Multi-layer caches are supported as dynamic map service layers (use ArcGISDynamicMapServiceLayer). The map server creates a dynamic map image by consolidating all the tiles from each layer cache and taking a "snapshot". While this is quicker than going back to the source data (traditional non-cached map service) the resolution of the tiles may not be optimal for the map extent requested... as a result map contents can appear coarse and pixelated. This same behavior will occur any time a client requests a dynamic map image from a cached map service (fused or multi-layer). Currently there is no way to force a cached service to go back to the source data to create a dynamic map image.

Support for multi-layer caches in the ArcGIS Silverlight API is undocumented at this time because we generally discourage their use. In fact, multi-layer caches are unsupported the other web clients, Flex and JavaScript. The ADFs provide limited support - similar to Silverlight in that a dynamic map image "snapshot" of the layer cache tiles is generated.

Why are multi-layer caches discouraged for use with Web clients? Basically, its about performance. Multi-layer caches generate a separate cache for each layer, so if you have 10 layers in your map, you have 10 separate caches. A fused cache with the same 10 layers has 1 cache. If a client requests 4 tiles in the fused cache, it would have to make 40 requests to show the same layers in a multi-layer cache, thus reducing client performance dramatically. With that in mind, in almost any situation multiple layers can be logically and intuitively grouped. Base map layers which will not change often can be consolidated into a fused cache. Operational layers which may change frequently can be hosted in a non-cached service. In fact in some cases data which changes often can still be cached if you employ efficient techniques for partial updates to the cache. You may want to take a look in the ArcGIS Server help or blog posts for additional information. Here are a few links: