Electrical Current and Microprocessors a Philosophy T
he diagram above was borrowed from an IBM White Paper explaining the CPU design/fabrication process known as SOI (Silicon On Insulator
PDF). The paper explains SOI technology by providing a historical perspective of the types of switches used in microprocessors. Early switches were classic bi-polar until the introduction of the CMOS-transistor (Complimentary Metal Oxide Semiconductor) which operates at a greater speed and cost less to produce. In essence a microprocessor is constructed of millions of tiny electronic switches or MOS transistors which have no moving parts; their size would preclude any "mechanical" operation and of course limit speed. In order to operate as fast as they do and consume as little power MOS transistors are fundamentally capacitance switches. These types of switches or transistors must be fully charged to operate and once charged either high or low voltage allow the switch to remain closed or open in order to conduct or prevent current flow, hence the term binary. From this point on let's consider the terms transistor and switch synonymous.
Transistors are wired together using very thin wiring known as interconnects which are constructed of copper, aluminum was the preferred material until recently. When CPU makers proclaim their processor was built using a 90nm or 65nm node (a.k.a. "process") this refers to the width of the smallest interconnect wires. Scaling in the microprocessor industry has created numerous headaches for engineers since it's never as simple as "shrinking" the entire microprocessor core and releasing it to the public. Uniformity in scaling is next to impossible with so many different materials working together. While smaller transistors operate at greater speeds and consume less energy logically it would follow with each "die shrink" power consumption and thermal output would decrease. In fact the opposite is true so long as CPU Maker's scale down similar types of architecture. Given the fact these are capacitance switches in very close proximity all maintaining a constant charge, increasing the number while decreasing the Real-Estate they share increases the potential for capacitance cross coupling
, static current discharge and of course leakage. This is where Silicon On Insulator (SOI) technology can aid in placing an insulating material (silicon on oxide) to ameliorate current leakage into the surrounding semiconductor substrata. Even with material advances the Laws of Physics simply cannot be circumvented and will exact a toll. HeatC
omplacency is a dangerous state of mind, lest we forget since the advent of the microprocessor the driving force behind the industry has been and continues to be miniaturization or scaling. The chart above taken from an EETimes
article entitled; Modeling and design techniques reduce 90 nm power
(2004) there looked to be no end in site. An entire industry has evolved around thermal management to keep pace with microprocessor thermal effects. Another challenge as a result of scaling has been more complicated current demands. This has had serious ramification on both the motherboard and Power Supply industries. Several events have conspired to bring power related issues as well as heat to the foreground. They are as follows; transition from 3.3V/5V Rails to 12V rails for CPU power, the shrink to 90nm (Prescott) eventually to 65nm (Presler) in Dual Core, changes in the graphic card industry (SLI) and the final catalyst which is the least understood yet had the greatest impact, 12V Power Supply Design Guides including EPS12V Guide 2.91
, ATX12V Guide v.2.01
, ATX12V Guide 2.2
. To simplify, these papers limit all 12V rails to 18A, with varying peaks based on a 240VA UL Labs safety considerations. EPS12V Guideline 2.91 Sec. 7.2 240VA Protection: System designs may require user access to energized areas of the system. In these cases the power supply may be required to meet regulatory 240VA energy limits for any power rail. Since the +12V rail combined power exceeds 240VA it must be divided into separate channels to meet this requirement. Each separate rail needs to be limited to less than 20A for each +12V rail. The separate +12V rails do not necessarily need to be independently regulated outputs. They can share a common power conversion stage. The +12V rail is split into four rails. Refer to section 6.4 for how the 12V rail is split between different output connectors.
