AMD, profile picture
Uploaded byAMD
PPTX, PDF4,335 views

Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core

The document presents an overview of AMD's Zen 2 microprocessor architecture, detailing improvements over its predecessor, including performance enhancements, increased core counts, and power efficiency achieved through various technological innovations. It highlights specific changes in architecture, core complex configurations, and cache hierarchy, as well as the challenges and solutions involved in the design and implementation process. The analysis culminates in demonstrating substantial gains in single-thread performance and reductions in power consumption relative to the previous generation.

Embed presentation

Downloaded 172 times
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 1 of 33 Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core T. Singh1, S. Rangarajan1, D. John1, R. Schreiber1, S. Oliver1, R. Seahra2, A. Schaefer1 1AMD, Austin, TX, 2AMD, Markham, ON, Canada Presented at ISSCC 2020
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 2 of 33 Outline • Motivation • Market Segments • Architecture • Core Complex • Technology • Implementation • SRAMs • Power • Silicon Results • Conclusion
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 3 of 33 Motivation • Zen was a huge lift • Zen2 compelling successor to Zen • Goals – Give above industry trend generational performance improvement – Enable 2x cores same socket – Improve single thread (1T) performance • How can we do this? – Technology port – Architectural changes – Physical design and methodology changes • AMD was aggressive and we did all of the above to achieve the goals!!
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 4 of 33 Zen 2 Market Segments
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 5 of 33 Zen 2 Architecture • Changes from Zen – New TAGE Branch Predictor – Optimized L1 Instruction Cache: 32K/8-way vs. 64K/4-way – 2X Op Cache Capacity: 4K vs. 2K ops – 2X Floating Point Data Path Width: 256b vs 128b – 3rd Address Generation Unit – Larger Physical Structures: Integer Scheduler, PRF, ROB, Store Queue, L2DTLB – 2X L1 Data Cache Read/Write Bandwidth – 2X L3 Cache: 16MB vs. 8MB per Core Complex (CCX) • +15%1 single thread (1T) IPC over Zen • ~9% switching capacitance (CAC) improvement over previous generation, technology neutral 1 AMD "Zen 2" CPU-based system scored an estimated 15% higher than previous generation AMD “Zen” based system using estimated SPECint®_base2006 results. SPEC and SPECint are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org.
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 6 of 33 Core Functional Units • 32KB IC • 32KB DC • ~20 blocks, ~400K avg instances • ROM for uCODE • 5 L1 RAM variants • Chip Pervasive Logic (CPL) – clock/test block Floating Point Data Cache Load/ Store ALU Scheduler Branch Prediction I-CacheDecode L2 Cache uCode CPL
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 7 of 33 L2/L3 Cache Hierarchy • Only 3 unique custom macros – Down from 8 on Zen • Each 4M slice is identical • Multi-stage clock gating in L3 to keep clock distribution power the same as 8M L3 from Zen • LDOs incorporated into the L3 to supply VDDM to L2 and L3 arrays – Loss of package distribution of VDDM meant LDOs had to be moved closer – Must reduce current on VDDM CTLL3Tags L3Data 4M Slice L2 Data L2 Tags L2 Status Shadow tag macros for serving external probes 512K L2 LDOs
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 8 of 33 Zen 2 Core Complex (CCX) • 4 core complex • L3 size increases to 16MB • Design for flexibility • Maximize # cores for server case
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 9 of 33 Zen 2 CCX Configs HEDT/Server 4 Core, 16MB L3 CCX APU 4 Core, 4MB L3 CCX Value 2 Core, 4MB L3 CCX • Zen 2 Core can be used in various configs covering a wide power range • Multiple CCX can be placed to achieved desired core count Cores Market TDP 8 Notebook 15W 6 Desktop 65 W 8 Desktop/Server 65-120 W 12 Desktop/Server 105-120 W 16 Desktop/Server 105-155 W 24 HEDT/Server 155-280 W 32 HEDT/Server 155-280 W 48 Server 200-225 W 64 HEDT/Server 200-280 W
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 10 of 33 Zen vs. Zen 2 Technology Comparison Zen Zen 2 Tech 14nm FinFET 7nm FinFET Cores/CCX 4 Cores, 8 Threads 4 Cores, 8 Threads Area/CCX 44 mm2 31.3 mm2 L2/core 512KB 512KB L3/CCX 8MB 16MB CPP 78 nm 57 nm Fin Pitch 48 nm 30 nm 1x Metal Pitch 64 nm 57 nm Stdcell Track Library 10.5 track 6 track Cu Metal Layers 11 w/ MiM 13 w/ MiM
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 11 of 33 Zen vs. Zen 2 Technology Comparison (cont) Zen (14nm) Zen 2 (7nm) Layer Name Pitch Layer Name Pitch M0 StdCell Internal n/a M0 StdCell Internal 1.0x M1 StdCell Internal 1.0x M1 Stdcell & BEOL 1.425x M2-M3 1.0x M2-M3 1.0x-1.1x M4-M7 1.25x M4-M7 2.0x M8-M9 2.0x M8-M9 2.0x --- --- M10-M11 3.15x M10-M11 (RDL) 11.25x M12-M13 (RDL) 18.0x
