New to this? Read it in simple words
- AMD’s CDNA 3 design uses chiplets, smaller parts inside one processor package. The MI300A combines CPU and GPU chiplets with shared HBM3 memory.
- In a traditional processor, one faulty area may mean the whole piece is lost. Chiplets let designers build parts separately and combine the working ones.
- The links between chiplets can matter as much as the chiplets themselves. They must be fast enough to stop software from waiting.
- Chiplet
- A small chip that works with other small chips inside one processor package.
- CPU
- A general processor that runs most of a computer’s everyday tasks.
- GPU
- A chip that does many calculations at once. Most AI runs on GPUs.
A modern AI processor can be built from several smaller pieces. This can improve design choices, but the pieces must connect and work well together.
One processor built from smaller parts
AMD’s CDNA 3 design uses chiplets, which are smaller parts inside one processor package. The MI300 family puts computing and input-output parts on separate pieces, then connects them with advanced packaging.
The MI300A combines CPU and GPU chiplets with shared HBM3 memory. A CPU is a general processor, while a GPU is built for large groups of similar calculations. Shared memory can reduce slow data copies between them.
A traditional processor uses one large piece of silicon. If one area has a defect, the whole piece may be lost. Chiplets let designers build different functions separately and combine working parts.
The package is more than a small circuit board. Distance, signals, power, and heat affect performance. The links between chiplets must be fast enough to stop the software from waiting.
Computing parts, input-output parts, memory, and fast links must work together as one machine.
The connections are part of the design
Chiplets can use different factory processes for different jobs. Companies can reuse parts in several products and build systems that would be difficult as one large piece. The trade-off moves to packaging, heat, and link speed.
Software becomes very important. Compilers turn code into machine instructions. Runtimes manage work while it runs. Both must make the separate parts feel like one computer.
Shared memory makes programming easier, but physical distance still matters. Data is usually faster and cheaper to use when it is close to the processor that needs it.
Heat also connects every part of the design. Memory and fast links can create hot areas that reduce speed during a long job. Cooling and power control help decide the real performance.
Compare the full machine
Designers are no longer improving only one piece of silicon. They must make a group of parts work together. Hardware and software both affect the final result.
Many supplies and skills now matter. These include factory access, memory, packaging, fast links, low-level software, and communication tools. A weakness in one layer can waste the power of all the others.
Buyers should test complete workloads on the full machine. Useful measures include finished jobs, usable tokens, power, and total cost over time. The best design is the one whose separate parts feel invisible in daily use.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
Product numbers and efficiency statements are AMD’s claims; compare them with independent tests for your own workload.
AMD’s chiplet design lets companies build AI chips from smaller parts, though heat and link speed remain limits. How we rate the mood