Instinct MI300 is coming.
Chiplets and the ability to stack such chiplets on top of one another have to be two of the most amazing advancements in chip manufacturing in recent years. As they say, the options are limitless. At CES 2023, AMD demonstrated how adding additional cache to a processor can improve gaming frame rates with the Ryzen 7000X3D CPUs(opens in new tab). However, AMD also had something as amazing for those who work in data centers.
AMD has combined a CPU and GPU onto one very enormous chip, the AMD Instinct MI300, utilising its 3D chip-stacking technology. It’s not only that this chip has a GPU in addition to a CPU. These days, a GPU is essentially built into everything we think of as a CPU, so this isn’t very notable. AMD is also no stranger; it has been producing APUs for many years, which are essentially chips that house both the CPU and the GPU. Actually, AMD refers to the MI300 as an APU.
The AMD Instinct MI300 is a data center accelerator that has an impressive 146 billion transistors, almost double the size of Nvidia’s AD102 GPU which is found in the RTX 4090 and has 76.3 billion transistors. This massive chip was held up on stage at CES by AMD’s CEO, Dr. Lisa Su, and is so large that it appears to be the size of a decently sized stroopwafel. The cooling requirements for this massive chip must be significant.
The AMD Instinct MI300 has a GPU based on AMD’s CDNA 3 architecture, which is a version of its graphics architecture designed specifically for compute performance. The GPU is paired with a Zen 4 CPU and 128GB of HBM3 memory. The chip is made up of nine 5nm chiplets and four 6nm chiplets, which are believed to include six GPU chiplets and three CPU chiplets, as well as a four-piece base die and memory around the edges. The chiplets are connected by a 4th Gen Infinity interconnect architecture, although no further details have been provided about the exact makeup of the chip. By consolidating all components onto a single package with fewer data transfer points, the efficiency of the product can be significantly improved. This is because it reduces the number of calls made to memory off-chip, which can slow down the process. In high-level computing, it is important to prioritize both bandwidth and efficiency, making this approach a logical choice. It is similar to the principle behind AMD’s Infinity Cache on its RDNA 2 and 3 GPUs, which stores more data locally to reduce the need to access data from external sources and maintain high frame rates.
There are several reasons why an MI300-style accelerator is not currently available for gaming. Firstly, the price of the MI300 is likely to be beyond the budget of most gamers. Additionally, it has not yet been figured out how to program a game to recognize multiple compute chips on a gaming GPU as a single entity without specific coding, as past attempts with SLI and CrossFire have not been successful.
It is worth noting that the MI300 is an extremely large and powerful chip. According to AMD, the MI300 delivers an eight times improvement in AI performance compared to the company’s own Instinct MI250X accelerator, which is also a multi-chiplet chip with 58 billion transistors. The MI300 is also five times more efficient than the MI250X. The MI300 is expected to be released in the second half of the year, although it is likely to be a high-priced product that is more suitable for those working in data centers or AI with access to significant funds.
If you’re waiting on a CPU, why not buy a mouse right now?