Before getting into technical writing, Abrash was a game programmer, having written his first commercial game in 1982, Space Strike for the IBM PC. The game was a PC booter. Other games he wrote were Cosmic Crusader (1982) and Big Top (1983) for the same system. After working at Microsoft on graphics and assembly code for Windows NT 3.1, he returned to the game industry in the mid-1990s to work on Quake for id Software. Meantime, Abrash also worked on the popular game Doom. Some of the technology behind Quake is documented in Abrash's Ramblings in Realtime published on the Dr. Dobb's Journal. After Quake was released, Abrash returned to Microsoft to work on natural language research, then moved to the Xbox team, until 2001.

In 2002, Abrash went to work for RAD Game Tools, where he co-wrote the Pixomatic software renderer, which emulates the functionality of a DirectX 7-level graphics card and is used as the software renderer in such games as Unreal Tournament 2004. At the end of 2005, Pixomatic was acquired by Intel. When developing Pixomatic, he and Mike Sartain designed a new architecture called Larrabee, which now is part of Intel's GPGPU project. Gabe Newell, CEO of Valve, has said that he has "been trying to hire Michael Abrash forever. [...] About once a quarter we go for dinner and I say 'are you ready to work here yet?'"

Michael Abrash wrote in the 90s several influential books and other publications which gave him a reputation as good technical writer among the software engineering community. In his 1994 book Zen of Code Optimization: The Ultimate Guide to Writing Software That Pushes PCs to the Limit, Abrash presents principles and theory applicable to today's programmers. The key point of the book was that performance must always be measured, and the book included a measurement tool called the Zen Timer to check if theoretical code optimizations actually worked. The presentation of step-wise program refinement empirically demonstrated how algorithm re-design could improve performance up to a factor of 100. Assembly language re-coding, on the other hand, may only improve performance by a factor of 10. Abrash also showed how elusive performance improvement can be. Simply improving performance in one sub-routine would only expose bottlenecks in other routines and so on. Finally, he demonstrated processor-dependent assembly-based performance improvements by comparing assembly language optimizations across X86 family members.

Another lesson Abrash offered was how hard-earned performance rewards by such micro optimizations could disappear or even make the program execute slower. Abrash was able to gain performance by carefully counting clock cycles of each instruction and ordering instructions in such a way that they used a minimal number of clock cycles. However, the number of clock cycles per instruction changed with new implementations of the x86 architecture and, as a result, his performance gains were lost from one generation to the next. For example, the bit-oriented "XOR EAX, EAX" (example in Intel assembler syntax) instruction was the fastest way to set a register to zero in the early generations of the x86, but most code is generated by compilers and compilers rarely generated XOR instructions, so the IA designers decided to move the frequently occurring compiler generated instructions up to the front of the combinational decode logic, making the literal "MOV EAX, 0" instruction execute faster than the XOR instruction. Still, carefully hand optimized assembler code will in general perform well on later x86 processors, even without retuning for the new architecture. The reason is that the optimization policies change only gradually between the architectures, so that many optimization assumptions valid for the original architecture still hold on the new architectures, for example "integer arithmetic is significant faster than floating point arithmetic" or "minimize memory accesses for maximum register usage".

In the year 1997 Abrash published the Graphics Programming Black Book, primarily based on his experiences and optimizations with the Quake 1 graphic subsystem.