Patent Application
20230230852
Cold plates offer an efficient way for processors to be cooled during operation, which can extend the lifetime of high-value integrated circuit chips. Processors such as those used in server racks located in large data centers can produce 600 W of heat during operation. Given the close spacing of processors in a server rack, as well as spacing between server racks, air cooling is not an efficient solution at scale, especially as the industry moves to more artificial intelligence and machine learning capability, which requires more processing power than other uses. Traditional direct-to-chip cooling plates consist of a copper plate with straight thin fins machined into the surface using, for example, a CNC or skiving machine. Such a structure can generally achieve a thermal resistivity of approximately 0.05° C./W at a flow rate of 1 L/min.
A cold plate includes a metal base plate and a plurality of fins on the metal base plate. The fins are porous and comprise an additively manufacturable material.
A cold plate cooling system includes a metal plate, a cold plate on the metal plate, and a manifold on the metal plate and over the cold plate. The manifold includes at least one first port for receiving a coolant into the manifold and at least one second port for exiting the coolant from the manifold. The cold plate includes a metal base plate and a plurality of fins or pins on the metal base plate, where the fins or pins are porous and comprise an additively manufacturable material.
A first method for making a cold plate includes providing a metal base plate and 3D printing a plurality of porous fins or pins on the metal base plate.
A second method for making a cold plate includes providing a metal base plate, 3D printing a plurality of porous fins or pins, and adhering the fins or pins to the metal base plate.
Embodiments of the present invention include a method to create a finned structure on a copper plate using 3D printing via binder jetting. This 3D printed metal structure comprises, for example, straight or curved fins of copper or a copper-silver alloy to be used for single phase direct-to-chip cooling for electronic applications. Binder jetting offers several advantages over traditional milling including potentially higher throughput during manufacturing and a final structure that is porous and has a high surface roughness, thus increasing the total surface area in contact with the cooling fluid.
In some embodiments, a non-transitory machine-readable medium is used in additive manufacturing of cold plates according to at least some embodiments. Data is typically stored on the machine-readable medium. The data represents a three-dimensional model of a cold plate or a series of two dimensional models, which when layered on top of one another comprise a three-dimensional model, which can be accessed by at least one computer processor interfacing with additive manufacturing equipment (e.g., a 3D printer, a manufacturing device, or other such devices). The data is used to cause the additive manufacturing equipment to create the cold plate. As used herein, the term “three-dimensional model” refers to both one model having three dimensions and two or more models each having two dimensions, which stacked on top of each other provide a three-dimensional model.
Data representing a cold plate can be generated using computer modeling, such as computer aided design (CAD) data. Image data representing the cold plate design can be exported in STL format, or in any other suitable computer processable format, to the additive manufacturing equipment. Scanning methods to scan a three-dimensional object may also be used to create the data representing the cold plate. One exemplary technique for acquiring the data is digital scanning Any other suitable scanning technique can be used for scanning an article, including X-ray radiography, laser scanning, computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound imaging. Other possible scanning methods are described, for example, in U.S. Patent Application Publication No. 2007/0031791. The initial digital data set, which may include both raw data from scanning operations and data representing articles derived from the raw data, can be processed to segment a cold plate design from any surrounding structures (e.g., a support for the cold plate).
Machine-readable media can be provided as part of a computing device. The computing device may have one or more processors, volatile memory (RAM), a device for reading machine-readable media, and input/output devices, such as a display, a keyboard, and a pointing device. Further, a computing device may also include other software, firmware, or combinations thereof, such as an operating system and other application software. A computing device can be, for example, a workstation, a laptop, a personal digital assistant (PDA), a server, a mainframe or any other general-purpose or application-specific computing device. A computing device may read executable software instructions from a computer-readable medium such as a hard drive, a CD-ROM, or a computer memory, or may receive instructions from another source logically connected to computer, such as another networked computer.
A flat plate of 101 Copper 12″×48″, ⅛″ was machined to 45×34×3 mm. Mounting holes were drilled at the four corners and a thermocouple slot was milled into one side. A flatness requirement was specifically called out to the extent needed or desired to compensate for warpage of the plate during machining.
Copper Powder (D90<20 μm) was plasma coated with silver as described in U.S. Pat. No. 7,727,931, col. 13 line 40—col. 14 line 39 with the gold target replaced with silver, to create a thin (˜20 nm) nonuniform coating to produce copper-silver powder. The powder was loaded into an Mlab binder jet 3D printer from ExOne (North Huntingdon, Pa.), and the printer was prepared for printing following a standard manufacturer recommended startup process. The printer used ExOne's Aqueous Binder.
The machined Copper Plate was inserted into a 3D printed adaptor in order to fit snuggly into the 50×70 mm build side of the printer and to make sure the finned structure would be centered on the plate. A CAD file containing a set of fins that were 1 mm thick and 0.6 mm tall was loaded into the printer software. Printing was carried out using the parameters listed in Table 1. The first layer was printed directly onto the copper plate, with each subsequent layer thus being adhered to the plate as well. The preparation and printing process is shown in
Following the printing step, the copper plate was carefully lifted out of the printer without disturbing either the printed finned structure or the powder surrounding it, and heat treated to remove the majority of the water content in the aqueous binder.
The heat treatment took place in a hydrogen atmosphere 1200 series furnace from CM furnaces. The following heat treatment cycle was used:
1. 20 minutes purge at room temperature with 100% nitrogen at a flow rate of 80 SCFH.
