Altair for GPU shares many points in common with the CPU version, it is a simplified version oriented parallel computing and rendering.
So Aldebaran G1 is a VLIW in order processor.
128 KiB L1 data Scratchpad memory.
128 KiB L1 instruction Scratchpad memory.
32 KiB L1 data cache (Set-associative 4 ways).
1 MiB L2 Texture cache (Set-associative 4 ways).
2 MiB L2 cache Scratchpad memory.
1 MiB L2 Buffer cache (Set-associative 4 ways).
The calculation unit can only execute 2 instructions / cycle.
- 16 VFPU (Vector Float Point Unit), 16 FDIV (Float Division), 16 EFU (Elementary Function Unit).
- 16 ALU (32 bits , no division), 16 LSU.
The EFU will have all the necessary instructions to speed up calculations and 3D rendering.
(sin, cos, arctan, dot, rasterization, Min / Max, Clip, Culling , Sorting etc etc).
L2 SPM (2 MiB) stores the vertex / textcoord / normal / index / color / other + code of the shaders.
If this is insufficient, it is stored on the L2 cache.
The Pixel Process Unit is used to manage the Frame buffer / Zbuffer and Blending.
The compute unit does the various necessary calculations and indicates a pixel line to write.
The PPU takes care of this, internally it will have a buffer and a queue to manage the different line draw commands of the different Compute Unit.
For the textures, here is what will have to manage:
- RGB/RGBA 32 bits/16 bits
- brightness/gray 8 bits
- DXT1/DXT3/DXT7
- VQ Compressed
- Palette 8 bpp and 4 bpp
