(Sorry that I am not an insider, but I did spend a half-day working on this answer.)
Various versions of the iMac G3 used the PowerPC 750, 750CX, and 750CXe. The RAD750 comes in two versions: "rad-tolerant" and "rad-hardened". According to a quote in this question, InSight used the rad-hardened version.
BAE (the manufacturer of the RAD750) says very little about how their products are radiation-hardened, but does like to tout that they are hardened.
- The size of the RAD750 die (130 mm2) is about twice the size of the PPC750 (67 mm2) and about three times the size of the 750CX/CXe (42.7 mm2). (Wikipedia)
- There are notably more (but not quite double) transistors in the RAD750 (10.4 million) than in the PPC750 (6.35 million). The PPC750CX/CXe has 20 million transistors, which is accounted by the on-chip L2 cache. (Wikipedia)
- There is some differences in the size of the fabrication process. The PPC750 was 250 nm, the 750CX and 750CXe were 180 nm (Wikipedia), the rad-tolerant RAD750 is 250 nm, and the rad-hardened is 150 nm (manufacturer sales sheet).
- Some special fabrication steps (from the PPC750 and from earlier BAE products) were eliminated for the RAD750, to help BAE contract out the fabrication to other companies. It has nothing to do with radiation, but is a difference between the processors. (article by Berger et. al.)
- Each circuit in IBM's original design was replaced by one with better radiation tolerance. Most of the new circuits were from Lockheed-Martin's 250 nm ASIC library, but many complex functions (32 bit incrementers, comparators, and rotators) had to be designed especially for the RAD750. Each circuit had the same function as the original IBM version, but was radiation-tolerant. (Berger)
- The PPC750 had over 800,000 storage elements, many of which were dynamic. All of these elements were upgraded to radiation-hardened, static versions. (Berger) Particular effort was put into the hardening of RAM cells, sense amplifiers, decoders, and latches. (presentation by Burcin)
- Combinational logic was redesigned to reject the brief transients that can be caused by radiation. (Berger)
- The phase-locked-loop (used to multiply the clock frequency) was redesigned to be more tolerant of radiation, and also to increase the chip's operating temperature range. (Berger)
- The PowerPCs were packaged in plastic; the RAD750s are packaged in ceramic. This helps to absorb some radiation.
- RAD750s run at slower clock speeds. The rad-tolerant runs at 110 to 132 MHz, and the rad-hard runs at 200 MHz (manufacturer sales sheet). The iMac clocked between 233 and 700 MHz.
- Both the iMac and RAD750 processors used a support chip that provided a PCI bus and some peripherals. However, these support chips are entirely different. In particular, the one for the RAD750 is designed to make it easy to upgrade from the RAD6000, and also has SpaceBus support. (Berger)
- One thing that BAE is very clear about is that each chip is extensively tested. The amount of testing that is done simply would not be economical for commercial chips.
- Related to design and testing, BAE's products are certified to meet various standards. You don't get that with the commercial PowerPC chips.
Confirmed not different:
- All of the processor discussed here are CMOS.
- The general chip layout of the RAD750 is the same as the PPC750. (Berger)
- All of the processors discussed here are single-core (Wikipedia). Furthermore, the RAD750 simply does not have enough transistors to account for a second (redundant) core. It's worth noting that Curiosity and Perseverance each have two full RAD750-based computers, for redundancy.
- Both the PPC750 and the RAD750 do not include an L2 cache on-chip; however, both allow the cache off-chip. The PPC750CX/CXe have a 256 kB L2 cache on-chip.
BAE's page on radiation hardening has a link to this Military and Aerospace article, which suggests that their radiation-hardening is due to design differences:
At BAE Systems, pure rad-hard by design work has gone away over the past 10 years, but the company still is in the business of providing the most advanced rad-hard processors and general-purpose computers for government space and military programs.
“Rad-hard by design is no longer necessary,” Gonzales says. “We can leverage commercial foundries to have the wafers fabbed, and bring the wafers back into BAE Systems in Manassas. We operate fabless, but all our rad-hard capabilities are still here.”
Another quote in the same article:
In the future, it’s likely that rad-hard chip designers will combine techniques in rad-hard by design with radiation-mitigation techniques to ensure the highest possible reliability, yet provide strategic weapons systems designers with the latest electronic technologies. “For strategic applications, we could combine a rad-hard-by-design part where we can get them, with radiation mitigation with analog and digital filtering, shielding, and running computations at different times and comparing them to enable the system to respond to potential corruption,” Wiley says.
Still, there remains a need for rad-hard by design to the highest standards, even though the market for these parts is shrinking. “The foundry is a national asset,” Wiley says. “It is the only place that makes digital parts like processors and memories that can tolerate radiation to these levels.”
The substrate is probably different. Some versions of the PPC750 used silicon-on-insulator, but it appears the kinds used in the iMac did not. The manufacturer sales sheet of the RAD750 does not specify the substrate, but does tout "latchup-immune". Latchup can be caused by radiation and can be fatal to a chip. It usually is solved by silicon-on-insulator or silicon-on-sapphire (the latter is fairly standard for spacecraft processors).
By dividing the die area by the number of transistors, it appears that the transistors are generally larger. Fabrication process size is not a good metric here; it tells you that a manufacturer can make smaller transistors, but does not tell you if the actual size of transistors are smaller. Larger transistors are less susceptible to radiation.
The PPC750 did not have a watchdog timer, and the RAD750 makes no mention of it. However, it is an easy part to add to a CPU, and standard in aerospace systems, so it is likely present in the RAD750.
The following measures are discussed in Wikipedia, but there is no evidence either way for the RAD750:
- Using boron depleted of boron-10 in the passivation layer
- Using edgeless CMOS transistors