#GPU_architecture

[phonemantra-blog]

So far only Radeon RX 460 You can connect a full desktop discrete graphics card   to a Raspberry Pi 5 single board computer for $60. [caption id="attachment_85125" align="aligncenter" width="780"] GeForce RTX 4090 to Raspberry Pi 5[/caption] This is exactly what channel author Jeff Geerling implemented by connecting a Radeon RX 460 adapter to his Pi 5. And it all worked. True, achieving this is not so easy. Of course, the Raspberry Pi 5 itself does not have a PCIe slot, and indeed there are no classic slots at all. But since this is a Raspberry Pi, its capabilities can be expanded with various modules, adapters, and other things. That is, you can even connect a GeForce RTX 4090 to Raspberry Pi 5? To connect the video card, the author of the project used the uPCity module, which is not yet commercially available, but will appear in time. In addition, of course, the card needed to implement additional power. https://youtu.be/BLg-1w2QayU It’s difficult to say whether someone needs such a solution for practical purposes, since the Raspberry Pi 5 was created for experiments, but the ability to connect a discrete video card in itself is already a big plus.

[phonemantra-blog]

Jensen Huang expressed commitment to the Chinese market Nvidia CEO Jensen Huang believes that the United States is still far from independence in the semiconductor manufacturing market. [caption id="attachment_85069" align="aligncenter" width="780"] Nvidia[/caption] According to Huang, it will be another 10 to 20 years before the United States stops being dependent on Taiwanese and Chinese chip makers. Of course, this is provided that all these years the country’s course will be aimed at getting rid of this very dependence. The head of Nvidia thinks so The United States has been working for quite some time to open chip production factories within the country, but it cannot be said that very many of these factories have already been built there. Huang also noted that, despite US sanctions, Nvidia will continue to work with the Chinese market, as it is simply the largest. Let us remind you that yesterday it became known that instead of the banned RTX 4090, the company will create GeForce RTX 4090 D (Dragon) especially for China .

[phonemantra-blog]

Many cards are created from used adapters In China, a full-fledged hunt has begun for video cards that can at least somehow be turned into accelerators for AI. We have already talked about redesigning the RTX 4090 and RTX 3090 , but the Chinese went further and began to add memory to even less powerful adapters.  [caption id="attachment_84870" align="aligncenter" width="780"] RTX 3080 20GB and Radeon RX 580[/caption] The companies involved began creating versions of 3D maps that had never been released. For example, GeForce RTX 3080 with 20 GB of memory and even Radeon RX 580 with 16 GB. The second one by modern standards and especially by the standards of accelerators for AI is very weak, but in some scenarios it is not performance that is more important, but the amount of memory. RTX 3080 20GB and Radeon RX 580 16GB have already been created  [caption id="attachment_84871" align="aligncenter" width="780"] RTX 3080 20GB and Radeon RX 580[/caption] Moreover, at least in the case of the RX 580, of course, we are talking about modifying far from new adapters. It is reported that many of the cards that are now getting new life were previously involved in mining. Moreover, interestingly, companies engaged in such production provide a two-year warranty on such adapters. Apparently, the demand is such that overall sales more than cover the risks of failure of devices that are no longer new.

[phonemantra-blog]

In monetary terms Nvidia is today's largest semiconductor manufacturer based on revenue and profit. We already talked about this the other day, and now another visual graph has appeared demonstrating the rapid success of the company.  [caption id="attachment_84806" align="aligncenter" width="780"] Nvidia[/caption] This is a graph of the revenue of data center chip companies. As you can see, Nvidia became the first here too, with a huge gap from its competitors, and the turning point occurred only at the beginning of this year.  Nvidia is turning into a monster. Before this, for several years (actually much longer, this data is just not on the graph), the leader was Intel, which accounted for 45% to 60% of revenue or more. At the same time, Intel's share gradually decreased, while the shares of Nvidia and AMD, on the contrary, grew. However, in 2023, due to the AI ​​boom, Nvidia's share rose sharply, while the shares of AMD and Intel fell sharply. This does not mean that companies began to earn less - AMD is actively increasing sales of server CPUs, but Nvidia's growth is so great that the shares of AMD and Intel have actually decreased.   [caption id="attachment_84807" align="aligncenter" width="780"] Nvidia[/caption] Today we already said that in the last quarter alone, Nvidia sold about half a million H100 and A100 accelerators  and sales will only grow in the near future.  

