ECE 172 Digital Systems. Chapter 15 Turbo Boost Technology. Herbert G. Mayer, PSU Status 8/13/2018

Transcription

1 ECE 172 Digital Systems Chapter 15 Turbo Boost Technology Herbert G. Mayer, PSU Status 8/13/2018 1

2 Syllabus l Introduction l Speedup Parameters l Definitions l Turbo Boost l Turbo Boost, Actual Performance l Turbo Boost, on DP Configuration l Turbo Boost, for SPEC JBB 2005 l Bibliography 2

3 Introduction l Turbo Boost (turbo) first used on Intel Core TM i7 processors and Intel Xeon processor 5500 Series l Turbo dynamically enables some temporary performance boost l Turbo temporarily increases processor core clock in defined, discreet frequency steps (AKA bins) for the benefit of higher, but safe performance increase l Permissible, only while conditions on chip safely preserve the physical health of the CPU! l How much can core frequency raise as a function of number of active cores and of other architectural parameters? l Not to be confused with overclocking, or its opposite thermal throttling! 3

4 Introduction Key learnings for ECE students, who wish to work in future Hi Tech Processor Development: 1. Lab setup differs significantly from typical product environment 2. For performance evaluation: Take into account open vs. close chassis air flow! E.g. see page 26! 3. Consider blowing fans vs. passive heat sinks 4. Measure each performance data point repeatedly, and devise consistency policies, how to use various data measured for supposedly identical test case 5. Have objective method for: Which cases to throw out, which to keep? Which ones to measure repeatedly; why? Always use same method! 6. Whichever EE policy is adopted: use it consistently! 4

5 Introduction Core i7 introduced turbo: Intel 45 nm High-K Silicon technology, launched as High-End Desktop platform with 1 socket; later used as server with 2 sockets Frequency & Voltage Independent Interface C O R E 0 C O R E 1 C O R E 2 C O R E 3 C O R E S DRAMs DDR3 Last Level Cache Pwr & Clk IMC QPI QPI U N C O R E QPI High Level Nehalem Architecture 5

6 Speedup Parameters l Without changing any other μp design parameters, the following technologies can increase compute speed of a typical uni-processor CPU: l Faster core clock l Artificially cooling CPU, then accelerating core clock l Wider memory bus l Faster bus frequency l Larger data and I-caches l Single clock caches l Multi cache hierarchy l Faster memory technology (with lower latency) l Specialized instructions, to replace multiple ops with a single l Instructions that avoid change in flow of control, e.g. CMOV l Uniform instruction format, e.g. RISC l Generally accompanied by increased HW cost! 6

7 Definitions Enhanced Dynamic Acceleration Technology, or EDAT l Prior to Turbo Boost in Core i7, Intel s previous generation Core 2 Duo introduced EDAT l Allows one core to automatically increase operating frequency, if other cores are in sleep state or idle; i.e. is a safe boost! l First verify number of active threads and electrical and thermal parameters, before taking advantage of a clock boost; keep within product constraints l Like Turbo Boost, EDAT is a Green technology, as it provides performance on demand, while keeping power consumption low, when other processors operation is temporarily not needed 7

8 Definitions Turbo Boost l Turbo boost (AKA Turbo) is increase of clock l Turbo is distinct from overclocking & EDAT l Is natural attribute of silicon product spec. l Does not change durability of part, i.e. increased clock rate is defined attribute of silicon product! l Overclocking, by contrast, increases clock frequency by running outside parts spec.; will overheat the part; requires special cooling l Turbo technology runs processor within product specification; aims to take advantage of optional thermal headroom available during underutilized conditions, i.e. when other cores are idle 8

9 Definitions Turbo Boost Hyper-Thread Looks Like Multi-Processor Under Right Conditions 9

10 Definitions Overclocking l Overclocking is a forced increase of the natural processor clock speed beyond specification! l Is an Unnatural Act! unhealthy for life of CPU l Overclocking results in running processor outside specified limits, outside safe limits l Voids Warranty! l Overclocking is not a Green Technology as it forces increased power consumption, raising core temperature beyond safe limits l Overclocking can break a processor, so done in Las Vegas speed competitions; to run safely, requires extraordinary cooling methods 10

11 Definitions Overclocking Overclocking: A Special Distraction Meant for Performance Enthusiasts 11

12 Definitions Overclocking Not Meant for Stable Execution! 12

13 Definitions Thermal Throttling l When thermal parameters change, or when number of active cores increases, prior clock increase is reversed, saving the chip from melt-down, from self-destruction, and saving power l Thermal Throttling results in decreased performance of any such tamed μp l Assumes μp to be running in steady state of execution, acknowledges that temporary hot spots are possible l Which then results in thermal throttling to save chip from melt down! 13

14 Definitions Thermal Throttling, Cont d l Happens when typical mix of IO-bound & compute-bound execution is replaced by pure compute-bound execution, resulting in more heat generation than is safe l Similar to safety action taken in Turbo mode, the frequency is throttled, resulting in less current, less heat generation, ultimately in diminished performance l μp architect decides, which safe technology of performance boosting should be realized in silicon: EDAT, Turbo Boost, Overclocking, Thermal Throttling? 14

