AMD Instinct MI300X Beats NVIDIA RTX 4090 as Top GPU in Geekbench
The AMD Instinct MI300X is a Data Center GPU engineered to meet contemporary AI demands. Competing primarily against NVIDIA’s Hopper AI GPU family, the MI300X has made headlines by dominating the Geekbench OpenCL benchmark, securing the top position effortlessly.
MI300X Specifications
Built on the CDNA 3 architecture, the AMD Instinct MI300X is a technological marvel featuring an advanced chiplet packaging design with 153 billion transistors. The GPU includes 28 dies and eight HBM3 packages, offering up to 192 GB of memory capacity-the highest among current Data Center GPUs.
GPU and Performance Specs
The Instinct MI300X is equipped with 304 compute units totaling 19,456 cores. It boasts a VRAM bandwidth of 5.3 TB/s and an Infinity Fabric interconnect speed of 896 GB/s. With a TDP of 750W, it consumes nearly double the power of NVIDIA’s RTX 4090, which has a TDP of 450W.
Benchmark Performance
In the Geekbench 6.3.0 OpenCL test, the AMD Instinct MI300X achieved an impressive score of 379,660. The benchmark was conducted on an AMD EPYC system featuring two 9754 CPUs with 128 cores each and 3 TB of system memory. For comparison, the fastest gaming GPU, NVIDIA’s RTX 4090, scores 319,697 points, while AMD’s 7900 XTX scores 207,354 points. This highlights the exceptional speed of modern Data Centers and AI GPUs.
Here’s a detailed comparison chart of the AMD Instinct MI300X and NVIDIA RTX 4090 across various parameters:
Market and Usage
Due to the high power and cooling requirements and exorbitant prices, GPUs like the MI300X are not designed for standard PC platforms. For instance, a single RTX 4090 costs around $1500 to $2000, whereas an MI300X chip costs approximately $15,000.
These GPUs feature specialized hardware designs and accelerators tailored for Data Centers, AI, HPC, and Cloud servers, making them unsuitable for gaming and traditional PC applications. Moreover, the necessary driver support is lacking. AMD has also unveiled its next-gen Instinct roadmap, which includes the MI350 and MI400 series.
Originally published at https://blog.spheron.network on June 19, 2024.
