SGI: Hardware

My Onyx2 currently has 2 RM9s. If I can get my hands on an RM10, would you say that one RM10 is superior to two RM9s? I know the RM10 has 256MB of texture memory, and the RM9 has 64MB. So one RM10 would still have twice the texture memory than two RM9s. I think the analogue question applies to RM10 vs RM11. Any thoughts (or even facts)?

Wrong.

When you have two RMs in an Onyx2, raster memory adds up (so you can have more features or larger bit depth), but texture memory is used in parallel to speed things up.

So any IR pipe with RM9s has 64MB texture RAM, regardless of the number of RMs. Any IR3 has 256MB, or four times what IR2E has.

As usual, JJ's answer is right on the money. And while the extra texture memory of an RM10 will indeed offer the benefits he mentions, it's also probably worth noting that the raster memory side of the RM10 offers no distinct advantage over the raster memory provided by the same number of RM9s. Each individual Onyx2 RM is actually a interconnected two board set, each comprised of a raster memory and texture memory board. RM9s and RM10s both use the same 030-1402-00x raster memory board.

I see. So one RM10 means actually four times the texture memory of two RM9s. In the same time it means a downgrade in raster memory to half of what I had before. And while the texture memory of one RM10 is four times that of two RM9s there is a speed disadvantage because it is only one board. Did I get it right?

The RM11 on the other hand has 1024MB of texture memory so again an increase by multiplier 4 as opposed to RM10 (no matter how many. But I understand the raster memory of the RM11 is substantially bigger. (2,5 GB opposed to 160MB?) so in that case one RM11 would mean an increase in raster memory and texture memory over any combination of RM10s. Is that correct so far?

Right on the money

Last, multiple raster managers results in more raster memory, but also a higher fill rate because the raster managers are interleaved. So, theoretically multiple RM9/RM10s may have a higher fill rate than a single RM11, but whether that is true remains to be seen. Very little benchmark data of IR4 is public.