Apple’s new M1 Ultra aims to beat Nvidia’s RTX 3090

$Apple (AAPL.US)$ ’s new M1 Ultra is a surprising combination of two M1 Max dies, fused together to create a single powerful chip. The M1 Max has a secret high-speed interface that allows Apple to combine two chips into one. The result is an M1 Ultra chip that has double the CPU cores, double the memory, double the memory bandwidth, and most importantly, double the GPU cores.
Apple calls this combination UltraFusion, and it’s effectively Apple’s own 2.5D chip packaging implementation. The chip industry has been turning to chiplets to design processors for years, with $Advanced Micro Devices (AMD.US)$ ’s Zen 2 and Zen 3-based Ryzen chips leading the pack for modern chiplet designs and until recently, performance. Apple rivals like $Intel (INTC.US)$ , Samsung, and $Qualcomm (QCOM.US)$ are in the early stages of working together on a new standard that could let companies build processors out of Lego-like chiplets. Apple is ahead of the pack with a chip that fuses together two separate GPUs.
$NVIDIA (NVDA.US)$ and $Advanced Micro Devices (AMD.US)$ have built similar solutions for combining two GPUs in the past, but as AnandTech points out, it looks like Apple has solved the holy grail of multi-GPU design here. Apple’s UltraFusion tech supports an impressive 2.5TB/s of bandwidth between the two M1 Max chips. That’s a huge bandwidth jump over what Nvidia offers with NVLink for SLI or AMD with Infinity Fabric, which are used as high-speed links between GPUs.