GPU and CPU news and discussions

[wjfox]

Nvidia announces RTX 3090 Ti with faster memory and performance

Jan 4, 2022, 11:35am EST

Nvidia is announcing a new flagship GPU today, the GeForce RTX 3090 Ti. While the wait continues for Nvidia’s Ampere successor, the RTX 3090 Ti is here to prove Nvidia can still squeeze more performance out of its existing 8nm GA102 chip.

Nvidia is only teasing the RTX 3090 Ti today, and it’s the familiar-looking triple-slot design we’ve seen on the RTX 3090. Both cards may look identical, but inside, the new RTX 3090 Ti will include 24GB of GDDR6X running at 21Gbps. That’s the same amount of VRAM as the RTX 3090 but with a nearly 7.7 percent faster memory clock, providing additional performance for 4K gaming and AI tasks.

Nvidia says the RTX 3090 Ti will also include 40 teraflops of GPU performance, around 11 percent faster than the RTX 3090 with 36 teraflops. There’s also 78 teraflops for ray tracing and 320 teraflops for AI tasks. Nvidia hasn’t detailed base and boost clocks, but it certainly looks like the RTX 3090 Ti will be around 10 percent faster than the RTX 3090 on paper.

Just how much this extra performance will impact power requirements isn’t clear yet, either. Rumors suggest the RTX 3090 Ti could require a 1000W PSU, with a TDP of 450 watts. If accurate, that’s 100W more than the RTX 3090 in terms of power draw and a big bump to the 750W recommended PSU.

All of these numbers mean very little without game benchmarks or pricing, though. Nvidia isn’t sharing these details today, so we’ll have to wait on detailed reviews to see how well this card performs against the RTX 3090 and AMD’s Radeon RX 6900 XT. Given the RTX 3090 debuted for $1,499, we wouldn’t be surprised if the RTX 3090 Ti launches around the $2,000 mark.

https://www.theverge.com/2022/1/4/22866 ... e-ces-2022

[Nanotechandmorefuture]

Nvidia announces RTX 3090 Ti with faster memory and performance

Jan 4, 2022, 11:35am EST

Nvidia is announcing a new flagship GPU today, the GeForce RTX 3090 Ti. While the wait continues for Nvidia’s Ampere successor, the RTX 3090 Ti is here to prove Nvidia can still squeeze more performance out of its existing 8nm GA102 chip.

Nvidia is only teasing the RTX 3090 Ti today, and it’s the familiar-looking triple-slot design we’ve seen on the RTX 3090. Both cards may look identical, but inside, the new RTX 3090 Ti will include 24GB of GDDR6X running at 21Gbps. That’s the same amount of VRAM as the RTX 3090 but with a nearly 7.7 percent faster memory clock, providing additional performance for 4K gaming and AI tasks.

Nvidia says the RTX 3090 Ti will also include 40 teraflops of GPU performance, around 11 percent faster than the RTX 3090 with 36 teraflops. There’s also 78 teraflops for ray tracing and 320 teraflops for AI tasks. Nvidia hasn’t detailed base and boost clocks, but it certainly looks like the RTX 3090 Ti will be around 10 percent faster than the RTX 3090 on paper.

Just how much this extra performance will impact power requirements isn’t clear yet, either. Rumors suggest the RTX 3090 Ti could require a 1000W PSU, with a TDP of 450 watts. If accurate, that’s 100W more than the RTX 3090 in terms of power draw and a big bump to the 750W recommended PSU.

All of these numbers mean very little without game benchmarks or pricing, though. Nvidia isn’t sharing these details today, so we’ll have to wait on detailed reviews to see how well this card performs against the RTX 3090 and AMD’s Radeon RX 6900 XT. Given the RTX 3090 debuted for $1,499, we wouldn’t be surprised if the RTX 3090 Ti launches around the $2,000 mark.

https://www.theverge.com/2022/1/4/22866 ... e-ces-2022

Next gen VR is going to be nuts never mind the gaming graphics!

[wjfox]

CES: Intel Announces Core i9-12900KS at 5.5GHz and 22 Other Alder Lake CPUs

By Joel Hruska on January 6, 2022

If you’ve been interested in Intel’s new Alder Lake platform but have been waiting for the company to launch its mainstream parts, we have good news. The chip giant took the lid off its mainstream Alder Lake family at CES this year, from the Core i9-12900 all the way down to some basic Celeron options.

