The Server SoC at the heart of Data Center almost don’t care about power consumption, at the opposite of Application Processors for smartphone. If you design a server multi-core SoC, you target the highest performance, in fact a combination of high frequency and lowest possible latency, and try to pack as many CPU core and embedded cache memory in a single chip. The first limitation is the die size, the chip should exhibit a yield compatible with semiconductor economics, and the power consumption compatible with physic laws (electro-migration, voltage drop and thermal dissipation). Please note that I didn’t mention power efficiency, as I don’t think it’s really a care about for this type of SoC design.
You will appreciate how challenging is the move for Qualcomm to penetrate this data center market. In fact, Qualcomm desperately need to find new market segment out of mobile. Samsung and Apple, the top two leaders in high end smartphones, are going vertical, integrating their internally designed AP. In the lower end smartphone segment, Mediatek and Spreatrum are now reaching double digit market share. To escape from this squeeze effect, Qualcomm has to move into new application. Qualcomm is good at designing application specific processor, the company has built an impressive IP port-folio (ARM architecture license, DSP, GPU, Network-on-Chip, etc.) as well as experienced designer team and is not afraid by the most advanced technology nodes like 16FF nm or even lower.
Should Qualcomm attack an hypothetic IoT market characterized by sub $5 processor price or the very dynamic data center market? I have made some market size evaluation, grapping data from IDC or Gartner. My evaluation is that the server SoC market weight $15 billion in 2015 (99% captured by Intel), SoC shipments are in the 25 million (per year) range, leading to an ASP in the $500-600 range…
25 million units is roughly 100 time smaller than the AP market size… but ASP is 20x larger. Moreover, instead of a myriad of competitors who need to gain market share (at any price?), there is only one competitor… Ok, it’s Intel.
If this news (Google will buy server SoC to Qualcomm, if the expected performance is verified) is confirmed, and we should know it by 2/11, this could ring the bell for a very interesting fight. On my right stands Intel, literally owning this market, and designing for the highest performance at any Silicon area and power consumption expense. On my left, Qualcomm who has been cleaver enough to kick out competitors like Texas Instruments, Nvidia or STMicroelectronics from a $30-40 billion AP market within 5 years (2005-2010). Which makes this move so interesting is the fact that Qualcomm not only brings its own IP, or ARM 64-b CPU, but also comes with a completely different design culture: design for power efficiency.
What’s happen when a SoC consume so many power like it does today in data center? This incredibly high power consumption impacts cost of ownership at every stage, not only electricity bill.
- The package should have exceptional Theta jA (power dissipation) characteristic, it’s more expansive than a standard package.
- Power should be dissipated by the means of a heat-sink mounted on the package.
- The rack itself has to be cooled
- The data center has to be cooled, amount of electricity spent to run the cooling system is higher than the electricity spent to power the data center chips
Last but not least, as for any IC, the SoC performance degrades when the junction temperature goes up.
I honestly don’t know by how much it would be possible to decrease server SoC power consumption. That I know for sure is that the Qualcomm designer culture is to design for performance and power efficiency. Will they do better than Intel for server SoC? Let’s say that Qualcomm is probably one of the very few companies able to address this challenge…
From Eric Esteve from IPNEST
Also Read: Submerging the Data Center
Next Generation of Systems Design at Siemens