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 12 of 33 Place and Route Design Optimization • 7nm FinFET presents unique route challenges – Lower layer jogs forbidden – Denser standard cells with reduction in track height – Increased lower level metal resistance • Deep collaboration between AMD CAD, foundry, and EDA partners – Cell density management – Advanced legalization techniques – Improved pre-route timing estimates – Wire Engineering and Via Ladders Same-Layer Jogs Forbidden Inter-Layer Jumpers Required
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 13 of 33 Placement Restricted by Large Cells • Multi-row cells benefit power and area, but create placement challenges • Clustering of flops has many benefits but can cause placement issues • Resulting small gaps are challenging to use and required innovation to exploit – New algorithms – Flexible power grid choices
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 14 of 33 Design RC Miscorrelation • Pre-route vs Post-route miscorrelation caused by length and layer assumptions • Pre-route miscorrelations for resistance and capacitance have differing root causes – Layer assignment for resistance – Length estimates for resistance and capacitance • Based on previously modeled trends, EDA tools may have challenges estimating delay • Required innovation to tackle Layer Normalized Resistance Normalized Capacitance M1 1.00 1.00 M2 3.17 0.96 M3 2.31 0.96 M4 0.72 0.75 M5 0.55 0.83 M6 0.52 0.83 M7 0.55 0.83 M8 0.52 0.83 M9 0.55 0.92 M10 0.16 0.96 M11 0.16 0.92
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 15 of 33 Pre-Route Correlation Improvements • Plots showing ClockTreeSynthesis vs Route timing • Large variance in initial results – Large number of paths have overly- pessimistic delay during pre-route steps. Tools waste resources trying to fix – Significant number of paths have optimistic delay estimates. These paths are under- optimized • Employed timing with targeted capacitance scaling and global route- based layer estimation – Standard deviation dramatically improved while keeping a slightly pessimistic mean cts_vs_route.slack.corr cts_vs_route.slack_delta.hist cts_vs_route.slack.corr cts_vs_route.slack_delta.hist Timing Slack Correlation Timing Slack Delta Initial Results Improved Results Pessimistic Optimistic
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 16 of 33 Wire Engineering Challenges • Lower layers getting more resistive with latest technology nodes – Very short routes in tight data paths need a buffer – Routes longer than Steiner due to complex rules – Challenging for optimization tools to comprehend • Critical signals need to get to higher layers quickly
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 17 of 33 Wire Engineering and Via Ladders • Team used selective layer optimization, buffering, pre-routes, and via ladders to exploit the fast layers for critical signals • Two types of via ladders – High Performance: for large buffers driving long wires – EM: for high-activity gates (e.g., clock drivers) – Mitigated EM issues on large fanout nodes with high activity Top Via Ladder View Side Via Ladder View
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 18 of 33 L2/L3 Cache Changes • Zen had an on-die LDO to generate VDDM supply for use by cache arrays • Zen 2’s package choices make using package layers for VDDM distribution impossible • Moved the bitline precharge from VDDM to VDD to reduce current VDDM VDDM BLT[] BLC[] WRCS[] RDCSX[] SAPCX SAEN WDT_X WDC_X XCENX SAT SAC SAC_INT SAT_INT BLPCX WL[N:0] BLT[] BLC[] WRCS[] RDCSX[] SAPCX SAEN XCENX SAT SAC SAT SAC BLPCX WL[N:0] NegBL Write DriverWDT_X WDC_X
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 19 of 33 • Moving bitline precharge to VDD creates both bitcell stability and writeability challenges • High level of configurability allows for silicon flexibility VDD Precharge Challenges SRAMSRAMSRAMSRAMs WLUdEn NegBlEn Assist configurations Fuses Assist controller System Management Programming details superVminEn superVmaxEn Voltage thresholds superVmaxEn=1superVminEn=1 VDDM VDDmax VDDmin VDD Controller pauses voltage increase and unsets superVminEn register before continuing to raise voltage Controller pauses voltage increase and sets superVmaxEn register before continuing to raise voltage VDD where VDDM-VDD=superVminThreshold VDD where VDD-VDDM=superVmaxThreshold
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 20 of 33 • Moving precharge to VDD reduced our current enough to allow on die-distribution but presents other challenges • Read before write timing challenges at low VDD, high VDDM VDD Precharge Timing Challenges WL@ constant VDDM BLPCX @ high VDD BLPCX @ low VDD WL on before Bitline precharge turns off at low VDD! Bitline precharge turns on before WL turns off at high VDD! Power races with WL Read before write challenges
