Product Info | User Reviews | Article Images(4) | Image Gallery | Comments | Forum Thread
Axe
From Glaze3D to Axe
The new tool set was making steady progress after several months; a workable version was ready just before the end of 2000. With these tools expected to reduce design time by many months and Glaze3D requiring several improvements to make it a marketable part, the chip went partially back to the drawing board. While much of the core architecture remained the same, Axe was created with a tweaking and recoding on the new design system. The total amount of EDRAM was increased from 9 to 12MB, the internal memory bus was expanded to 1024-bits, and Matrix Anti-Aliasing and a programmable vertex shader were added as well. While making a fresh start, the Glaze architecture had provided the necessary insight to allow Axe development to move forward quickly.
Axe Specifications
600-800 Mpixlels/sec
Single Pass Quad Texturing
Single Cycle Dual Texturing
MatrixAA
12 MB EDRAM on 1024-bit Bus
DirectX8 1.1 Vertex Shader
50 Million Vertices/sec Sustained
Dual Chip Support over CBI (Custom Bus Interface)
Block diagram of Axe
Axe prototype board
Another Axe prototype card
Axe development
Bitboys set an extremely aggressive schedule for Axe, anticipating tape out in less than 6 months. The original schedule called for the release of Axe during the 2001 Christmas seasons under the name Avalanche3D. With the new tools in place, this appeared to be a realistic timeframe. This, however, would be the first time that Bitboys would take a chip design to final silicon, so there were still a few important things to learn.
Axe, as with every chip, had bugs early on in development. These bugs were rooted out of course, but only with time. The majority of the bugs were found relatively quickly, though some remained for several months, as additional little issues were discovered. In calculating the original schedule, Bitboys had not set aside a sufficient amount of time for verification. Testing and debugging required several months more than had been originally anticipated.
During the summer and fall of 2001, place and route was occurring by a partner company. Initially, early versions of the netlist were used for this, with newer verses as they became available. This place and route took considerable time, with the end result requiring considerably more time than expected. With the end of 2001 approaching, more bad news was on the way.