clock of 1560MHz, and a boost clock of 1860MHz. That’s a bump over the 36 teraflops on the RTX 3090, the base 1395MHz clock, and 1695MHz boost clock speeds. The RTX 3090 TI also includes 10,752 CUDA cores, 78 RT-TFLOPs, and 320 Tensor-TFLOPs.

40 teraflops, Jesus.
It's not that much of an improvement over the RTX 3090, but still...


Sweeney: 40 TFLOPS Can Render Photo-Realistic Dynamic Scenes, But Humans Require More Than Computing Power

[Yuli Ban]

[Yuli Ban]

GeForce RTX 40 series board could smash 100 TFLOPS barrier as leaker doubles down on Lovelace AD102 GPU capability

Leaker Greymon55 has doubled down on an old claim of how capable the high-end AD102-powered RTX 40 series boards could be. A leak from the end of 2020 suggested cards with a Lovelace AD102 GPU (potentially RTX 4080, RTX 4080 Ti, RTX 4090, and RTX 4090 Ti) could churn out 66.4 TFLOPS in FP32 (single-precision floating-point format) at 1.8 GHz. A later claim, made by Greymon55, stated a board with an AD102 GPU could offer 85 to 92 TFLOPS, with clock rates from 2.3 GHz to 2.5 GHz. Now, the same leaker has posted a reminder of this astonishing graphical compute performance, saying “if Lovelace gets up to 2.5 GHz, you’ll get about 90t of FP32”.

Although TFLOPS (tera (trillion) floating operations per second) rate is far from the only way of measuring a GPU’s compute performance and general handling of graphics output, it has become more accessible as a form of quick comparison tool for the wider public, as seen in the recent competition between the PlayStation 5 (10.28 TFLOPS) and Xbox Series X (12.15 TFLOPS) current-gen consoles. Nvidia’s RTX 2080 Ti took the PC gaming world to 13.45 TFLOPS and then Ampere came along and basically tripled that amount with the RTX 3090 delivering 35.58 TFLOPS in FP32 (single-precision floating-point format) and the upcoming RTX 3090 Ti supposedly managing 40 shader TFLOPS.

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

Early last year, graphics card makers like Asus announced they would be raising prices amid an intensifying GPU shortage. This price hike was partly a result of tariffs imposed on some Chinese products by the Trump administration; the tariffs had been waived for several years but were allowed to go into effect at the end of 2020. Last week, the Biden administration reinstated many of those waivers, and Asus announced Monday that the company would lower its GPU prices as a result.

The price drops will go into effect on April 1 and will apply to the company's entire lineup of GeForce RTX 3000-series GPUs, including the 3050, 3060, 3070, 3080, and 3090 series cards. "Consumers should expect prices to decline up to 25% on different models throughout the springtime," Asus said in a statement.

Asus and other GPU makers have revived (or simply continued selling) many older GPU models—including the RTX 2060 and 1660 and the GTX 1050 Ti—to try to address the ongoing GPU shortage. There's no word on whether we can also expect to see price reductions for these models or if they'll eventually go away as the availability of newer GPUs improves.

[Tadasuke]

I posted about HPC cards, now it's time for desktop cards. GTX 680 costed $500 in 2012 and RTX 4090 will cost $2000 in 2022. $500 from 2012 is $620 in 2022. Even with this large price increase, performance grows with higher acceleration than price. We can safely assume over 1 petaflops in 2032, but for an even higher price. Imagine owning a PC that would be #2 fastest supercomputer in 2008. This will become reality during the 2030s.

[Tadasuke]

12900KS is 4% faster in apps and 2% faster in games than 12900K
it's also more expensive, more power hungry and hotter, so needs better (more expensive) cooling

[Tadasuke]

Today, I quickly made this graph, to answer the important question, of how have Intel desktop CPUs evolved in multi-threaded performance, since the famous i7-2600K. I took Cinebench R15 scores from HWBot.org, so that maximum overclocks are compared. Comparing max OC makes more sense, because 2600K could be OCed by more % than 12700K. Boost clocks have been getting closer and closer to max OC since the i7-4790K. Actually, 12700K overclocks to a lower point than 3770K, which is a shame. As you can see, things are starting to accelerate more so than before. 13700K will double the amount of E cores, while improving performance per clock. I will update this graph once results from overclocking 13700K become available.

EDIT1: I hope this doesn't continue like that, because 4x in 10 years is ... really bad! 4x should take 3 years. If it took 5 years, then I would be still ok with it, but, seriously, 10 long years? It has to change. It just has to. The graph was improved and re-uploaded.

[Tadasuke]

Ok, so today I made probably a more important graph - how has multi-thread performance of Intel server CPUs changed over the years and how it will change until 2026. Xeon E5540 in 2009 costed only $774, E5740 in 2010 costed $1,980 and E7-8870 costed $4,616. As you can see, price increased by 6x in just 2 years! That's not the end however, because Xeon Platinum 8180M in 2017 was priced at whoppingly 13,011!! That's almost 17x more than 8 years earlier!! However, thankfully to increased competition, it came down to $3,950 in Q1 2020, while performance improved by 8%. Sapphire Rapids will cost about $10,000 and Diamond Rapids once again $13,000 (but it will have 144 cores). So performance/price (without taking inflation into the calculation) will improve by ~10x between 2009 and 2025, which is not a lot for 16 years in my opinion. YoY performance growth between 2020 and 2025 will be substantially better than between for example 2015 and 2020. Expect around 10x between Ice Lake (2021) and Diamond Rapids (2025). In fact, it will be remarkably good in pure performance metrics. 10x in 4 years could after all mean 100x in 8 years, but we'll see.

