Our hypothetical design combines an x86 and a PowerPC 604 processor on one chip. By using the same package, the processors can share all external memory, support logic, the FPU, and even the on-chip cache to reduce costs.
We chose a 486 core because it is less than one-third the size of the Pentium core. Even so, the 486 with necessary modifications would make the die size of this 615 about 20 percent larger than a standard 604 chip. The 486 should be able to keep up with the clock speed of the 604; both could reach 133 MHz in current IC processes. At this speed, the combined chip would deliver roughly 75-MHz Pentium performance for x86 software and 150-MHz Pentium performance on native-PowerPC software. By the second half of 1996 (when IB
M's 615 may debut), process advancements should allow both the 486 and 604 cores to reach as high as 180 MHz, giving a proportional increase in performance.
The Pentium's FPU is significantly faster than the 486's. But our design leverages the fast PowerPC FPU to achieve results competitive with a Pentium's, even with some 486-related overhead in the shared design.
The 486 core natively executes x86 instructions. This approach is faster and simpler from a design standpoint than alternatives that involve decoding and converting x86 instructions into RISC instructions. For a first-generation device, we think it makes sense to use a simple x86 core rather than adding a bulky translation unit. Translating x86 code into the PowerPC's RISC instruction set is difficult and may require modifications to the PowerPC core to handle x86 condition codes and other features.
Finally, a native-PowerPC OS would handle such things as the PC BIOS and core logic. DOS applications that directly access the hardwa
re would have to be emulated, as in SoftWindows today. Performance would be mediocre for these types of programs.
As we have designed it here, the 615 would cost about 30 percent more to build than a standard PowerPC 604. At current margins, this would be an additional $30 to $40. However, business considerations would probably encourage a vendor such as IBM to sell the chip without charging a premium.
Building a 615
illustration_link (20 Kbytes)
We Build a 615