[...]

Separately from its large Alder Lake launch, Intel announced a new, ultra-high-end CPU: The Alder Lake Core i9-12900KS. This 8+8 CPU will feature a 5.5GHz maximum clock speed and an all-core boost of 5.2GHz. The latter is more impressive than the former. Single-core boost clocks are all but meaningless in an age when most applications are at least dual-threaded. An all-core 5.2GHz boost implies that customers should see at least that speed under full load.

https://www.extremetech.com/computing/3 ... -lake-cpus

[Tadasuke]

There is a really interesting article on Next Platform on the past 10 years and the next 5 years of Intel server CPUs. Price/performance got worse between 2012 and 2019 (!!). Sandy Bridge top bin part costed only $2057, while Skylake top bin part costed a whooping $13011 (!!). Cascade Lake in 2020 brought the price down to $3950 (because of AMD EPYC). The dark ages of CPUs seem to be behind us. In front us is Sapphire Rapids, Emerald Rapids, Granite Rapids and Diamond Rapids which will have 144 cores at 2.7 GHz, AVX-1024 instructions, AMX 2 matrix engine, 288 MB of L3 Cache, DDR6, PCI-E 6.0 and will be 10.15x faster than current Ice Lake or 65,56x faster than Sandy Bridge from 2012. Xeon SP v7 on the "Mountain Stream" platform in 2025 will be 167.18x faster than the Nehalem Xeon X5570 server processor from March 2009. That's over 7 doublings in a span of about 16 years. Price will be 938% higher (not adjusted for inflation) than the 45nm 4 core 8 thread Xeon X5570. If not for competition, Diamond Rapids would probably cost around $30,000. IBM and Fujitsu also offer alternatives. There are upcoming ARM server processors with more performance and more cores. IPC improvements in 2024 and 2025 will be greatest since the 1990s for Intel. It is likely that CPUs will reach 1024 cores in the 2030s. AMD is equipping server CPUs with 3D stacked extra cache. It is also likely that L3 cache will be counted in gigabytes in the 2030s. I think it would be better to sum L2 and L3 cache, Intel has similar amounts of L2 and L3. Intel increased L2 cache from 2 to 14 (!) MB between i9-9900K and 12900K. L3 cache nearly doubled. Corecount increased from 8 in 2018 to 16 in 2021 to 24 in 2022 (Raptor Lake i9). Frequency is improving by 100 MHz a year on average. From 5 GHz in 2017 to 5.5 GHz in 2022 and 6 GHz in 2027. I just hope that prices will remain sane and only increase as much as inflation.

https://www.nextplatform.com/2022/01/20 ... -draw-one/

[Tadasuke]

There is a substantial progress in AMD GPU single-core performance in recent years. From 1050 MHz in 2015 to 2815 MHz in 2022. Vega architecture in 2017 added half-precision and RDNA 1.0 in 2019 improved IPC (instructions per clock) by 25%. Higher single-core speed allows for more calculations per pixel or per object. To counter Amdahl's Law, GPUs also need better single-core throughput and not only more compute units. Total cache capacity has also increased exponentially, from 3 MB in 2015 (Fury X) to 136 MB in 2020 (6800 XT). Unfortunately, prices of new graphics cards have been on the rise.

[Tadasuke]

Cost of Intel's 10 cores has been dropping exponentially:

I opted for using Xeons instead of i7-6950X. I think it's likely that i5-13400 will be a 10-core and will cost just that by the end of this year. Later we can expect 10-cores falling into i3 territory.

[Tadasuke]

VRAM Capacity chart (10 years) in lowest-end gaming graphics (dGPUs) cards from Nvidia and their AIB (Add-in boards) manufacturers. Assuming there will be no lower capacity version of RTX 3050. I personally can attest that with less than 8 GB of VRAM there are problems in multiple games, even older ones when you add mods. Some games do require 16 GB to work without problems. More is better.

[wjfox]

Thanks for these excellent graphs, Tadasuke. Really shows the amazing progress.