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 21 of 33 • Solving these multiple voltage timing challenges required a number of techniques – Dual voltage clock shapers to average two voltage domains • Can alter the number of these buffers on VDD or VDDM or remove them entirely to make timing more or less dependent on either supply – False read before write problem can be mitigated by compressing the front end of the WL during a write operation Solving Timing Challenges ISOX@VDDM Input@VDD shapedFallInput @VDD VDDM LS LS VDD Psuedo-dynamic level shifter WREN WLCLK WLCLK_shape WLCLK WL during read WL during write
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 22 of 33 CAC Comparison • 3% decrease in flop power allocates more budget for combinational logic
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 23 of 33 FLOP Palette Improvements • Rich flop library, balance timing/power needs by driving right flop mix • Up to 8% Fmax benefit from high speed flops in timing critical loop paths Best for Performance Best for Power
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 24 of 33 Low Power Gater Latch Energy with AvgApp Activity (fJ) State LP Latch Regular Latch Ratio E=1 0.22 0.18 121% E=0 0.17 1.61 10% Total 0.38 1.79 22% E TE CLKB CLKBB CLKBBCLKB CLKBB CLKB qf_x qf Q Dbar Dbar qf CLK • 90% Power savings in latch for common case of E = 0 through internal self gating • Clock gater latch power contribution from 22% in Zen to 13% in Zen 2 for an average application
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 25 of 33 Zen 2 Clock Optimization • Multi mesh plan for the core supported by configurable clock tree construction – FP level mesh gating enabled with minimal timing/area overhead – 15% Mesh power savings in Idle and Average App • Tight clock skew distribution • Relocated clock spines and technology shrink (vs. Zen) achieves similar skew profile while reducing CAC
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 26 of 33 Zen vs. Zen 2 CAC Comparison • Primary sources of CAC reduction – 14 nm to 7 nm scaling – 6 track library – Aggressive microarchitectural CAC optimizations
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 27 of 33 Generational Leadership Perf/Watt • Performance/Watt driven by a combination of technology and design improvements • Timing – Improved scalability by optimizing at a wider voltage range compared to Zen – Multi-corner optimization • Library choice and optimization – 6 track library enabled additional CAC/leakage savings in addition to default technology entitlement • Design CAC – MBFF, low power clock-gater library optimization – RTL improvements – CAC aware downsizing methodology Zen power @ 100% IPC Zen2 power @ 115% IPC 7nm CAC Savings 7nm Timing Design CAC Savings Library Choice Power Improvements – ISO Frequency
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 28 of 33 Frequency/Power Silicon Results • 4 cores active with 2 threads per core • The combined effect of lower Vmin for the same frequency and reduced CAC enabled a 50% reduction in power for a given frequency throughout most of the F(P) curve • This enables 2x cores in the same socket!! 50% power reduction
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 29 of 33 Frequency/Voltage Silicon Results • 1 core active with two threads per core, 3 cores idle • F/V curve improved over all voltages • Design worked to improve the low voltage performance for improved linearity • Wide voltage range
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 30 of 33 Conclusion • Met Goals – Moved to energy efficient TSMC 7nm finFET – Made huge architectural changes – Improved PD and methodology • Results are clear – Scalable across 15W mobile to 280W Server – 50% reduced power at iso-frequency – Enable 2x cores in same-socket – >15% 1T IPC over previous generation – ~9% CAC improvement over previous generation technology neutral – Enables peak frequencies up to 4.7GHz (+350MHz generationally) • Zen2 delivers generational performance uplift!!
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 31 of 33 Acknowledgements • We would like to thank our talented AMD design team across Austin, Fort Collins, Santa Clara, Boston, Markham, and India who contributed to Zen 2 • Please stay for our chiplet paper next
2.2 : AMD Chiplet Architecture for High-Performance Server and Desktop Products © 2020 IEEE International Solid-State Circuits Conference 32 of 33 Disclaimer and Endnotes DISCLAIMER The information contained herein is for informational purposes only, and is subject to change without notice. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein. No license, including implied or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations applicable to the purchase or use of AMD’s products are as set forth in a signed agreement between the parties or in AMD's Standard Terms and Conditions of Sale. GD-18 All rights reserved. AMD, the AMD Arrow logo, EPYC, RYZEN, Threadripper, Radeon, Infinity Fabric, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.
2.1: Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 33 of 33

