In 1997, NVIDIA was preparing the RIVA 128 graphics chip, on which the company's future depended. Its first chip, the NV1, did not bring the expected success, and the development of the next project, NV2, had to be halted due to a series of technical failures. According to Jensen Huang's recollections, the team had only one attempt left to release a new product. If a serious error had required redesigning the microchip, the company could have run out of money before receiving the corrected version. Jensen Huang's interview with Sequoia Capital.
What happened next? Read this article in full.
Before sending a chip to production, engineers create a digital description of its design: they define the necessary logical operations and connections between the circuit components. The design was verified using the circuit simulation tools available at the time, and upon completion of development, it was sent to the factory for microchip manufacturing. However, design verification does not guarantee the absence of errors in the final device.
Some issues are discovered during the joint operation of the chip and software. For example, a specific sequence of commands may lead to an incorrect result or a system halt. If the root cause lies in the hardware circuit and cannot be bypassed via software, simply installing an update will not be sufficient.
Engineers would have to modify the design and order corrected microchips.
Such a redesign would incur additional costs, and the product release would be delayed until the new samples are received and verified. Cadence on design verification and the cost of chip redesigns.
To conduct more tests before manufacturing the RIVA 128, NVIDIA acquired an Ikos hardware emulator. According to the IMD business school, the purchase cost one million dollars. Huang described the device as a refrigerator-sized machine that was connected to a computer in place of the future chip. IMD analysis, Huang's recollections.
The digital description of the circuit was loaded into the emulator, after which its hardware reproduced the operations provided by the design. The computer could send it commands, emulating the operation of a game and a driver, and receive results just as it would when working with a real device. This made it possible to test the behavior of the future chip together with software. Cadence on the principle of hardware emulation.
The team tested drivers for controlling the chip, as well as games and other applications. According to Huang's recollections, the system worked so slowly that they had to wait a very long time for each frame to appear on the screen. But even such a result helped engineers understand whether the future RIVA 128 was processing commands and rendering images correctly. Huang's interview with Acquired.
| Where the circuit error was discovered | How to verify the fix |
|---|---|
| In a finished microchip | Modify the design, order new samples, and wait for their manufacturing. |
| On the emulator | Modify the design, load it into the emulator, and conduct the test again. |
Time savings were achieved by reducing repeated manufacturing cycles. Cadence on the capabilities of hardware emulation.
A second advantage emerged: engineers and programmers could work in parallel. While some refined the circuit, others prepared and tested drivers.
By the time the microchips were produced, a significant portion of the joint hardware and software debugging had already been completed. IMD on the RIVA 128 development process.
After the chip was manufactured, it was still necessary to verify its actual characteristics, including operating frequency and power consumption.
The emulator reproduced the circuit's logic, so successful tests on it did not replace the verification of the finished device. Cadence on the differences between a model and a physical chip.
NVIDIA applied an already existing technology: in an Ikos announcement from April 1997, the company is listed as a user of hardware emulation. In this story, there were not yet video cards designing other video cards. Developers were assisted by specialized equipment for testing digital circuits. Ikos announcement published by EDN.
According to Huang, the manufactured RIVA 128 worked on the first try. NVIDIA managed to avoid a redesign that could have derailed the product's release. In 1997, the chip hit the market, and within the first four months, the company shipped over a million units. Huang's recollections, NVIDIA's official timeline. Thus, NVIDIA became a pioneer in hardware emulation of graphics accelerators, a practice the company still uses today.
The success of the RIVA 128 concludes the first story in our series. In the next article, we will move on to the RIVA TNT and GeForce 256 to understand how NVIDIA's graphics chips evolved after the company was saved and what tasks they gradually took on.