EEVblog #1341 - AMAZING $250,000 IBM Processor TEARDOWN!

Wow! This brings back memories. I worked at the IBM East Fishkill plant back in the 1980's. I worked in the manufacturing technology areas where the ceramic modules were created. At its peak, the plant employed about 15,000 people, and the weekly production could be driven away in the back of two station wagons. An unbelievable amount of engineering went into the design of these modules and the manufacturing processes to make them. The ceramic green sheet material that formed each layer of the module was created by a long sheet roll process on a machine that was approximately 63 feet long. The roll was then taken to a die machine that precisely cut the dimensions of each sheet. The sheets were precisely punched for layer-to-layer vias, then fed into screening machines that spread molybdenum paste through masks to create the pattern on each layer. The next step was a stacker that aligned and stacked each of the 63 layers so that vias from layer-to-layer would align. The stack was then pressed and readied to go through the sintering process. The sintering or firing, of the ceramic substrates was done in huge sintering furnaces that had a hydrogen forming gas. Each furnace was loaded by robotic arms that moved assemblies that weighed hundreds of pounds. When the process was running, the excess hydrogen gas was burned off in a stack. It looked like something from the Wizard of Oz. Occasionally, there would be an air leak into one of these furnaces where the oxygen/hydrogen ratio would be just right to cause an explosion of the entire furnace, shutting it down for an internal rebuild. The ceramic module would then go to the thin films processing section, where it was CMP polished for flatness and the multiple thin film layers were deposited using technology adapted from wafer manufacturing. In the newest generation of TCM, there were six layers of thin film redistribution wiring on top of the ceramic, before the chips were attached. This process had to provide for corrections in lower layer yield defects, and had to achieve 100% yield over the entire substrate. There were multiple customization laser repair processes that added/subtracted single trace defects between the chip sites. Wow! I could go on, but I'm running out of steam... maybe someone else can pick up and describe the additional complexity in the manufacturing processes of this unique assembly.

I thought about taking it apart for years. So glad I did not. When I discovered Dave I immediately hoped he would enjoy it for mailbag. I learned so much more from watching Dave disassemble this that I would have doing it myself.

Wow yea the memories are flying now!! I worked at I.B.M. East Fishkill N.Y. for 31 years and for about 2 years during the late 70s to about 1981 I worked in one the many lines that made these TCMs when the TCM had 33 layers!! now about the Molybdenum , The MOLY was deposited by an electroless plating process that used 3 or 4 steps on the pads that supported the pins, the final process plated the pin pads Gold. The internal wiring was made of Moly paste that was "screened onto" each individual ceramic layer to "wire" the chips together!! The real trick was to keep all the via holes aligned for each layer to have continuity with the layer above and below itself when the TCM was being "fired" to cure the ceramic layer's that formed the TCM the via hole would misalign and cause loss of continuity in areas of the TCM due to inconsistency of the shrinkage rate of the layers of ceramic!!!!! This was a HUGE PROBLEM for a while but the engineering departments figured it all out and solved this daunting problem and done the fix!! In later years the layers went from 33 to 64 and the via problems remained solved due to changes in the ceramic mix!! And yes I seem to remember the name Knickerbocker as one of the top engineers in the MLC (Multi Layer Ceramic) program but never actually met the fellow! John A Bellas KC2UVN

I don't know why youtube recommended this video to me , but watching this guy being so happy over a processor is making me happy!

How often do you change your processor oil? Once every 2 years, or 50 billion quadrillion instructions, whichever comes first.

About the processing power: The 390 is not a scientific computer but a business machine. It was not designed to run calculations but to run data through. "Want to run a batch job over all 50 million accounts of you bank's customers? Sure, will be done in 5 minutes."-type stuff. Or to connect a couple thousand end user terminals to one system and have guaranteed response times below 1/10 second. The missing chips are due to increased load, as far as I know. Those modules were highly redundant, so the advertised capacity probably was only half of what was possible with the extra chips as backups and to compensate for production errors. I wouldn't be surprised if a third of those chips didn't work in early production. But there were enough of them for the complete module to still be functional and have enough backups for plenty of them to die later. Later when the yield got better they could leave out some chips and get the same result.

