4 ms·
The article notes some advantages of being able to use different processes for different wafers, but there's not much more on that. It might be helpful if you w
by Animats 4y ago
The article notes some advantages of being able to use different processes for different wafers, but there's not much more on that. It might be helpful if you wanted different fabs for memory and compute, or for flash type memory. Anyone know what they're getting at here?
There are situations when you really need that, but they mostly involve imagers. The Advanced Scientific Concepts flash LIDAR had a two-chip stack, with the light detectors made with InGaAs technology. The counters and timers were ordinary CMOS. This also shows up in some IR sensors.
- yaantc 4y agoMemories have their specialized processes indeed, but there are other reasons to specialize. New nodes are fine for logic, but it takes time for analog IPs to move to new nodes (and some may not). So what AMD did, using an advanced node for compute/logic and a less advanced one for I/Os should be typical. You can also see this in broadband cellular modems, where the baseband part is on an advanced node and the RF on an older one. When you look at processes offering, you often have variants optimized either for peak performance (frequency) or maximum efficiency. The peak performance would be the natural choice for (big) CPUs, and an efficiency node better suited for a GPU or any massively parallel accelerator where efficiency is more relevant than peak frequency (on this, I think Intel planned to use TSMC for their HPC GPU, could be related: they can focus on high perf for their CPUs).
- to11mtm 4y ago> The article notes some advantages of being able to use different processes for different wafers, but there's not much more on that. It might be helpful if you wanted different fabs for memory and compute, or for flash type memory. Anyone know what they're getting at here? Well, one advantage is that it can be a cost/efficiency savings on a few levels. For example, in the case of Zen2/3, the main CPU chiplet is at 7nm, but the I/O die is at 12nm. If I had to guess, generally speaking it is useful in cases where parts of the final module would benefit from higher transistor density versus others; Memory controller tech generally gets fewer updates than the CPU/APU itself, so it allows faster design cycles; you already know the existing I/O controller works, one less portion to re-qualify.
- cm2187 4y agoMy understanding is that one of the key benefits is that if you have an imperfection on a wafer, you lose one small chiplet instead of losing a big monolithic CPU. So for equivalent imperfections you get a better yield.
- loonster 4y agoA much better yield. If the yield is 1/8, reducing the chiplet to 1/16 the size will increase yield to 7/8
- atq2119 4y agoDifferent functions scale differently. With the latest processes, logic scales best, while analog scales worse: SRAM (which is internally pretty analog) still scales decently, but less than logic; and I/O scales very little at all (think of it this way: the size of transistors that drive outputs is pretty much determined by the current you need to drive, and the current is determined by e.g. the PCIe spec, which is itself subject to the physical constraints of relatively long wires). As a consequence, if your design has CPU dies and I/O dies, using a smaller process only for the CPU dies is likely to be a good trade-off.
- deepnotderp 4y agoExamples could be GPU chiplets using processes optimized for throughput (density * peak perf product) vs CPU chiplets optimized for peak perf
- ksec 4y agoIt is just cost reduction and speed of iteration in silicon. You can now mass manufacture a new CPU Core in a new Node, while your Media Engine, the Hardware Video Encode and Decode stays the same on an old node. Your Media Engine is the same across the entire product line up, so you get economy of scale with Fabs just printing it, you can also iterate on your Media Engine for new functions or errata individually without stopping progress on the larger SoC Design. The same on I/O where there are speciality node much better suited for I/O, and their rate of progress is quite slow compared to other part of the system. But most of these are only good for Desktop and Server. And the world is now largely Smartphone, or Laptop Energy efficiency chips. Where integrated SoC still provides the best results in efficiency.