So far, NVIDIA’s H100 GPU has been the most energy-intensive chip for data centers since its launch, as it had a maximum rated power of 700 W. However, this situation will change with the introduction of the Instinct MI300X, which will have even higher power consumption s0ll. The AMD Instinct MI300X GPU is listed in the presentation footnotes with a power consumption of 750 W.
In comparison, the Instinct MI250X GPUs based on the CDNA 2 GPU architecture consumed between 500 and 560 watts. This means that power requirements have increased by 34 to 50 percent in just one generation This increase in power requirements is to be expected due to the design of the chip itself and the performance offered. Diw AMD Instinct MI300X offers an impressive 8x performance increase in AI workloads compared to its predecessors. At the same time, the chip is 5 times more efficient, which means it can deliver the same performance while consuming significantly less power. These improvements were likely achieved through technological advances in chip architecture, the use of better materials, and power consumption optimization. AMD has focused on improving the performance and efficiency of its GPU accelerators to meet the increasing demands in AI workloads. Says.
The AMD Instinct MI300X GPU is a high-performance graphics processing unit that is expected to have 304 CDNA 3 processing units. In comparison, its predecessor MI250X only had 220 compute units. This 38% increase in the number of processing units means that the MI300X will offer a significant performance boost. Another impressive feature of the MI300X is the HBM3 memory with a capacity of 192 GB. In comparison, the MI250X had 128 GB of HBM2e memory.
The capacity of the HBM3 memory has been increased by 50%, which is a significant improvement. Since HBM3 is a newer memory technology, much faster transfer speeds are also expected. A major advantage of the Instinct MI300X GPU is the enormous 192 GB of VRAM. This allows it to run large language models with fewer GPUs than the competition. For example, the MI300X can handle up to 540 billion LLM parameters with fewer GPUs than competing products that use 80 GB H100 chips.
Increasing performance requirements in terms of GPUs and HPC accelerators are an ongoing development. As these components become more powerful, the demand for computing power is continuously increasing. The aforementioned Gigabyte server roadmap indicates that CPUs, GPUs and APUs are approaching the 1000 watt limit. AMD in particular has invested in various chip and packaging technologies to significantly reduce power consumption.
Although AMD is known as the company with the most power-hungry chip, they have still developed technologies to reduce power requirements. At the ISSCC (International Solid-State Circuits Conference) in 2023, Dr. Lisa Su, CEO of AMD, commented on energy efficiency in zettascale computing. She explained that without new designs, a nuclear reactor would be required to meet the power needs for such a system. This underscores the increasing challenge companies face to meet rising power demands. To address this need, AMD is expected to launch the MI300X GPUs. These GPUs will also be available in more conservative, lower-power variants, particularly in the PCIe form factor. Although the exact details have not been announced yet, it is likely that these variants will be available by 2024.
Source: WccfTech
16 Antworten
Kommentar
Lade neue Kommentare
Urgestein
1
Urgestein
Mitglied
Urgestein
Urgestein
Urgestein
Veteran
Veteran
Veteran
Mitglied
Urgestein
Urgestein
Urgestein
Mitglied
Urgestein
Alle Kommentare lesen unter igor´sLAB Community →