[Tadasuke]

PCI-Express 4.0 came out in first PC motherboards in 2019 (Zen 2, X570 chipset), PCI-Express 5.0 in 2021 (Alder Lake, Z690 chipset) and PCI-Express 6.0 is expected in 2023 (or 2024) on desktop and 2024 on server. For example new Transcend M.2 2TB SSD achieves up to 7.4 GB/s read speeds trough PCI-E 4.0 x4. Some GPUs have offer slight improvements in performance through PCI-E 4.0 x16 over 3.0 x16 and larger improvements when the number of lanes is lower than 16.

PCI-E 6.0 interface can be useful for servers, AI/ML, networking and storage in data-intensive markets like data center, HPC, industrial, automotive, military and aerospace. Hard to say if this exponential trend will continue uninterrupted after PCI-E 6.0 which will allow for 256 GB/s bandwith over 16 lanes (64 GB/s over 4 lanes which is the common number of lanes in SSDs). PCI-E 6.0 specification has already been officially released by PCI SIG : https://pcisig.com/pci-express-6.0-specification.

[Tadasuke]

Jarrod'sTech made a helpful comparison between different i5 processors. As you can see from this YouTube screenshot, the new i5 is 2x faster than the i5 from 2017. Perhaps we can expect the same 2x improvement from i5 four years later (16600K?). I'd like it to be more than that, but I don't know if that's realistic. Maybe if the upcoming IPC gains are >30%. I'd say that at least 2x and it's not impossible to be 2.5x or 3x.

[Tadasuke]

My predictions for AMD Zen 4 desktop CPUs:
> come out about 2 years after Zen 3
> improve IPC by 40% (quite a lot)
> at least one model will go over 5 GHz on turbo clocks
> feature 1.5x more cores and 4.5x more cache than Zen 3 (because of V-cache)
> no 2 core CPUs anymore
> Athlons will get 4 cores 8 threads (>$100)
> Ryzen 3 will get 6 cores 12 threads ($100-200) - will be the 3rd most popular
> Ryzen 5 will get 8 cores 16 threads ($200-300) - will be the most popular
> Ryzen 7 will get 10 cores 20 threads and 12 cores 24 threads ($300-400) - will be the 2nd most popular
> Ryzen 9 will get 18 cores 36 threads and 24 cores 48 threads ($500-800)
> Threadripper will get 32-96 cores 64-192 threads
> highest Threadripper model will feature over 1 GB of cache
> Zen 4 will fight against Intel Raptor Lake
> I personally may upgrade to Zen 4 and you might consider that too (if prices are reasonable)

[Tadasuke]

This chart shows show much interconnected can layers in a 3D chip be, depending on year and bump density. Currently 3D SoC are only beginning, because interconnect density doesn't even allow for cores on cores or cores on uncore. At best, we will see L3 Cache layered on top of CPUs this year, tripling their L3 Cache size (which can be important for some workloads). In 2022, TSV technology process is at 9000 nm compared to 5nm for transistors. So you see, how far engineering needs to go, to match interconnect size with transistor size (which will be obviously smaller) to link everything together, allowing for true 3D architectures. Doubling every 2 years, it won't be possible even in 2050. However, we can be sure that performance will be improving every decade and 3D chips will be more and more popular and more and more interconnected. Both Intel and AMD are working on that. Fabs are working on that too. It's all official info.

[wjfox]

Intel to blend CPU, GPU cores into monster supercomputing chip

By Joel Khalili published 1 day ago

Intel has teased new Xeon chips that will collect CPU and GPU hardware into one socket to maximize performance across high performance computing (HPC) use cases.

Codenamed Falcon Shores, the new line of processors will combine x86 CPU cores with Xe-HPC GPU cores, with varying ratios depending on the intended workload. These cores will connect up to shared “high-bandwidth memory developed by Intel”.

The company says it expects Falcon Shores processors to arrive some time in 2024, paving the way for zettascale supercomputing systems.

https://www.techradar.com/news/intel-to ... uting-chip

[Tadasuke]

I don't believe in fp64 zettascale in 2027. I do believe in int4 (AI inferencing, neural nets) zettascale in 2027 or 2028.

