Inside ATI & NVIDIA: How they make frames flyby Anand Lal Shimpi on September 23, 2002 2:14 AM EST
- Posted in
Finally, after all of the testing the chip is "synthesized" meaning that the HDL is translated into a gate-level netlist or schematic that is used to physically manufacture the chip. The first cut placement route is done, which is the real layout or how the gates will look in silicon. At this point there's still no chip to play around with but the chip can be "produced" using what is known as a FPGA (Field Programmable Gate Array).
A FPGA is a generic logic device consisting of a large number of gates; the gates can be configured in such a way to effectively emulate the chip before production. The benefit of doing this is that the chip can be fully tested and operational for only the cost of the FPGA which is on the order of thousands of dollars, whereas it costs millions to get a chip to and back from a foundry. A particular type of FPGA is used by both ATI and NVIDIA and it's manufactured by a company called IKOS; the IKOS box, as it is often referred to, is effectively a very large FPGA used to emulate designs like the R300 or NV30.
The IKOS box can fully emulate a design however it isn't able to run anywhere near as fast as the final GPU will operate. While today's GPUs are running at speeds above 300MHz, the IKOS box can emulate the design running at a small fraction of that - in the KHz frequency range (1000 KHz = 1 MHz). With the GPU running that slowly, you can't gauge performance but you can run a system with the OS installed, test drivers, and even run games on the IKOS box (albeit at ~0.2fps).
After the verification process is complete and the layout is done then comes the elusive tape-out. This is the preparation of everything necessary to be sent out to the foundry for actual production of the chip. About 4 weeks later you get your first chip or A0 silicon as it is often referred to as. The testing doesn't stop during those four weeks however, simulations continue as well as verification of the PCBs (Printed Circuit Boards) that the GPU will eventually be soldered onto.
Once the first silicon (A0) is back verification of it starts immediately; all of the functional units are tested and any unexpected behavior is immediately noted and debugged. Focused Ion Beam (FIB) tools are sometimes employed to fix bugs in the chip; we introduced the concept of using a FIB tool in our Inside Intel article, but as a brief refresher, a FIB tool allows you to effectively perform surgery on a microprocessor without ruining the functionality of the chip. The tool allows you to cut or lay down new wires on a chip, down multiple metal layers, to fix a bug in the design without destroying the chip. This way ATI or NVIDIA can attempt to fix a bug and see whether the bug is actually fixed before sending the revised design back to the foundry; remember it takes about 4 weeks to get a chip back after tape-out and you don't want those 4 weeks to be wasted if it turns out that the fix doesn't really get rid of the bug.
Once all of the bugs are fixed the finalized design is taped-out and sent to the foundry for production. These chips are then sent through qualification where they are put through all sorts of tests for compatibility, thermal and voltage stresses, signal integrity, etc
After qualification is complete, it's time for production and now you know how GPUs are made.