Data centres and servers
Xeon, EPYC, Graviton, H100 and the chips of the cloud. 296 chips on EveryChip so far.
Data centres and cloud servers use the largest and most power-hungry chips: server CPUs with up to a few hundred cores, AI accelerators drawing 700 W to over 1,000 W each, network chips and DPUs, and huge amounts of memory. Intel's Xeon and AMD's Epyc dominate general-purpose servers, while Arm designs such as Amazon's Graviton have grown fast. Nvidia's GPUs lead in AI.
Server CPUs
A server processor is built for many tasks at once and for long, reliable running. It supports large amounts of error-correcting memory, many PCIe lanes for storage and network cards, and two or more sockets on one board. AMD's Epyc, launched in 2017, used chiplets to fit more cores than single-die designs, and took a large share of the market from Intel's Xeon.
Custom cloud chips
The largest cloud companies now design their own chips. Amazon has Graviton CPUs and Trainium and Inferentia AI chips, Google its TPU and Axion CPU, and Microsoft its Cobalt CPU and Maia accelerator. They buy fewer standard parts this way and tune chips to their own software. The power used by AI servers has become a constraint on where data centres can be built.
Famous data centres and servers chips
Still to be written: AWS Graviton.
Further reading
🇺🇸 United States3D XPoint
Non-volatile memory from Intel and Micron, sold as OptaneMemory · 2015
WorldElectrochemical RAM
Non-volatile memory for analog deep learning accelerationMemory · Other
Photo credits on this page
The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.
- DDR5 SDRAM: Jacek Halicki, CC BY-SA 4.0, source
- Epyc: Smial ( talk ), FAL, source
- Granite Rapids: Intel Corporation, Public domain, source
- DDR2 SDRAM: An-d, CC BY-SA 3.0, source
- Electrochemical RAM: Goldenvu, CC BY-SA 4.0, source
- PC133: Vitaly Zdanevich, CC0, source
- 3D XPoint: Trolomite, CC BY-SA 4.0, source
- DDR4 SDRAM: smial, FAL, source
- Intel 1103: Thomas Nguyen, CC BY-SA 4.0, source
- DDR SDRAM: CrazyD, CC BY-SA 3.0, source
- DDR3 SDRAM: Kjerish, CC BY-SA 4.0, source
- NvSRAM: Raimond Spekking, CC BY-SA 4.0, source
- Magnetoresistive RAM: Daderot, CC0, source
- Tensor Processing Unit: Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan,, CC BY 4.0, source
- High Bandwidth Memory: Fritzchens Fritz, CC0, source
- DRAM: Appaloosa, CC BY-SA 3.0, source
- DDR2 SDRAM: Raimond Spekking, CC BY-SA 4.0, source
- Hopper: 极客湾Geekerwan, CC BY 3.0, source
- IBM 2361: Bubba73 , Jud McCranie, CC BY-SA 4.0, source
- ULLtraDIMM: FlashStorage, CC BY-SA 3.0, source