![\"Differential](\"https:\/\/blog.zuken.com\/wp-content\/uploads\/2018\/07\/john-berrie-differential-pairs.jpg\")
<\/p>\nYou can tell when something isn\u2019t as clear as it should be. The same questions come up time and again. You ask three experts and get three different answers. Routing differential pairs can be like that. Why? Because \u201cit depends\u201d – on exactly what signals those pairs are carrying and what kind of PCB you\u2019re creating.<\/p>\n
What can I tell you that\u2019s useful? Well I hope I can help to answer some of those repeat questions, so you can concentrate on the hard stuff.<\/p>\n
<\/p>\n
Let\u2019s start at that differential driver on the left ( Balance fan-out and end routing patterns separately at both ends<\/strong>. When you reach that receiver on the right, do the same as you did at the driver end. Don\u2019t consider the lengths within the coupled section at all when doing this. Once TRUE<\/strong> and COMPLEMENT<\/strong> are coupled, the signals travel in a different way. I think of coupled sections as apples and fan-out and end routing as oranges \u2013 and you can\u2019t add apples to oranges unless you want a fruit salad!<\/p>\n If you only want good signal integrity, make no corner tighter than 135\u00b0. Many device application notes tell you the same. That angle works well for 45\u00b0 routing. You\u2019ll hear that curves give you better signal integrity, but for most PCB tracks – even for very fast signaling – \u00a0that isn\u2019t true. Get curves wrong and you\u2019ll make it a whole lot worse – and what \u201cwrong\u201d means isn\u2019t always obvious.<\/p>\n If you\u2019re routing on a flex, you need curves, but that\u2019s for mechanical reasons. You have to follow mechanical constraints, so the skill there is in making sure you don\u2019t break signal integrity in the process.<\/p>\n Add length in coupled sections only, and keep the pair coupled all the way, because the differential impedance and delay-per-unit-length are different from the single-ended values \u2013 even within the same pair.<\/p>\n In this case, you almost invariably need to add length to the pair, not the single-ended signal. As with matching between pairs, add length in coupled sections only. The application notes or standards docs will tell you how much.<\/p>\n Now you\u2019ll remember that we matched the fan-out and end routing separately, so that\u2019s not what I\u2019m talking about here. This is about differences that arise within the coupled routing sections. This is sometimes called phase matching<\/strong>. The ideal is that TRUE<\/strong> and COMPLEMENT<\/strong> travel together in perfect complementary harmony, as shown on the left in the pictures below. There\u2019s no such thing as perfection, so we get as close to it as we can. The faster your rising and falling edges get, the fussier phase matching requirements become. When slower edges get out of phase, the phase difference doesn\u2019t affect the signal so much. You can see that in the pictures in the center and on the right. The phase difference matters most when TRUE<\/strong> or COMPLEMENT<\/strong> change state from low to high or high to low.<\/p>\n Differential impedance and delay-per-unit-length depend on phase matching. Without it, those numbers are different. To get good signal integrity, you need pretty good matching. You shouldn\u2019t be breaking out from the coupled pattern, so the usual culprits where phase mismatches are concerned are corners. The outside track goes a bit further than the inside track.<\/p>\n The trap is to think you can just add the differences in length to the shorter sides, somewhere near the corners. Those are sometimes called phase bumps<\/strong>. There\u2019s more to it than that, because the bump couples differently and the coupling round those corners gets quite fuzzy. You have to follow tight recommendations for exactly what those bumps should look like. If you don\u2019t really need phase bumps, then my advice is not to add them.<\/p>\n The best place to start is to balance the number of 135\u00b0 right and left turns and to try to minimize the number of turns. After that you can compensate for any remaining difference with a phase bump if you need to.<\/p>\n These PCI Express signals are making their way from a high-speed FPGA I\/O bank to their connector. I guess you can see there\u2019s a lot more to consider besides what I\u2019ve mentioned, but I\u2019d like to think this short blog post has helped answer a few of those golden-oldie differential-routing questions.<\/p>\n You can tell when something isn\u2019t as clear as it should be. The same questions come up time and again. You ask three experts and get three different answers. Routing differential pairs can be like that. Why? Because \u201cit depends\u201d – on exactly what signals those pairs are carrying and what kind of PCB you\u2019re […]<\/p>\n","protected":false},"author":31,"featured_media":11334,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"om_disable_all_campaigns":false,"footnotes":"","_links_to":"","_links_to_target":""},"categories":[228],"tags":[237,265,277,11659,11660,20141],"class_list":["post-10062","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-pcb-design","tag-fpga","tag-high-speed","tag-routing","tag-signal-speed","tag-trace-length","tag-verification-and-validation"],"acf":[],"yoast_head":"\n![\"Differential](\"https:\/\/blog.zuken.com\/wp-content\/uploads\/2018\/07\/differntial-driver-left.jpg\")
). Imagine the TRUE<\/strong> ( ![\"TRUE](\"https:\/\/blog.zuken.com\/wp-content\/uploads\/2018\/07\/differential-driver-true.jpg\")
) and COMPLEMENT<\/strong> ( ![\"Differential](\"https:\/\/blog.zuken.com\/wp-content\/uploads\/2018\/07\/differential-driver-complement.jpg\")
) outputs are adjacent balls on a fine-pitch BGA, fanned-out to inner layers. Whatever else you do, keep the track lengths to fan-out vias equal, and then keep the lengths after those vias equal until you start the parallel track pattern. That parallel track pattern is the coupled section, because there\u2019s electromagnetic coupling between the two sides.<\/p>\nCorner angles and curves<\/h3>\n
Matching to other differential pairs<\/h3>\n
Matching between differential pairs and single-ended signals<\/h3>\n
Matching within a differential pair<\/h3>\n
![\"Three](\"https:\/\/blog.zuken.com\/wp-content\/uploads\/2018\/07\/differential-pairs-3-waves.jpg\")
<\/p>\nThe real world<\/h3>\n
![\"Differential](\"https:\/\/blog.zuken.com\/wp-content\/uploads\/2018\/07\/differential-pairs-featured-img.jpg\")
<\/p>\n","protected":false},"excerpt":{"rendered":"