[phonemantra-blog]

The largest customers bought 150,000 H100 adapters each In recent months, we've seen a lot of news about how Nvidia is selling its AI accelerators left and right, especially the most powerful H100. But only now has it become known how many of these adapters were actually sold. [caption id="attachment_84778" align="aligncenter" width="780"] Nvidia[/caption] According to Omdia, Nvidia managed to sell almost half a million H100 and A100 accelerators in the third quarter alone. There is no data on the sales ratio of these models, but probably the main emphasis now is on the H100. Nvidia sells half a million H100 and A100 accelerators per quarter In the current quarter, analysts expect that Nvidia will be able to sell more than half a million of these accelerators. If you look at the chart only for the H100, you can see that by the end of the entire current year, Nvidia is expected to supply about 650,000 of these adapters to its largest customers alone. [caption id="attachment_84779" align="aligncenter" width="780"] Nvidia[/caption] Moreover, the demand is so great that now the waiting time for deliveries reaches 52 weeks, that is, a whole year! You can also see that the largest buyers of the H100 are Microsoft and Meta*, each of which purchased approximately three times as many accelerators as its closest competitors such as Google, Amazon, Oracle and Tencent. * The Meta company is recognized as extremist in Russia and is banned.

[phonemantra-blog]

Parameters are still unknown As you know, new US sanctions prohibit the supply of various accelerators for AI to China, including the H800 and A800 models, which, in fact, were created for China. However, Nvidia is simply going to pull the same trick again with the creation of special versions that will not fall under the next sanctions. [caption id="attachment_81201" align="aligncenter" width="600"] Nvidia[/caption] HGX H20, L20 PCIe and L2 PCIe models expected soon Nvidia is reportedly preparing three accelerators at once: HGX H20, L20 PCIe and L2 PCIe. They can be presented in the very near future: after November 16. It is still unknown what characteristics the new accelerators will have, but it is worth remembering that the new sanctions limit the supply of solutions with performance above a certain level, so Nvidia will clearly somehow limit this very performance. Let us also recall that the gaming RTX 4090 was also banned, but at least Nvidia has other options within the same line, and new ones will appear soon .

[phonemantra-blog]

January 8 It looks like Nvidia GeForce RTX 40 Super video cards will be presented in early 2024 at CES 2024. At least Nvidia has already announced a special event. [caption id="attachment_80889" align="aligncenter" width="600"] Nvidia's[/caption] It will take place on January 8th. Nvidia, of course, does not say anything directly about the new video cards, but the event is called GeForce Special Address, which indicates specifically 3D gaming cards. Nvidia's "super graphics cards" appear to be announced in January. In addition, you can remember that last year at CES the company announced the mobile RTX 40 Laptop and the desktop RTX 4070 Ti. As numerous leaks indicate , at least three models await us: RTX 4070 Super, RTX 4070 Ti Super and RTX 4080 Super. Whether there will be an RTX 4060 Super and RTX 4090 Super (or Ti) later is still unknown.

[phonemantra-blog]

Early next year At the beginning of next year, new video cards of the GeForce RTX 40 Super line should appear on the market, as reported by VideoCardz with reference to insider kopite7kimi. The source is confident that the announcement of GeForce RTX 40 Super gaming video cards will take place at the famous CES exhibition, which will take place early next year in Las Vegas. It was previously reported that the line would feature three models. The GeForce RTX 4080 Super will be based on the AD103 chip with 10,240 CUDA cores and will receive 16 GB of GDDR6X video memory with a 256-bit bus at a TBP level of 320 W. A line of new GeForce RTX 40 Super video cards will be released at CES 2024 [caption id="attachment_80345" align="aligncenter" width="780"] GeForce RTX 40[/caption] The GeForce RTX 4070 Ti Super will receive an AD102 or AD103 GPU equipped with 8448 CUDA cores and 16 GB, and the GeForce RTX 4070 Super will receive an AD103 or AD104 with 7168 cores and 12 GB of memory. Insider kopite7kimi previously accurately reported the characteristics of the GeForce RTX 30 and RTX 40, and also told us in what order the RTX 40 video cards will be released.