Intel CPU roadmap:
Q4 2021 - Q1 2022: Alder Lake (laptops are just coming out)
Q3 - Q4 2022: Raptor Lake
Q2 - Q3 2023: Meteor Lake
2024: Arrow Lake
2025: Luna Lake
2026: Nova Lake

AMD CPU roadmap:
Q1 2022 - Zen 3+ (for laptops)
Q3 - Q4 2022 - Zen 4
Q3 2023 - Zen 5
2024 - Zen 6

[Tadasuke]

Intel's CPUs Linpack results in gigaflops (gflops) from 2013 to 2022. 13900K is supposed to be 40% faster than 12900K. Finally a mainstream 1 teraflop CPU. As you can see, performance grew 5x in 9 years since the introduction of AVX-2 instructions in Haswell microprocessors. Vertical axis is logarithmic.


Nvidia and AMD HPC (supercomputer) GPUs from 2007 to 2022. As you can see, performance roughly doubles every 2 years. Unfortunately, prices don't remain the same, but are rising.

Memory bandwith and transistor count are also growing exponentially and predictably in supercomputer accelerator cards. We still don't know the specs of upcoming Nvidia's flagship HPC card.

[Tadasuke]

From what I heard (don't know if that's entirely correct), Amdahl's Law https://en.wikipedia.org/wiki/Amdahl%27s_law for GPU scaling limits the justifiable and commonsensical core count to 131 072. This year, AMD is going to triple (!) their top-end compute units count from 5120 to 15360 (61.44 Tflops), roughly doubling effective performance in games. Diminishing returns can already be felt by looking at scaling from Nvidia Turing to Ampere. If you were to double compute units count from 131072 to 262144, you would probably gain only 10% in framerate, because of the impossibility of 100% parallelizing computer graphics or physics, thus ending the era of large gains.

Frequency of silicon GPUs is probably going to improve exponentially up to about 4 GHz, after which we can perhaps expect maybe +100 MHz a year. Intel Pentium 4 Extreme achieved 3.8 GHz in 2005 and since then, they were able to on average raise the frequency by 100 MHz a year every year to 5.5 GHz in 2022 and probably 6.5 GHz in 2032.

The thing that will continuously add more performance is performance per cycle or per clock or per flops (commonly referred to as IPC). For example: AMD from GCN 5.0 (Vega) to RDNA 1.0 (Navi) added +25%. Intel Alder Lake similarly adds +25% to gaming performance at the same clocks compared to Comet Lake (if you look at 10100F vs 12100F). But it won't be sufficient for doubling of performance every 2 - 4 years. So just like today Alder Lake i7 is only usually 2x faster in games than an old Sandy Bridge i7, GPUs of 2045 might be only 2x faster in games than GPUs of 2035.

Unless of course a revolution happens and we get computers which work in a different way or use completely different materials and photons instead of electrons. Until then, we will have to be content with using 1 petaflops GPUs which won't even be 32x faster in practice than 32 teraflops GPUs.

[Yuli Ban]

It’s a given that Nvidia’s upcoming Lovelace graphics cards will be more powerful than their current Ampere counterparts, but more information is emerging on just how much more powerful they’ll be as we draw closer to the RTX 4000 release date. Now, new leaks suggest the GeForce RTX 4090 will offer a generational performance leap over the company’s current flagship GPU, the RTX 3090.

Sources close to Moore’s Law Is Dead (MLID) are highly confident that RTX 4090 cards built with the full AD102 GPU die will boost fps in videogames using standard rasterization by 80-110%, compared to the RTX 3090. We can also expect ray tracing performance to “at least double”, but MLID is unable to provide any exact metrics at this point in time.

However, in its bid to create the best graphics card in its Lovelace lineup, it appears that Nvidia may push the RTX 4090’s TDP up to an alarmingly high 450-600W. So, don’t be too surprised if you need to upgrade your gaming PC with a pricey power supply if you snag yourself team green’s soon-to-be most powerful pixel pusher.

[Tadasuke]

I have more graphs on computing to share. I create them to better understand how things are and aren't rolling. In 2022 we do have a substantial amount of data to look at, so we can better predict the future than people in the past (even 5 years ago). Super optimists were wrong, but super pessimists were also wrong. There is progress, but not as steep as some have been predicting since the late 80s.