More Related Content

PPTX
3D V-Cache
byAMD
 
PPTX
AMD Chiplet Architecture for High-Performance Server and Desktop Products
byAMD
 
PDF
AMD Zen 5 Architecture Deep Dive from Tech Day
byLow Hong Chuan
 
PDF
ISSCC 2018: "Zeppelin": an SoC for Multi-chip Architectures
byAMD
 
PDF
The Path to "Zen 2"
byAMD
 
PDF
AMD and the new “Zen” High Performance x86 Core at Hot Chips 28
byAMD
 
PPTX
“Zen 3”: AMD 2nd Generation 7nm x86-64 Microprocessor Core
byAMD
 
PPTX
Hot Chips: AMD Next Gen 7nm Ryzen 4000 APU
byAMD
 
3D V-Cache
byAMD
 
AMD Chiplet Architecture for High-Performance Server and Desktop Products
byAMD
 
AMD Zen 5 Architecture Deep Dive from Tech Day
byLow Hong Chuan
 
ISSCC 2018: "Zeppelin": an SoC for Multi-chip Architectures
byAMD
 
The Path to "Zen 2"
byAMD
 
AMD and the new “Zen” High Performance x86 Core at Hot Chips 28
byAMD
 
“Zen 3”: AMD 2nd Generation 7nm x86-64 Microprocessor Core
byAMD
 
Hot Chips: AMD Next Gen 7nm Ryzen 4000 APU
byAMD
 

What's hot

PDF
AMD EPYC™ Microprocessor Architecture
byAMD
 
PPTX
AMD Radeon™ RX 5700 Series 7nm Energy-Efficient High-Performance GPUs
byAMD
 
PDF
AMD: Where Gaming Begins
byAMD
 
PPTX
Hot Chips: AMD Next Gen 7nm Ryzen 4000 APU
byAMD
 
PDF
Delivering a new level of visual performance in an SoC AMD "Raven Ridge" APU
byAMD
 