15 Definitions Turbo Boost, sometimes referred to as Dynamic Overclocking: 15

16 Detail on Turbo Boost 16

17 Turbo Boost l A number of dynamic parameters dictate upper limit of Turbo Boost speedup l E.g. the core s temperature, momentary frequency, the overall current, momentary power, and total number of active cores l Typical frequency step of turbo boost is MHz. For each SKU, fuse values are set during chip manufacturing, to define upper bound, by how many frequency steps maximally a core can grow safely l Parameters d-c-b-a mean: If 1 core is active, single core s frequency may increase by a bins; else if 2 cores are active, these 2 can grow by b frequency steps; etc. 17

18 Turbo Boost l Applying that encoding principle, this time starting at opposite (right) end, entry means: l For 3 or 4 cores being active, core frequency may increase by just 1 frequency step; note, frequency step AKA bin l But if only 2 cores are busy, clock speed may grow up to 4 steps, or 4 bins l And if only a single core is active, that core may grow by 8 frequency steps, amounting to 1.06 GHz of additional clock ticks (Hertz) 18

19 Turbo Boost l However, boost is reversed, if for any reason a predefined envelope of maximally allowable current or temperature is exceeded l Decrease is designed not only to save the microprocessor from thermal stress, but to save power and run more green l Similarly, as bound shows, other cores may become active, forcing the current boost rate to decrease, again to protect the processor l When Turbo Boost Technology promotes core to a higher frequency, the processor will draw more current than it would while running at nominal frequency 19

20 Turbo Boost l User incurs incremental cost for electrical power consumed in turbo mode l Such cost is minor compared to the power used by system as a whole l Users may choose to manually adjust the balance between performance and power consumption through OS power policies l Performance teams at Intel focused on workloads known to be CPU-centric, they concentrated on single- and multi-threaded workloads for turbo performance data... 20

21 Turbo Boost, Actual Performance l Then proceeded by running 3 baseline frequencies without enabling turbo l Base frequencies were 2.66 GHz, 2.8 GHz, and 2.93 GHz to simulate lower and upper bounds of the same workload l Setting affinity manually, and forcing workloads to run on a single CPU allowed maximum benefit from Turbo l Affinity: either associate a thread (SW) with a dedicated core (HW) l Or associate a thread (SW) with a hyperthread (HW) 21

22 Turbo Boost, Actual Performance Operating Frequency as a Function of the Number of Running Cores. TDP: Thermal Design Point 22

23 Turbo Boost, Actual Performance l Setting processor affinity: an application manually tells OS scheduler where to run. Restricts available cores or hyper-threads where this workload shall not run l E.g. setting Affinity = p3, tells the OS: to only run scheduler on Processor 3 l Setting Affinity = p0, p2, p3 allows an app to run on hardware thread 0, 2, or 3, not on 1! l Restricting affinity to few, even to a single core, improves performance for singly threaded SW, as thread migration and thus overhead is minimized l Can improve performance when practiced by sophisticated user! 23

24 Turbo Boost, on DP Configuration l Table next page summarizes DP Turbo setup l Used in engineering validation board, called Green City: has open top configuration -no lid l Creates different thermal conditions compared to end-user production environment in standard chassis l Each processor has individual heat sink with active fan attached, in addition to 4 external fans on the sides l If workload does not hit Turbo constraints, the core frequency can increase up to 3.33 GHz dynamically on μp used (Green City), depending on number of active cores 24

25 Turbo Boost, on DP Configuration DP Configuration Parameters for Experimental Setup 25

26 Turbo Boost, on DP Configuration DP Configuration Physical Setup 26

27 Turbo Boost, for SPEC JBB 2005 l Evaluated SPEC JBB 2005 benchmark l Quantifying impact of Simultaneous Multi- Threading (SMT) under Turbo mode l Turbo provides upside with and without SMT while best performance is achievable with SMT and Turbo for this workload l In this case, the benchmark rarely hit TDP, which is indicated by the unconstrained Turbo frequency Def: TDP is Thermal Design Power. AKA Thermal Design Point; is the maximum heat, generated by μp that the cooling system can handle 27

28 Turbo Boost, for SPEC Turbo Performance for SPEC JBB

29 Bibliography Markus Mattwandel, Herbert Mayer, et al. for Intel SW College: Architecture and Early Performance Results of Turbo Boost Technology on Intel Core TM i7 Processor and Intel Xeon Processor 5500 Series November 3, Intel website pressroom/kits/corei7/pdf/intel%c2%ae%20core %E2%84%A2%20i7_Overview.pdf Intel Core i7 Microprocessors, The Best Processor on The Planet 3. General SPEC website August, SPEC website for integer component of SPEC CPU2006: October, SPEC website for floating point component of SPEC CPU2000: 6. Intel Turbo Boost technology, technology/turboboost/ 7. Intel Turbo Boost technology 2.0: chttp:// content/www/us/en/architecture-and-technology/turbo-boost/ turbo-boost-technology.html 29