2. Gas switched to 100% hydrogen at a flow rate of 10 SCFH.
3. Heat to 195° C. at a rate of 5° C./min.
4. Hold at 195° C. for 2 hours.
5. Cool to 80° C. at a rate of 5° C./min (see below).
6. Gas switched to 100% nitrogen at a flow rate of 80 SCFH.
7. Purge with nitrogen for 20 minutes.
While a nominal cooling rate of 5° C./min was programmed into the furnace, the furnace did not have the means to cool itself down that quickly. The program instead allowed the furnace to cool as quickly as possible, with a 10° C. holdback from the programmed rate.
Following the heat treatment process, the loose powder was manually brushed off the machined Copper Plate, leaving behind just the printed structure. Small remaining amounts of loose powder were blown off with a low-pressure air hose.
The cleaned “green” part was then subjected to a second heat treatment (in the same furnace) to sinter the structure together. The following heat treatment cycle was used:
1. 20 minutes purge at room temperature with 100% nitrogen at a flow rate of 80 SCFH.
2. Gas switched to 100% hydrogen at a flow rate of 10 SCFH.
3. Heat to 500° C. at a rate of 5° C./min.
4. Hold at 500° C. for 1 hour.
5. Heat to 900° C. at a rate of 5° C./min.
6. Hold at 900° C. for 10 hours.
7. Cool to 100° C. at a rate of 5° C./min (see below).
8. Gas switched to 100% nitrogen at a flow rate of 80 SCFH.
9. Purge with nitrogen for 20 minutes.
While a nominal cooling rate of 5° C./min was programmed into the furnace, the furnace did not have the means to cool itself down that quickly. The program instead allowed the furnace to cool as quickly as possible, with a 10° C. holdback from the programmed rate.
Stand-alone finned structures with a base were printed using the same printing parameters outlined in EXAMPLE 1.
Instead of using a copper plate as the base for the first layer, the printed parts were suspended in loose powder during and after the print. The printed structures consisted of straight fins and a square base 1 mm thick. The printed structures were oversized 27% in the x-y plane in order to compensate for shrinkage during sintering.
The parts were printed and heat treated using the parameters described in EXAMPLE 1, with one exception: the hold time at 900° C. during the second heat treatment was reduced from 10 hours to 4 hours.
Following the second heat treatment process, the finned structures were brazed onto clean copper plates (machined according to EXAMPLE 1). Silverbraze 45 brazing material was used in two different forms: 1) thin foil; 2) paste/slurry.
The braze materials were applied to the machined Copper Plates (
The stack of machined Copper Plate, braze material, and printed finned structure were heat treated to form a plate/braze/printed structure assembly according to the following program (same furnace as used in EXAMPLE 1):
1. 20 minutes purge at room temperature with 100% nitrogen at a flow rate of 80 SCFH.
2. Gas switched to 100% hydrogen at a flow rate of 10 SCFH.
3. Heat to 635° C. at a rate of 5° C./min.
4. Hold at 635° C. for 30 minutes.
5. Heat to 780° C. at a rate of 5° C./min.
6. Hold at 780° C. for 10 minutes.
7. Cool to 100° C. at a rate of 5° C./min (see below).
8. Gas switched to 100% nitrogen at a flow rate of 80 SCFH.
9. Purge with nitrogen for 20 minutes.
While a nominal cooling rate of 5° C./min was programmed into the furnace, the furnace did not have the means to cool itself down that quickly. The program instead allowed the furnace to cool as quickly as possible, with a 10° C. holdback from the programmed rate.
The hold temperatures were chosen according to the solidus and liquidus temperatures of the braze material (655° C. and 745° C., respectively, as reported by the supplier).
The steps from EXAMPLE 1 were repeated with the following two changes:
A custom benchtop setup was used to measure the thermal properties of the cold plates. The setup replicates the thermal environment of a processor, while allowing for more granular control and better repeatability across tests than a real processor.
The test setup comprised:
Heating convection from the microstructures to the fluid was calculated from the difference in temperatures between the thermocouple mounted closest to the bottom of the cold plate and the inlet water temperature:
where the heating flux density φ was equal to power divided by cross-sectional area of the pedestal. Power through the pedestal was deduced from the temperature difference of the thermocouples embedded within the pedestal, such as x1 and x3 in
Table 2 provides the thermal resistivities measured for the Examples described above.
Computational Fluid Dynamics (CFD) model was used to simulate thermal resistance of microchannel cold plate at different flow rate under the same condition as experimental setup. The Finite Volume Method (FVM) was employed to compute thermal and flow field in the computational domain.
The model was validated by fabricating and testing thermal performance of three different cold plates. Thermal resistance data was collected at flow rates between 200 and 1400 mLPM and compared to simulation results.
The first sample was a flat copper plate no microchannels.
Second and third samples had microchannels fabricated with Electro Discharge Machining (EDM) method to generate the microchannels, the channel width of which are 152 um and 203 um, respectively.
This computation demonstrates simulation predicting heat sink performance within 3% of experimentally measured thermal impedance at flow rate of 1000 mLPM. Thus, the simulation provided the effective means of performance prediction with high confidence.
Table 3 compares the performance of heat sink manufactured using the method of EXAMPLE 3 to predicted performance of machined-copper heat sinks with the same (EXAMPLE 4) and 2× smaller (EXAMPLE 5).
The thermal performance of the sample EXAMPLE 3 was superior (thermal resistance was ×2.4 smaller) than in a comparable machined-copper sample. Its thermal resistance was even significantly smaller than the resistance of a machined-copper heat sink with channel width of 250 μm. This was an unexpected result with the improvement achieved by using binder jetting to fabricate the heat sink channels.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/IB2021/056185 | 7/9/2021 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 63049745 | Jul 2020 | US |
