AMD was planning a major upgrade with its RDNA 4 graphics architecture, expected to significantly increase the number of shader engines compared to the current RDNA 3 models. The new Navi 4X models were anticipated to feature up to 9 shader engines, 50% more than the 6 shader engines in the existing Navi 31.
Such an innovation promised to boost graphics performance, crucial for demanding games and applications. However, according to recent updates, AMD has decided to cancel these plans. Reasons could include technical challenges and high production costs, which would have made these GPUs unaffordable for most consumers.
Thus, gamers will have to wait for the next significant update in AMD's lineup that can offer comparable performance improvements.
The personal computer market is experiencing a revolution with the adoption of artificial intelligence, leading to a significant increase in demand and, consequently, a rise in prices. By the end of this year, the share of computers with AI technologies is expected to reach 19%, and by 2024, it will increase to 60%. This trend is particularly noticeable in consumer segments such as gaming systems, where manufacturers are eager to integrate AI features to attract buyers.
Companies such as MSI and Lenovo are actively incorporating the latest AI technologies into their devices, including Microsoft Copilot, which accelerates this process. As a result of this demand, the cost of AI computers may rise by 15% soon.
The integration of artificial intelligence is also affecting the upgrade of computer architectures and components. Manufacturers are adapting their applications and components in a way that convinces consumers of the need to switch to new technologies. PC sales this year are expected to reach 267 million units, confirming the "green zone" for the industry.
Intel has confirmed in its latest update that instability issues with the 13th and 14th generation K SKU processors may be related to operating conditions that exceed standard specifications. This includes maintaining high voltages and frequencies during periods of increased heat. Analyzed cases show that shifts in minimum operating voltages may be associated with operation beyond these specified conditions. Although the specific cause has not yet been determined, the majority of reports on this issue have come from users with unlocked motherboards that support overclocking.
Intel observed that motherboards with 600/700 series chipsets often set BIOS defaults that disable temperature and power delivery safeguards, potentially exposing the processor to prolonged periods of high voltage and frequency. The company urges system and motherboard manufacturers to provide users with a default BIOS profile that adheres to Intel's recommended settings.
Additionally, Intel strongly recommends that users stick to standard BIOS settings to ensure operation within recommended parameters. Motherboard manufacturers are also advised to implement warnings for users about the risks associated with using overclocking features. Intel continues to actively investigate the root cause of this issue and promises to provide further updates as new information becomes available. A public statement regarding the issue's status and BIOS setting recommendations is expected to be published in May 2024.
AMD has released an updated version of its Radeon Adrenalin 24.4.1 driver, which includes several improvements and fixes as well as support for new games like the strategy game Manor Lords. This update aims to enhance the gaming experience and address issues faced by users in previous versions. Key improvements include optimizations for games such as HELLDIVERS 2 and Starcraft II, where users previously experienced crashes and stutters. The update also corrects texture display errors in Cossacks 3 among other technical enhancements.
Additionally, the new driver offers enhanced support for AMD HYPR-Tune features in games like Diablo IV and Suicide Squad: Kill the Justice League, ensuring higher performance and stability. Noteworthy are the improvements in loading times in World of Warcraft and fixes for image corruption in Dying Light 2 Stay Human: Reloaded Edition when using Radeon Boost and ray tracing.
The latest benchmark results from Geekbench reveal that AMD's new Strix Point APU significantly outperforms previous Phoenix and Hawk Point models. Equipped with up to 12 next-gen Zen 5 cores operating at 2.3 GHz, which is almost 1 GHz faster than its predecessors, Strix Point demonstrates notable performance improvements.
Identified by the code name 100-000000994-14_N and as "AuthenticAMD Family 26 Model 32 Stepping 0," the APU features 12 cores and 24 threads. It boasts a base frequency of 5.11 GHz, although it typically operates between 2.3-2.4 GHz. The total L3 cache is 24 MB, reflecting the inclusion of both Zen 5 and Zen 5C cores.
In the single-core test, the processor scored 1224 points, while it reached 12754 points in the multi-threaded assessment, marking a 59% improvement over previous tests. These scores place the Strix Point ahead of the fastest Phoenix (Ryzen 9 7940HS) and Hawk Point (Ryzen 9 8945HS) models, highlighting the superiority of the new Zen 5 architecture. Strix Point processors are expected to hit the market later this year, starting with mobile platforms and followed by more powerful chiplet-based models.