As a result of the guidelines above regardless of PSU wattage, albeit 500W or 1100W that power supply's 12V-Rails cannot exceed 18A. Where does this lead us, or more importantly where does this leave Power Supply makers? Although Intel unofficially dropped the 240VA current limit 2005 power supply makers are still left in an awkward position. For the PSU maker they're "required" to meet these guidelines insofar as they're desire to advertise on their packaging they meet those guidelines. The question I've been asking for over a year now is why? Why comply with standards which are ultimately detrimental to the evolution of the modern PSU? The answer to this question is an obtuse tautology at best and in the best scenario a paradox. Perhaps the gist of the problem is there are very few companies whom are willing to challenge these standards and have the expertise to do so. The only PSU maker I know of whom has contested aspects of the Form Factor and Server System Infrastructure guidelines has been PCPower&Cooling
. The following PSU article at ExtremeTech
should shed some light on why PCPower is different and why power supplies which meet the SLI standard are still unable to power such a system. The graph below borrowed from X-Bit Labs article Grand Clash for Watts: Power Consumption of Premium Graphic Cards
may be the most accurate to date, based on the integrity of the test system Power Consumption of Contemporary Graphic Cards...(Measuring Power Consumption)
and the author Tim Tscheblockov's expertise.W
ith graphic card power levels reported
to exceed 300W we are approaching the limitations of the combined 75W PCIe slot and 6-pin external connector either of which may share or are limited by the 18A (20A Peak) 12V-Rail Guidelines. It's claimed after this peak GPU's will follow the same trail Intel has "backtracked" when they made what seemed as dramatic as a Nativistic Movement
to the days of the PIII. With respect to architecture and power efficiency GPU's today are analogous in their demands on power circuitry as the original Socket-478 Prescott was to Socket-478 motherboard power supply circuitry back then. Given the conditions discussed above restrictions on PSU circuitry have hamstrung manufacturers, many SLI ready PSU's are simply unable to power high end SLI and QUAD SLI systems. As an ad hoc
fix, several PSU manufacturers have released (AC fed) VGA Power Boosters which are basically 12V power supplies which integrate at the external 6-pin VGA connector relieving the burden on the system PSU. The pre-production samples we have from Clever Power?s
are passive in that they integrate between the 6-pin connectors and graphic card in the VGA Booster. And between the 24-pin ATX and 4-pin baseboard connector in their CPU Stabilizer, both products aid primarily in storing and smoothing current. The idea has merit once the devices are "charged" they act as a large capacitance bank supplying large amounts of current on demand which would normally stress the PSU. CapacitanceU
ntil recently supplying clean uninterrupted DC-current to the CPU and/or GPU wasn't so much a problem. As explained above an industry decision to transfer CPU source current from a shared 5V and 3.3V rails to the 12V-Rails began to complicate matters from several perspectives. Motherboard power circuitry increased from a few "phases" (Mosfett pairs) to multiple (redundant) phases in which these Mosfett's share the task of smoothing, reducing and increasing voltage to the CPU as needed. The VRM (Voltage Regulation Module) responsible for this DC-dance has now become the hottest component on your motherboard and some Mosfetts run nearly three times as hot as your CPU. With the "balance of power" now falling almost completely upon the 12V-Rail there may be a time when Power Supplies are purely 12V devices, hopefully with multiple transformers running in parallel yet each its own proprietary 12V circuitry.
Motherboard makers have gone to great lengths to lesson the load on over-taxed power circuitry on with some innovative options and in doing so thought they might find a solution that was "future proof" meeting the demands of future CPU's to be released on the same socket. Gigabyte
pioneered the use of an add-in power card named "DPS" (Dual Power Supply) which purportedly transformed their 3-phase on-board power circuitry into a 6-phase model. Unfortunately marketing claims went unsubstantiated since each 3-phase power circuit (the add-in card and on-board) has their own VRM controller. Without a single chip "regulating" voltage among all the Mosfett?s, all we actually have are redundant 3-phase systems.
Ironically DPS did aid in smoothing current while sporting its own active Mosfett cooler and bears similarities on a small scale to the products Clever Power has given us today. Later Gigabyte DPS versions even used a heat-pipe cooler which I was able to test when Gigabyte sent me the GA-8N-SLI board last year. While no PSU can compensate for inadequate motherboard power circuitry the option of choosing a single 12V-Rail PSU which is SLI certified such as PCPower&Cooling Silencer 750 EPS12V
look to be promising. Designed around a Single 12V-Rail this 750W silent PSU offers 60A on that line. Fortunately the microprocessor landscape is beginning to change so the chart below will most likely be defunct in the next few years.
Undoubtedly advances in microprocessor design including GPU's will help to remedy power and heat related issues. As I write this Conroe owners have already begun to find relief from the heat in that area. Unfortunately Conroe and Allendale cores have been touted as two of the best friends a powerful VGA card ever had, a title once attributed to AMD's FX cores. Shorter pipelines and more efficient IPC processors have placed Intel back in the good graces of those "in the microprocessor know" whom began to despise all things from Santa Clara which exploited the "need for speed" sales edict. Since the Net Burst bubble has been burst things are looking up for Intel and all they had to do was unlock the marketing department storage closet whether the PIII architectural plans were hidden for years. We can expect a similar metamorphosis in the graphic card industry, although in the interim we'll have to contend with the R600 and G80 which are reported to consume as much as 300W. Clever Power looked to those End-user's whom were experiencing all sorts of system instability and anomalies which were power-circuitry and power supply related, these are they're pre-production solutions.