IBM had been working on electronic arithmetic prior to WWII, and the IBM 603 Electronic Multiplier (1946) was IBM's first deployed product with large numbers of tubes. It was built to find out if tube machines could be successfully deployed in a business environment and used commercially without constant tending by skilled technicians. That was the commercial limitation - the ENIAC and UNIVAC sized machines required large numbers of on-site techs to keep the hardware going. The 603 didn't. All IBM machines at the time were rented by IBM, and rental included on-site servicing by IBM Customer Engineers in white shirts and ties. IBM had to keep the servicing cost down. IBM was only willing to deploy tubes if the operating and maintenance costs were acceptable.
The 604 was a more commercial version of the 603. The 603 looked like a very large suitcase built of black perforated metal. The 604 looked like IBM's other tabulating machines and had better access to the components. This was the first high-volume product with tubes.
None of these machines had much memory. Just a few registers. There were no good memory devices yet suitable for field deployment. They were all programmed with plugboards. Everybody involved knew they needed better programmability, and there was an effort at Columbia University in the 1930s to do that. It involved Eckert and Mauchley, who went on to design the ENIAC and the UNIVAC I. But there was no good place to put the program.
The memory problem persisted for years. Williams tubes were unreliable. Acoustic delay lines had temperature stability problems. Rotating drums could work at a commercial price point. The IBM 650 (1953) was IBM's first real business computer, with drum main memory and all-electronic computation. Knuth's first book is dedicated to the IBM 650.
This side of computing gets less attention in histories, which tend to focus on the military machines. But this was the side that shipped in volume and made real money. It came from the engineers with relays and clutches in Endicott and the accountants at Armonk, who had to develop something IBM Sales could sell to large and medium sized businesses. No need for Turing, von Neumann, etc.
The IBM 1401 (1959) ended the 600 line. At last, all-transistor, all core memory, in a reliable box that could be deployed to customer sites. It's a very strange machine, with variable-length everything, but it works well. Then, of course came the IBM System/360, which standardized architecture for decades to come.
Arguably, that is the ROM.
I suspect quite a few as other "modular" systems in the transistor era like the later IBM Standard Modular System and DEC Flip-Chips ended up with plethora of specialised modules, but I'd be interested if that growth had already begun in the tube era.
It's clever. It should reduce time for some repairs.
Both the 604 and the G3 were bit serial to save components.