Have anyone seen sales numbers for the Sam Coupe, MSX2 and Amstrad plus/GX4000. I assume the first and only batch of Amstrad pluses/GX with were several hundredthousands machines.
As the thread has been revived anyway: I was a Sam owner back in the day and the number I have seen most often for it is just 13,000 machines produced, quite possibly all by MGT. Different people have different opinions as to whether SamCo manufactured only, but West Coast Computers' were definitely just cheaply-modified unsold stock.
I think the main issue with the Sam Coupe is the shared memory pool.
Same as the CPC but it was more severe on the Sam resulting in an effective CPU speed of only 14% faster than the speccy.
Also considering that the Sam had to push more pixels than the speccy, that percentage went down even further.
Same number of pixels as a Spectrum (except in its Mode 3), but at four bits per pixel rather than 1.125. Absolutely crippled by memory contention though, as you say: that's 128 bytes for the pixel area in the main two modes, with the active area being about 2/3rds of the display. So the display requires exactly 50% more memory bandwidth than does the CPC's.
The same rule as the CPC applies anywhere that pixels are not on display: the CPU may make one memory access every four cycles. Where pixels are on display, it may make an access only once every eight cycles.
Mode 1 is the Spectrum-compatible mode, and slows things down even further by applying the once-in-eight penalty approximately half the time even while pixels aren't being fetched. Mode 2 is a Spectrum-esque mode but pixels are linear and attributes apply to 8x1 regions. There is no additional slowdown.
Ordinarily the pixel portion of the display is 192 lines high out of 311, so pixels in general are being output close to 41% of the time.
Assuming everything's being run in RAM, that the display is normal, and that memory accesses are the limiting factor then, given the nominal CPU clock speed of 6Mhz, for almost 59% of the display the processor is running about 150% as fast as a CPC. For the other 41% it's running about 75% as fast.
So I make that about 20% faster than a CPC. But with a fixed 24kb frame buffer, and no hope of getting hardware help at all. Picking which 32kb page it falls in is the limit of your ability to pick video modes and it is not believed that there is any way to trick the ASIC into deferring its work.
In net: think of the best software scroll you've seen on a CPC and imagine it being about 25% slower.
There are per-line interrupts and you can disable the display entirely and get that processing time back so it's possible to create a shorter display but it's not common because the display is already quite wide compared to its height: pixels are about 125% as wide as tall, so the aspect ratio with all pixels allowed is approximately 5:3.