AMD is set to expand its server CPU lineup with the new EPYC 4004 series, compatible with the AM5 platform. These processors are based on the Zen 4 architecture and utilize the same "Raphael" dies found in Ryzen 7000 and Ryzen 7000 X3D chips. Targeted at single-socket (1P) configurations, the EPYC 4004 will offer up to 16 cores and 32 threads with a maximum cache of 144 MB. Both standard and enhanced versions with 3D V-Cache are expected in the lineup.
As of now, the exact launch date and model details remain undisclosed, but AMD is expected to provide this information in the coming weeks. Support for the new CPUs on AM5 platforms promises to extend capabilities for mainstream-class servers.
While competitors continually strive for innovation, AMD seems to have chosen a conservative path in developing its new Radeon RX 8000 series based on the RDNA 4 architecture. Instead of following market trends and integrating the latest GDDR7 memory as NVIDIA is doing with its new GeForce RTX 5090 and RTX 5080 models, AMD continues to use the outdated GDDR6 memory with a speed of just 18 Gbps.
This conservative choice might undermine the competitiveness and appeal of the new cards, considering that consumers expect not just affordable prices but cutting-edge technology. All models in the Radeon RX 8000 series, for both desktops and laptops, will be equipped with this memory, potentially limiting their performance and innovative potential.
Despite intensified competition from AMD and Intel, NVIDIA continues to maintain a leading position in the artificial intelligence market. The company is particularly focused on developing its software capabilities, which are crucial for leveraging the power of its hardware. Software optimizations, such as TensorRT, are key to maximizing the potential of NVIDIA’s chips.
The company recently announced the Hopper and Blackwell GPUs, which, despite their high specifications, require adequate software support to function effectively. Statements by Bryan Catanzaro, Vice President of Applied Deep Learning Research, affirm that the computational challenges of AI are the most significant in history. NVIDIA remains committed to actively developing and optimizing its software stack, including CUDA, to maintain its industry leadership amidst fierce competition.
In a recent interview, former NVIDIA researcher Rihard Jarc highlighted the critical role of the CUDA software ecosystem in reinforcing NVIDIA's position in the artificial intelligence market. He emphasized that CUDA has enabled NVIDIA to surpass competitors and achieve significant financial success.
According to Jarc, the primary difference between NVIDIA and competitors like Intel and AMD lies in the existence of CUDA, which remains an insurmountable "wall" for others despite their efforts to integrate their solutions and make them more open to developers. For example, AMD's ROCm platform and HIPIFY tools allow for the conversion of CUDA code into portable HIP C++, making it compatible with ROCm.
However, the use of such code porting tools can significantly increase costs for clients who opt for platforms other than NVIDIA. Jarc argues that investing in the CUDA ecosystem requires a one-time expenditure that delivers cutting-edge performance and access to essential features, making NVIDIA the more favorable choice for AI applications.
Companies like Tiny Corp recommend NVIDIA's AI systems over AMD, emphasizing that the NVIDIA platform "Just Works," providing users with all necessary functionalities without additional effort and expense.
New variants of NVIDIA's A100 GPUs have been spotted on the Chinese market, featuring an increased number of CUDA cores compared to the original specifications. These GPUs boast 7936 cores, which is 15% more than the standard A100 models equipped with 6912 cores.
These enhanced GPUs also come with 96 GB of HBM2E memory, utilizing a 6144-bit bus interface distributed across six stacks with 16 GB each. The graphics cards are in PCIe form factor, appearing to be prototypes initially designed for GRID systems.
The listings specify that these GPUs have a base clock speed of 1260 MHz, although boost speeds were not mentioned. The performance is expected to reach around 20 TFLOPs for single-precision and 10 TFLOPs for double-precision computing, comparable or superior to the A100 PCIe 80 GB model, which delivers 19.5 TFLOPs and 9.74 TFLOPs, respectively.
The reason for these GPUs hitting the market remains unclear, but it's speculated that NVIDIA may have produced these units in 2020 for specific clients, and they have now surfaced in the marketplace. Some of these GPUs are also available on eBay, priced between $17,000 and $20,000.