I Used to be a mainframe programmer. Really enjoyed seeing the brain that executed my code. Thanks!

wow , we came a long way just in 25 years

They are called TCM. This brings memories :). I use to know so much about them from time I have worked in Dismantling IBM facility in the Germany. My job was to take special parts from preowned machines and package them for the clients that could not afford new or refurbished parts. TCM's where most expressive parts ever. Some of them costing in millions of DM (Deutsch Mark, German money before Euro for the new kids :P ) . I have build those out of the main frames. There is whole procedure for taking them out. You need special key that unlocks the cooler block. There is a special pen like tool if pins bend. There is a special conductive plastic case with special foam inside for caring them. There was a huge call it a main board with 9 sockets. and it was massive. 3x3. Some frames have had two of those one on the top one on the bottom. There was a special tool that you screw in those to use it as a handle when taking those big boards out. IT was two man job tro carry it out of the frame. I was young and stupid and was doing ti alone until my back gave up. board with TCM's in it was 90KG. It cost me months to fix my back and it involved putting electrodes and some current true my back. First you take a hex bit and unscrew all water blocks from all of the TCM's. Then you would Pull out whole tubing and cooling blocks. then you would unscrew huge power rails from the so called main board. Then you crew in handles and unscrew the mounting screws. You take the thing out and place it on the table. Needles to say you are grounded with a wrist bracelet and you wear special anti-static shoes and on grounded table and special painted floor that can not generate static. Teh you unscrew TCM and then you use special key tool to unlock the TCM from it's socket. Turning the key was moving the TCM sideways unlocking the pins. Then you inspect it and make sure pins are all good and correct them if needed. You stick your bar code label on the side so people can trace your work.Then you place TCM in the ESD bag and then you place it in the special container. Then you lock it and seal it with special wire and lead seal with special tool. To open the container you have to cut the seal or wire off. I remember looking at part list and pricing. And some of those costed like 4,5 Million DM. I think Cheapest where like 1 million DM. Some of the TCM's where processor units having in them 11x11 CPU cores. Each core have had a little cylinder with a spring making a connection to a silicone core for direct cooling and the top block would be like heat spreader. Water block would crew on it. Imagine those pig pumps. Those have had huge electric motors few KW each. Tubing was a maze. Individual tubes where like size of a little finger and they would all collect to bigger hoses. Some TCM's where memory only and would have up to you guested it 11x11 memory chips. So every TCM have had it's function. Whole system was made of 7 main frame units/closets. One of them would have two gigantic pumps and special reservoir that decriminalizes liquid used in the cooling system. BTW it was cooled with some alcohol mixture. Now imagine 6 out of 7 frames would be all full stock with 18 TCM's and one frame would be Cooling. Yes no permanent storage!!!! Whole system needed storage solution that was again contained in more frames/huge closets. Imagine costs when Bank or some Government institution orders a system. And yes many have ordered multiply systems in the cluster working as one back then gigantic system. And yes client side units where thin clients connected with Token ring adapters or later BNC 10baseT networks and having black and green CRT monitors up to big 14" :). They have had big mechanical keyboards with heavy metal base. I was taking them out of the 3900 series. There where some systems for the storage and they have had floppy disks drive. 8" fasynchronous

You need to take it back to IBM and tell them it was like that when you bought it.

Probably the happiest I’ve seen Dave in 5 years of watching him.

Ken Shirriff — Let's take a chip die apart layer by layer with chemicals, then do a detailed circuit analysis of the photolithography and metal deposition layers. Dave Jones — Let's scratch the whole thing up with wood fibre paper towels, and rubber gloves, and then smash the dies with a huge pair of pliers...

I never worked on mainframes that used these back in my IBM New Zealand days. But they were so expensive that the IBM Maintenance Parts Centre had an aircraft on standby to fly replacement parts across the country, to maintain service level contracts. It was much more cost effective to have a small central stock and a plane to freight the parts around the country.

Some IBM engineers where just crying when you started poking at the caps

And suddenly, processor porn turned into processor gore.

5 microns flatness is only ~10 wavelengths of visible light. That's meh for an optical flat made out of glass, but absolutely incredible for what is effectively a multilayer ceramic circuit board on steroids.

I worked at IBM in Poughkeepsie in 2000 and they were still making these large server modules. This is when they were transitioning to all copper interconnect. The ones I remember were liquid cooled and a refrigerator sized machine had rows of them. The AC unit that cooled it was larger than the computer! They had these huge shake tables that would test these systems for sensitivity to vibrations in another building.

And right now, i bet ASRock is trying to come up with a microATX board for it

MIPS, in case anyone is wondering , stands for "Meaningless Indication of Processor Speed" :-)

9:00 This is borderline porn territory. There's slugs going in and out of a hole. And copious amounts of lube