[phonemantra-blog]

But she continues testing her new adapters Fresh US sanctions affected not only the supply of various GPUs to China. Also on the American blacklist were some Chinese companies, including Moore Threads, which is known for its video cards. There is no exact data, but usually being blacklisted means that a company on it loses access to modern technical processes, since semiconductor manufacturers are prohibited from producing products for such companies. However, for now, Moore Threads does not despair and is preparing for the release of its new video cards.  [caption id="attachment_71704" align="aligncenter" width="780"] video cards[/caption] Are we expecting powerful Chinese video cards shortly? The company is already testing the MTT S90 model for consumer PCs and the MTT S4000 accelerator for data centers. The devices have device IDs 0301 and 0323 respectively, which may indicate they are 3rd generation Moore Threads GPUs.    Their characteristics are not yet available, but we can recall that the same MTT S80, although it cannot boast of high performance, is equipped with 16 GB of memory and offers a PCIe 5.0 interface and a GPU with 4096 ALU units. Current graphics cores for Moore Threads are produced by TSMC using the 7 nm process technology. If sanctions do mean the end of the two companies' collaboration, Moore Threads will have one option in the form of SMIC with its acclaimed 7nm process technology. 

[phonemantra-blog]

The error has already been fixed Just yesterday we reported that GeForce RTX 30 video cards can run the Frame Generation frame generator from DLSS 3, but Nvidia has already responded to the situation and reported that the technology does not work. [caption id="attachment_68775" align="aligncenter" width="780"] GeForce RTX 20 or 30 video cards[/caption] No, Nvidia did not cheat, and frame generation does not work on GeForce RTX 20/30 video cards. We have noticed that some members of the PC gaming community, such as the Discord community before this, are experiencing a bug in the old DLSS DLL version 1.0.1.0 that is erroneously generating frames on 20/30 series hardware. There is no generation of DLSS frames - it's just a raw duplication of existing frames. Frame counters show higher FPS, but the smoothness and playability of the game do not improve. This can be verified by evaluating duplicate frames using tools such as NVIDIA ICAT, or by looking for an error message in the Remix logs indicating an error in frame generation Thus, the technology not only does not work correctly, as it might seem, but does not work at all, and the performance improvement that special software displays is a mistake. Nvidia added that it has already fixed the DLL, and this trick will not be possible in the upcoming update.

[phonemantra-blog]

The new product shoots even better video DxOMark authors got to the base iPhone 15 and 15 Plus to evaluate their cameras. Of course, they rated them absolutely equally, so we’ll talk about the result of the iPhone 15, including the Plus version. [caption id="attachment_68037" align="aligncenter" width="780"] iPhone 14 Pro Max's[/caption] So, the new Apple product received 145 points, which puts it in 12th place in the overall ranking. In fact, if we do not count separately the devices with the same number of points, then we are talking about ninth place, which is very good for the basic iPhone. The base iPhone 15's camera competes well with the iPhone 14 Pro Max's camera. For comparison, the iPhone 14 Pro and 14 Pro Max, despite the presence of a “telephoto”, have only 1 point more! And the best result now, let us remind you, is the Huawei P60 Pro with its 156 points. It can also be noted that the iPhone 14 received 133 points a year ago, so the improvement is noticeable. If we look at the result in more detail, it is worth noting the record 80 points for background blur. And the worst indicator of the new product is expected for its zoom capabilities.