First RAM capacity in iPhones 2010 - 2030 graph. As you can see, RAM capacity goes up by roughly 10x a decade. The optimist in me wants to expect cheaper iPhones to have 32 GB and more expensive iPhones to have 48 or even 64 GB in 2030, but the realist says that it might be lower: 16 GB in iPhone SE and 24 GB in iPhone. iPhone SE 2022 has 4 GB and the upcoming iPhone 14 will have 6 GB.


Then GPU processing speed of Qualcomm Snapdragon SoCs 2010 - 2030 graph. As you can see, flops went up very steeply between 2010 and 2014, but then stagnated up to 2018, since when, they go up in an ~ordinary way. The trendline suggests 100 teraflops in 2030, but I think that 20-25 teraflops (around RTX 3070 Ti) is more likely (if things go fairly well). That will be more than PS5 anyway, but by then there will be PS6 (even PS6 Pro isn't out of the question).


And the last graph is mid-range (XX60 class) Nvidia GeForce graphics cards (GPUs) VRAM capacity 2010 - 2030. As you can see, VRAM went up by 12x in a decade. It is theoretically possible that in 2030, a 60 class Nvidia GPU will have 144 GB of VRAM (but more commonsense, down-to-earth, it will have 40 GB), but it may have an SSD (like Radeon SSG) in addition to VRAM. So it may have less VRAM, but a really fast SSD to make up for it. RDNA 2.0 and 3.0 Radeon cards do have something called Infinity Cache, which is L3 Cache, which supplements VRAM. 6900 XT has 128 MB of that stuff, while 7900 XT will have 512 MB (because of V-Cache or second layer of transistors). So Nvidia might also go that way - giving cards layers of extra cache (and SSD) instead of investing heavily in RAM. The hierarchy would look like this: L1 Cache, L2 Cache, L3 Cache, VRAM, SSD and all of this on one card. Currently, there are efforts to improve interoperability between SSDs and GPUs (Big Accelerator Memory), which essentially enables Nvidia GPUs to fetch data directly from system memory and storage without using the CPU, which makes GPUs more self-sufficient than they are today. Here you can read a paper about GPUs directly connected to SSDs: https://arxiv.org/pdf/2203.04910.pdf
Long story short I expect the following:
RTX 8060 in Q1 2031 offering 40 GB of VRAM, no SSD (but direct connection to PCI-E SSDs), 2 GB of L3 Cache, around 120-130 teraflops and 20x higher ray-tracing performance compared to current 3060. Will see how it goes. I think this prediction is realistic and likely to pass.


And finally, there are rumors that Chinese x86 CPUs could catch up to Intel in 2025. I don't know if that's going to happen that soon, but it will ultimately result in more competition and supply (which hasn't caught up to demand). Russians were trying to create competing arm chips, but they are not going to in current situation.

[Tadasuke]

Transistor count continues to grow exponentially. Nvidia H100 has 2.1x or 110% higher efficiency than A100 from 2 years ago, when measuring PCI-E card performance/watt. PCI-E uses 350 watts and SXM5 version uses 700 watts. H100 SXM5 can output 60 trillion standard floating point operations per second, but its power draw is really high. I used to think that GPUs would never cross 500 watts. Now I think they will cross even 1000 watts. Efficiency gets worse as you clock the card higher. For example MX570 has 4.73 tflops for 25 watts, while 3070 Ti has 21.75 tflops for 290 watts (both are Ampere), which means that MX570 has 2.5x higher efficiency than 3070 Ti (it's 2.5x more environmentally friendly in other words). I wish we would use max efficiency more often, instead of overclocking things, so that their efficiency becomes worse than it could be.

[Yuli Ban]

We'll have to wait for benchmark tests to see whether Intel is right, and the lead might change hands again when AMD releases its much anticipated Zen 4 Ryzen 7000 desktop chips in the second half of 2022. It would be a bit of a pyrrhic victory for end users, however, as the Core i9-12900KS will cost (at least) $739 when it arrives on April 5th — nearly $300 more than the AMD model. That's a pretty high price to pay for bragging rights, particularly when they could be short lived.