PDF
Delivering the Future of High-Performance Computing
byAMD
 
PDF
AMD Ryzen CPU Zen Cores Architecture
byLow Hong Chuan
 
PDF
7nm "Navi" GPU - A GPU Built For Performance
byAMD
 
PDF
Deep learning: Hardware Landscape
byGrigory Sapunov
 
PDF
If AMD Adopted OMI in their EPYC Architecture
byAllan Cantle
 
PPTX
03_03_Implementing_PCIe_ATS_in_ARM-based_SoCs_Final
byGopi Krishnamurthy
 
PDF
Evaluating GPU programming Models for the LUMI Supercomputer
byGeorge Markomanolis
 
PDF
Shared Memory Centric Computing with CXL & OMI
byAllan Cantle
 
PDF
FPGA Hardware Accelerator for Machine Learning
byDr. Swaminathan Kathirvel
 
PPT
PCB
byPranaviya
 
PPTX
Q1 Memory Fabric Forum: Using CXL with AI Applications - Steve Scargall.pptx
byMemory Fabric Forum
 
PDF
Zynq ultrascale
byel10namaste
 
PDF
DDR, GDDR, HBM SDRAM Memory
bySubhajit Sahu
 
PDF
Network Programming: Data Plane Development Kit (DPDK)
byAndriy Berestovskyy
 
PPSX
10. GPU - Video Card (Display, Graphics, VGA)
byAkhila Dakshina
 
AMD EPYC™ Microprocessor Architecture
byAMD
 
AMD Radeon™ RX 5700 Series 7nm Energy-Efficient High-Performance GPUs
byAMD
 
AMD: Where Gaming Begins
byAMD
 
Hot Chips: AMD Next Gen 7nm Ryzen 4000 APU
byAMD
 
Delivering a new level of visual performance in an SoC AMD "Raven Ridge" APU
byAMD
 
Delivering the Future of High-Performance Computing
byAMD
 
AMD Ryzen CPU Zen Cores Architecture
byLow Hong Chuan
 
7nm "Navi" GPU - A GPU Built For Performance
byAMD
 
Deep learning: Hardware Landscape
byGrigory Sapunov
 
If AMD Adopted OMI in their EPYC Architecture
byAllan Cantle
 
03_03_Implementing_PCIe_ATS_in_ARM-based_SoCs_Final
byGopi Krishnamurthy
 
Evaluating GPU programming Models for the LUMI Supercomputer
byGeorge Markomanolis
 
Shared Memory Centric Computing with CXL & OMI
byAllan Cantle
 
FPGA Hardware Accelerator for Machine Learning
byDr. Swaminathan Kathirvel
 
PCB
byPranaviya
 
Q1 Memory Fabric Forum: Using CXL with AI Applications - Steve Scargall.pptx
byMemory Fabric Forum
 
