As we move to  widescale deployment, we believe that  UCS X-Series  architectural message is resonating extremely well as deployment feedback has been fantastic.

In the first post part of the blog series, we discussed how heterogeneous computing  is causing paradigm shift in computing shaping UCS X-Series  architecture. In this blog we will discuss the electromechanical aspects of the UCS X-Series architecture.

Formfactor remains constant for a life cycle of the product hence  electromechanical aspects that shape the enclosure design are  very critical in design phase.

Electro-mechanical Anchors

Some of the anchors are:

1. Socket density per 'RU".
2. Memory density
3. Backplane less IO
4. Mezzanine  options
5. Volumetric space  for  logic
6. Power footprint & delivery
7. Airflow (Measured as CFM) for cooling

Socket and memory density is very important when comparing different vendor's product and in general is an indication of a how efficiently the platform mechanical has been designed in a given "RU" envelop.   The ratio of  volumetric space required for mechanical integrity vs  logic is another important  criteria.  These criterion helped us to zero down to 7RU as chassis height &  at the same time offering more  volumetric space available for logic compared to the  similar "RU" design from industry.

Previous generation of compute platforms relied on Backplane for connectivity.  UCS X-Series, does not use backplane but direct connect IO Interconnects.  As the IO technology advances , nodes and the IO Interconnect  can be upgraded  seamlessly as the elements that needs to be changed are modular and not fixed on backplane.   As the IO Interconnect speed increases its reach decreases making it harder and harder to scale in electrical domain. UCS X-Series has been designed with hybrid connector approach that  supports  electrical IO by default and be ready for optical IO in future.  This  optical IO option is optimized for intra  chassis connectivity. Direct connect IO without a backplane helps to reduce the airflow resistance and helps to remove the heat from inlet to outlet efficiently.

Power distribution

Rack power density per "RU" is hitting  1KW and soon will go beyond that.   Majority of the existing server design uses well established 12V distribution to simplify down conversion for CPU voltages. However   as current density increases  using 12V distribution would add to the  connector costs, PCB  layer count and routing challenges.  UCS X-Series, seeing the need for next generation of  server power requirements chose to use higher voltage distribution of 54V instead of 12V. Higher voltage distribution  reduces current density by 4.5 times and ohmic losses by 20 times lower compared to 12V.  Moving from 12V to 54V DC output helps in simplifying main PSU  design and makes onboard power distribution more efficient.

Server Power Consumption

We are seeing  CPU TDP (Thermal Design Power)  increasing by 75-100W at a  2 year cadence level.  Compute nodes will soon start seeing 350W per socket and they need to be ready for 500W+ by 2024.   A  2 Socket  server with GPU and storage  requires close to  2.2KW power not accounting for any  distribution losses.   To cool this  2 Socket server, fan modules alone  will consume  around 240W ,  11 % of total power.  Factoring distribution efficiencies at each intermediate stage of conversion from AC input  we are looking at around 2.4KW power draw. So, in a RACK servers with 20 x 2RU servers ,  Fan power alone will consume 4800W !!.   Modular blade platform  like UCS X-Series  with its centralized cooling  and bigger fans,  offer much higher CFM's at a  lower power consumption.  However fan  power consumption  is indeed  becoming substantial portion of the total power budget.

Cooling

Advances in semiconductor and magnetics allows us to provide more power  in the life time of chassis. However, it is difficult to pull off a dramatic upgrade on airflow ( measured as fan CFM) as technological advances are slow & incremental.  Additionally,  cost economics dictates  use of passive heat transfer techniques  to cool the  CPU in Server. This makes defining fan CFM requirements for cooling the compute nodes a multi variable problem.

Unlike standard rack design which uses spread core CPU placement,  UCS X-Series uses shadow core  design principle complicating cooling even further.

Banks of U2/E3 storage drives with power upwards of 25W each and accelerators   on front side of the blade will  restrict air going to the CPU as well as it will pre-heat air.

UCS X-Series design approached these challenges holistically.  First and foremost  is  the  development of the state of the art  fan module delivering the class leading CFM.  The other being  the dynamic power management  coupled with  fan control algorithm that will adapt & evolve as cooling demand grows and ebbs.  Compute  nodes are designed with high and low CFM   paths channeling appropriate airflow for cooling. Additionally,  power management options provides customer with configurable knobs to optimize for fan power  or high performance mode.

Emergence of Alternate Cooling Technologies

Spot cooling of CPU/GPU at 350W is approaching limits of air-cooling  Doubling airflow results in 30% more cooling but it would add 6-8 times more fan power with  diminishing return.

Data centers are not yet ready for liquid cooling on wholesale basis.  Immersion cooling  requires complete overhaul of the RACK.  Hyperscalers will lead  early adoption cycle  and  eventually Enterprises customers will  get there but the  tipping point from air to liquid cooling is still unknown.  Air cooling is not going away as we still need to cool memory, storage  and other components which are operationally difficult  for liquid cooling. We need to collect more data and answer following critical questions before  liquid cooling becomes attractive.

• Do we really need  liquid cooling for all RACKs or only few RACKs which hosts high TDP servers.
• Is liquid cooling more for green field deployments as a means to reduce fan power/acoustics  than for high-TDP CPU/GPU enablement?
•  Any Compliance/mandates  that  targets energy reduction  by certain dates in data center?
• TCO analysis of fan power saving vs  the total cost of  liquid cooling  deployment?
• Is customer OK to spend more on fan power cooling than retrofitting the infrastructure for liquid-cooling?
• Is liquid cooling going to  help  deploy more  high TDP servers without upgrading  power to  the RACK?
  • For ex: Saving 100W per 1U  in fan power translates to  3.6KW (36x 1U server) additional available power

UCS X-Series  however does support  a hybrid model -a combination of air/liquid cooling when air cooling alone is not sufficient. Watch out for more details in upcoming blogs on liquid cooling in UCS X-Series.

In the next blog, we will elaborate on trends drove the UCS X-Series internal architecture.

Resources

UCS X-Series -The Future of Computing Blog Series -Part 1 of 3