Article: Transistor Count: A Flawed Metric
By: Doug S (foo.delete@this.bar.bar), May 18, 2020 2:26 pm
Room: Moderated Discussions
I get what you are saying about how they shouldn't be included in a "count" for bragging rights (i.e. first to 100 billion or whatever) but what's wrong with counting them in the total when you are dividing by area to determine density when comparing processes? Those transistors are there, even if they are not active, and therefore the numbers for density are correct in including them, IMHO.
Look at it this way, if you have two designs, one of which has 20% non-active transistors to fill whitespace and another which has 2% non-active transistors because their designers did a lot of hand layout (or used magic, or whatever) and mostly avoided white space in their design. The latter design may be superior, but having a density that uses 18% less area for the same number of active transistors doesn't indicate the process is better. It indicates the designers are better.
The real problem with comparing processes this way of course is that not all chips are the same. As you say, the different optimization criteria of smartphone SoCs make them look "denser" all else being equal. I wonder though if there's any reason we shouldn't be able to compare Intel x86 and AMD x86 as a way to compare density of Intel and TSMC processes? They are both targeting high performance, and the chips have similar blocks - there is some variation in the amount of cache and so forth but it is a lot closer than comparing smartphone SoCs vs x86 CPUs.
Look at it this way, if you have two designs, one of which has 20% non-active transistors to fill whitespace and another which has 2% non-active transistors because their designers did a lot of hand layout (or used magic, or whatever) and mostly avoided white space in their design. The latter design may be superior, but having a density that uses 18% less area for the same number of active transistors doesn't indicate the process is better. It indicates the designers are better.
The real problem with comparing processes this way of course is that not all chips are the same. As you say, the different optimization criteria of smartphone SoCs make them look "denser" all else being equal. I wonder though if there's any reason we shouldn't be able to compare Intel x86 and AMD x86 as a way to compare density of Intel and TSMC processes? They are both targeting high performance, and the chips have similar blocks - there is some variation in the amount of cache and so forth but it is a lot closer than comparing smartphone SoCs vs x86 CPUs.
| Topic | Posted By | Date |
|---|---|---|
| New article: Transistor count: A Flawed Metric | David Kanter | 2020/05/18 07:04 AM |
| Non active transistors | Doug S | 2020/05/18 02:26 PM |
| Non active transistors | Ricardo B | 2020/05/18 09:12 PM |
| Minor quibble about fixed-performance ASIC | Paul A. Clayton | 2020/05/19 03:59 PM |
| Minor quibble about fixed-performance ASIC | David Kanter | 2020/05/21 06:58 AM |
| A complementary article about xtor density | Paul A. Clayton | 2020/06/02 07:07 AM |
| Low leakage transistors | David Kanter | 2020/06/02 07:53 AM |
| Transistor count: Metric is often GE | Chris L | 2021/01/03 09:39 PM |
| Transistor count: Metric is often GE | David Kanter | 2021/01/04 09:48 AM |
| Transistor count: Metric is often GE | Chris L | 2021/01/08 12:38 AM |