[phonemantra-blog]

The version with Radeon RX 6750 GRE 12GB is slightly more expensive than the version with RTX 4060 Just a few hours ago we talked about the Radeon RX 6750 GRE 12GB video card, and now MSI has presented ready-made PCs with this adapter. [caption id="attachment_68032" align="aligncenter" width="780"] GeForce RTX 4060[/caption] They are presented so far only in China, where they cost 700 or 890 dollars, depending on the CPU. More precisely, these are the prices with the Radeon RX 6750 GRE, although you can also choose a configuration with a different 3D card. Another competitor for the GeForce RTX 4060, but not the most modern. In terms of price, the version with Radeon RX 6750 GRE is almost identical to the version with RTX 4060 and cheaper than the version with Radeon RX 6750 XT. That is the performance of the Radeon RX 6750 GRE is at least slightly lower than that of the Radeon RX 6750 XT. That is, approximately the same as the Radeon RX 6700 XT. If you look at the latest tests, you can see that the Radeon RX 6700 XT is about 10% faster than the RTX 4060 in Full HD and up to 15% in 4K. So, if all the assumptions above are correct, a PC with MSI's Radeon RX 6750 GRE will be slightly better off than one with an RTX 4060, but that's without taking into account support for the latest DLSS and ray-traced performance. Unfortunately, we still don't know the specifications of the Radeon RX 6750 GRE, although available MSI data suggests that there will be minimal differences between the Radeon RX 6700 XT and RX 6750 XT.

[phonemantra-blog]

The cooling system is similar to that of the GeForce Founders Edition Galax company presented the same video card with a hole that it showed yesterday. This is indeed an RTX 4090, and the hole is indeed for cooling. [caption id="attachment_68019" align="aligncenter" width="780"] Galax RTX 4090[/caption] The card is called the RTX 4090 20th Anniversary Edition and was released in honor of the 20th anniversary of Galax itself. True, the company was founded in 1994, that is, it will turn 20 years old only next year. [caption id="attachment_68020" align="aligncenter" width="606"] Galax RTX 4090[/caption] The card is called the RTX 4090 20th Anniversary Edition and was released in honor of the 20th anniversary of Galax itself. True, the company was founded in 1994, that is, it will turn 20 years old only next year. Galax RTX 4090 20th Anniversary Edition Unveiled [caption id="attachment_68022" align="alignnone" width="594"] Galax RTX 4090[/caption] The adapter looks quite unusual due to the two fans shifting to the right. This same hole allows air to pass through the radiator of the cooling system, and on the reverse side, there is another fan. As a result, the CO configuration is similar to what Nvidia uses in its GeForce Founders Edition models. [caption id="attachment_68024" align="aligncenter" width="594"] Galax RTX 4090[/caption] Another feature of the card is a hidden power connector. It can also be noted that the adapter is quite large: 344 mm in length and 60 mm in thickness. However, the GPU frequency is not a record and is 2595 MHz. The new product will go on sale on November 11. This will probably be a limited edition. [caption id="attachment_68025" align="aligncenter" width="715"] Galax RTX 4090[/caption]

[phonemantra-blog]

And almost 15 times faster than AI At the System LSI Tech Day 2023 event, Samsung talked about the upcoming Exynos 2400 single-chip system. [caption id="attachment_64092" align="aligncenter" width="600"] GPU[/caption] For example, it is now known for sure that the graphics core of the new SoC is called Xclipse 940, and it is based on the AMD RDNA 3 architecture. Let us recall that new data suggests that Samsung will continue to develop its GPUs based on AMD technologies. CPU 1.7 times faster than Exynos 2200, and GPU based on RDNA 3 architecture The performance of the Exynos 2400 processor has increased by an impressive 70% compared to the Exynos 2200, but it is worth remembering that the Exynos 2200 was released not last year, but almost two years ago. AI performance increased 14.7 times. For now, that's all we've been told, but from other leaks we know that the new SoC will have a 10-core CPU, and the performance will be significantly lower than that of the upcoming Snapdragon 8 Gen 3.