Intel Pentium – A Conversation with Don Alpert
Donald “Don” Alpert managed the architectural team of the original Intel Pentium processor. The interview was conducted by Henner Thomsen of Stay Forever as part of the research into a podcast episode about the Intel Pentium.
Stay Forever: What was your specific role in the Pentium processor development project (which started out as the 586 project of course)?
Don Alpert: I was hired by Intel in Jan 1989 while the design team was finishing up the 486 design, so I was the first engineer assigned to work on the next-generation P5 design. I started by running benchmarks on the 486 RTL to identify cycle-by-cycle opportunities for improvements through parallel execution, branch prediction, floating-point speedup, and other ideas. Eventually, I was the architecture manager with a team of about 5 designers, including Bob Dreyer, Jack Mills, and Tejpal Chadha.
Stay Forever: Was there ever a consideration to adopt a RISC architecture for the Pentium processor? If so, were there discussions about licensing the Alpha instruction set?
Don Alpert: There was no consideration of using a RISC architecture. Intel was just introducing the i860 RISC architecture at the same time as the i486. The i860 was oriented toward high floating-point performance with an exposed pipeline for vector performance. Intel’s PC system manufacturers valued total software compatibility, and in fact Compaq in particular invested in startups designing x86 compatible chips to avoid total dependency on Intel.
Stay Forever: What was the primary target market for the Pentium processor? Was it high-end server/multi-processor systems, business desktop PCs, or home user PCs?
Don Alpert: All of those. For example, graphics performance was important for home users and desktop PCs, both 2D for Windows and 3D for workstations, and there was an emphasis on making the processor (cache and bus) multiprocessor ready for servers. Intel had customers addressing all these markets, such as IBM, HP, Compaq, and Sequent.
Stay Forever: I understand that you collaborated with the marketing department to identify product requirements. What insights did they provide about customer needs? Were gaming or multimedia CD-ROMs considered during this process? It’s worth noting that both the first CD-ROM PC game and the first CD-ROM multimedia encyclopedia were released in 1989, around the time when the Pentium project commenced.
Don Alpert: I don’t recall any processor requirements for gaming or CD-ROMs, but it is likely there was influence on the chipsets and peripherals, especially as the PC was evolving for Windows GUI and storage standards.
Stay Forever: Were there any major changes or revisions of the original plans for the Pentium architecture during development?
Don Alpert: The biggest change I recall was adding System Management Mode. Intel had redesigned the 386 CPU for static design and SMM to support the 386SL line of processors for power-managed notebooks. The strategy changed to designing the processors to include SMM and static design (for the P54C compaction) instead of having to invest in modifying the processors for these features.
Stay Forever: Did you collaborate with game developers to optimize the Pentium processor specifically for gaming? For instance, Id Software significantly optimized their Quake engine for the P5.
Don Alpert: I did not, but it is possible the technical market group promoted game optimization. There was an emphasis on improving floating-point performance for graphics, but the design team was focused on workstation graphics.
Stay Forever: In your opinion, which architectural change contributed most significantly to performance improvements: was it the superscalar architecture, 64-bit bus, or something else?
Don Alpert: This was studied carefully, and there was synergy among the features, but I don’t recall the results. I would guess that superscalar execution was the key development, which allowed some of the disadvantages for x86 compared to RISC to be reduced: parallel cache accesses to support the limited number of registers.
Stay Forever: Would you consider the superscalar architecture as a precursor to modern multi-core processors? Was there any consideration given to multi-core CPUs during that time?
Don Alpert: No, superscalar was a separate design decision from MP. The budget of 3M transistors would not have supported multi-core efficiently and mainstream PC software would not have been able to use multi-core efficiently.
Stay Forever: Was there ever a consideration to abandon backward compatibility with legacy x86 code? What level of performance improvement could have been realized without it?
Don Alpert: No consideration. I think the performance difference was minimal, just an incremental cost for circuitry and design complexity, such as parallel decode of common byte-variable x86 instructions.
Stay Forever: Reflecting on it now, do you have any regrets about the design of the Pentium processor, aside from the FDIV bug?
Don Alpert: No, the design effectively addressed the targets we set at the time, and enabled rapid advances in the PC market during the mid-1990s while selling hundreds of millions of CPUs.
Stay Forever: Some sources suggest that the original P5 Pentium wasn’t particularly successful even before the FDIV bug surfaced. Could you provide more context on this?
Don Alpert: The 5-year plan documented in 1990 called for selling 1M chips in 1995. That estimate was off by 10X, as almost 10M were sold. So this was a success, though, of course, the FDIV bug resulted in a $500M financial write-off.
Stay Forever: In your opinion, did media outlets and PC vendors like IBM exaggerate the impact of the FDIV bug?
Don Alpert: No. The context at the time was that Intel had started a program of advertising the benefits of Intel Inside to end-users, offering a promise of performance and quality. That $200M program was a conflicting marketing message to the implicit engineering analysis that said errors were unlikely to occur. This period was an interesting example for Intel to address conflicts in its values: “results orientation” said to produce and sell as many chips as possible, “quality” said that the chips should work correctly according to specification, and “customer orientation” said to satisfy customers’ requirements and desires. When the FDIV bug hit, Intel was more oriented to business/engineering results, and that evolved to more of an end-user customer orientation.
Stay Forever: The P5 introduced active cooling to processors. Was this requirement anticipated during its development?
Don Alpert: Not to the extent it was required. The power to achieve the performance targets at high frequency was more than anticipated, so the airflow to cool the chips was more of a system-design challenge than anticipated.
Stay Forever: What impact did the Pentium have on the PC market?
Don Alpert: The Pentium showed that performance competitiveness to RISC could be realized with x86 compatibility, and attempts to bring up Windows on RISC processors were not successful commercially. The Pentium processor platform, including chipsets and peripherals, made for a very attractive environment for developers and users, such as graphics GUI, CD-ROM software distribution, multimedia applications, and USB peripherals.
Stay Forever: Given that ARM-based processors are now dominating many market segments in today’s landscape, would you say that RISC has triumphed over CISC?
Don Alpert: I think the application markets have become fragmented, with x86 still dominant for personal and business computing, ARM for cell phones, and a variety of microcontrollers for embedded applications. I would not consider ARM to be RISC, there are too many modes and idiosyncrasies in the instruction set. May of the microcontrollers also have older architectures (e.g., 8051) or architectures that emphasize compact code rather than following strict RISC principles. In fact, the 1980s RISC architectures, like i860, MIPS, SPARC, and Alpha have disappeared or are limited to niche markets. Power/PowerPC has continued commercial success for IBM’s systems. So, I think that despite the technical advantages of RISC over CISC, the value of compatibility for applications and I/O drivers has proven stronger for x86 in personal computer and data center applications.