[Tadasuke]

Yesterday, Intel expanded its Fab in Oregon, is going to build a fab in New Albany in Ohio in the USA (3.5 football fields of clean room area) to start operation in 2025, also is going to build 17 billion € Intel Halbleiterfabrik in Magdeburg in Germany (to be opened in 2027, 7000 construction jobs and 3000 permanent jobs + suppliers, 2 football fields of clean room area) and in France plans to build its new European research hub, creating 1000 new high-tech jobs.

Intel wants to do 5 process nodes in 4 years (!). The U.S. share of chip-making business has dropped from 37% in 1990 to 12% in early 2022, according to the Semiconductor Industry Association. Taiwanese TSMC plans to spend 40-44 billion $ on new fabs in 2022 alone. So a lot is happening, but remember that building fabs is a multi-year giant undertaking. Current supply issues are going to be gradually resolved, unless demand grows faster. GPU, CPU and DDR5 prices are slowly dropping.

[Tadasuke]

Ryzen 5800X3D reviews are out. It looks to be 20% faster in games than 5800X, but can't be overclocked and core voltage is limited to 1.35v. L2+L3 cache add up to 100 MB. $800 12900KS uses much more energy and is barely faster in games (also motherboards for it and DDR5 RAM are more expensive). Exact results vary depending on reviewers. More cache doesn't improve productivity according to Linus Tech Tips. 5800X3D is an upgrade for people on AM4 platform. This year, AM5 platform will be released along with Zen 4. Upgrading $400 1700X to $450 5800X3D will mean over double the framerate in video games after 5 years. For most people, $300 5700X will be sufficient.

[weatheriscool]

Intel’s Next-Gen Raptor Lake Flagship CPU Rumored to Run at 5.8GHz

By Josh Norem on April 15, 2022 at 10:52 am
https://www.extremetech.com/computing/3 ... -at-5-8ghz

Now that Intel’s Alder Lake CPU lineup is complete with the addition of the Core i9-12900KS, it’s time to look to the future. Its upcoming CPUs will be called Raptor Lake, and they’ll be a refinement of Alder Lake made on the same Intel 7 process. It’s similar to how Intel used to follow a Tick-Tock strategy, way back when. Although the 12900KS is the highest clocked CPU available currently at 5.5GHz for two cores, rumors indicate Intel will be going even further with Raptor Lake.

According to a tweet from CPU rumor monger OneRaichu flagged by Wccftech, Intel will push towards 6GHz with its next-gen. Though it already holds a large clock advantage over AMD, Intel may attempt to extend its lead and push Raptor Lake has high as 5.8GHz.

That would apply to the flagship CPU only though, the Core i9-13900K (or maybe even the KS version). For this chip Intel will replace Alder Lake’s performance cores (P-cores), dubbed Golden Cove, with updated cores named Raptor Cove. In a recent demo of Raptor Lake, it showed it running eight performance cores (hyper-threaded), and 16 efficiency cores. This makes it a 24-core, 32-thread CPU. It’s expected that the efficiency cores (E-Cores) will remain unchanged as far as clock speeds are concerned.

[Tadasuke]

10 GT/s or 10 GHz system memory (DDR5) is now a fact:
https://twitter.com/msigaming/status/15 ... 7510465536
Of course it's overclocked memory for now, not default clocks from the store, but that's still an achievement.

[Tadasuke]

Mid-range graphics cards prices have risen from $199 in 2016 (RX 480) to $549 in 2022 (RX 6750 XT) or 175% in MSRP. $199 GPU today is as low-end as $79 GPU in 2016.

[Tadasuke]

Intel 4 will bring 20% improvement in performance per watt through EUV technology compared to Intel 7 production process. That's not a lot, but still slightly significant. It could theoretically mean 12 hours of battery life instead of 10 or 24 instead of 20, but in practice it will be less than that, because a computer is not only a CPU or a GPU.

Future Zen CPUs will receive AI inference FPGAs, thanks to Xilinx.

[weatheriscool]

AMD Releases New Details on Zen 4-Powered Ryzen 7000 Family, Upcoming AM5 Chipsets

By Joel Hruska on May 23, 2022 at 2:30 am
https://www.extremetech.com/extreme/336 ... 5-chipsets

It’s Computex this week, which means companies like AMD, Nvidia, and Intel will be demoing some of the systems and improvements they’ve got planned for the year. AMD’s Lisa Su is giving the keynote at Computex again this year, and the company is kicking things off by giving us a bit more detail on Zen 4 and its upcoming platform, Socket AM5.

First, to quickly recap what we know: Zen 4 is built on a 5nm process. It will arrive later this year along with Socket AM5. As befits the name, AM5 will introduce AMD’s support for PCI Express 5.0 and DDR5. Both of these technologies debuted on Intel’s Alder Lake last year, though compatible hardware has been nonexistent. That’s not unusual — PCIe 4.0 drives were in short supply for months after AMD debuted support for that platform — but it means both companies should have added support by the time that PCIe 5.0 hardware is practically in market.

[Tadasuke]

Zen 4 CPU Blender rendering seems to be 45% faster (31% shorter time) than Alder Lake, according to AMD Computex Keynote. We don't know the clockspeeds, but that looks to be substantial. Hopefully the CPUs come out in September, along with Raptor Lake. Progress in the 2020s might be better than in the 2010s.

[Tadasuke]

[Tadasuke]

Datacenter becomes Nvidia's largest business:

but their gaming hardware revenue is also growing.