Larry,
Here's some info on the HP stuff :
From the press release :
Each PixelFlow "flow unit" contains two powerful PA-8000 microprocessors and 8,192 pixel processors that can handle up to 16 million triangles per second.
This is a distortion so large it's scary. Assuming that the pixel processors are still running at 100 MHz, 16 million triangles per second is 6.25 cycles per triangle. This would only be true if all of the triangles sent down are rejected on the first edge-test as being outside the region being rasterized.
Unless the design has been extremely reworked from the UNC design (I don't believe HP put much engineering into this), peak performance on the rasterizer for the best-case triangles, (flat-shaded, non-textured) is no better than 20 cycles per triangle, 5 million triangles per second. Gouraud-shaded triangles are probably 2/3 to 1/2 of that. Textured triangles where the texture will affect the transparency of the pixels will be laughably slow, and any in-order requirements for triangle processing will obviate most of the parallelism benefits.
There is no programmable shading. A high end system was quoted at something like $1.5M because it needed a 16 cpu Exemplar to host it. The performance is okay, but the price is not competitive with an equal performing Onyx2 IR.
Another article said
Analysts said the system, which Hewlett-Packard dubbed the HP
Visualize PxFl, will be between 40 and 100 times faster than the best systems currently available. It is due to be be shipped in the
first quarter of 1998
This assertion is pure vapor, and unfounded, AFAIK.
The other thing to consider is that this machine has got to be forward compatible with Merced, as the projected lifetime for machines shipped in Q1 98 straddles Merced's hypothetical release date. Assuming HP is still working on Merced and planning to release it, this will impose some rather hard compatibility problems in the design of any new high-end machines.
-justinb |