Zynq ultrascale
byel10namaste
 
DDR, GDDR, HBM SDRAM Memory
bySubhajit Sahu
 
Network Programming: Data Plane Development Kit (DPDK)
byAndriy Berestovskyy
 
10. GPU - Video Card (Display, Graphics, VGA)
byAkhila Dakshina
 

Similar to Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core

PDF
AMD Zen 2 Core Microarchitecture Overview
byRobert Hallock
 
PDF
[IGC2018] AMD Don Woligroski - WHY Ryzen
by강 민우
 
PPTX
AMD & Intel Latest Processor
byABU TALEB
 
PPT
Intel Core i7
byMd Ajmat
 
DOCX
The end of the line for single-chip processors_.docx
byEthan145044
 
PDF
Amd ces tech day 2018 lisa su
byTeddy Kuo
 
PPTX
corei7anaghvjfinal-130316054830-.pptx
byPranita602627
 
PDF
AMD Ryzen Threadripper in Techday
byLow Hong Chuan
 
PPT
SDC Server Sao Jose
byRoberto Brandao
 
PPT
Core i71
byisha_rana
 
PDF
Amd Barcelona Presentation Slideshare
byDon Scansen
 
PPTX
Intel i7
byJustin k.
 
PDF
Lakefield: Hybrid Cores in 3D Package
byinside-BigData.com
 
PDF
Intel 8th Core G Series with Radeon Vega M
byLow Hong Chuan
 
PDF
IDF'16 San Francisco - Overclocking Session
byHWBOT
 
PPTX
Core i 7 processor
bySumit Biswas
 
PPTX
intel core i7
byAleem Pasha
 
PDF
AMD 2014 computex
byLow Hong Chuan
 
PPTX
AMD-Ryzen-Processors-Architecture-Performance-and-Security.pptx
byaltwajreBaseem1
 
PPTX
Azure Local AMD Solutions Offering 20250131.pptx
byUrielrodzmig
 
AMD Zen 2 Core Microarchitecture Overview
byRobert Hallock
 
[IGC2018] AMD Don Woligroski - WHY Ryzen
by강 민우
 
AMD & Intel Latest Processor
byABU TALEB
 
Intel Core i7
byMd Ajmat
 
The end of the line for single-chip processors_.docx
byEthan145044
 
Amd ces tech day 2018 lisa su
byTeddy Kuo
 
corei7anaghvjfinal-130316054830-.pptx
byPranita602627
 
AMD Ryzen Threadripper in Techday
byLow Hong Chuan
 
SDC Server Sao Jose
byRoberto Brandao
 
Core i71
byisha_rana
 
Amd Barcelona Presentation Slideshare
byDon Scansen
 
Intel i7
byJustin k.
 
Lakefield: Hybrid Cores in 3D Package
byinside-BigData.com
 
Intel 8th Core G Series with Radeon Vega M
byLow Hong Chuan
 
IDF'16 San Francisco - Overclocking Session
byHWBOT
 
Core i 7 processor
bySumit Biswas
 
intel core i7
byAleem Pasha
 
AMD 2014 computex
byLow Hong Chuan
 
AMD-Ryzen-Processors-Architecture-Performance-and-Security.pptx
byaltwajreBaseem1
 
Azure Local AMD Solutions Offering 20250131.pptx
byUrielrodzmig
 

More from AMD

PPTX
Heterogeneous Integration with 3D Packaging
byAMD
 
PPTX
AMD EPYC Family World Record Performance Summary Mar 2022
byAMD
 
PPTX
AMD EPYC Family of Processors World Record
byAMD
 
PPTX
AMD EPYC Family of Processors World Record
byAMD
 
PPTX
AMD EPYC World Records
byAMD
 
PPTX
AMD EPYC 7002 World Records
byAMD
 
PPTX
AMD EPYC 7002 World Records
byAMD
 
PPTX
AMD EPYC 100 World Records and Counting
byAMD
 
PPTX
AMD EPYC 7002 Launch World Records
byAMD
 
PPTX
AMD Next Horizon
byAMD
 
PPTX
AMD Next Horizon
byAMD
 
PDF
AMD Next Horizon
byAMD
 
PDF
Race to Reality: The Next Billion-People Market Opportunity
byAMD
 
PDF
GPU Compute in Medical and Print Imaging
byAMD
 
PPTX
Enabling ARM® Server Technology for the Datacenter
byAMD
 
PPTX
Lessons From MineCraft: Building the Right SMB Network
byAMD
 
Heterogeneous Integration with 3D Packaging
byAMD
 
AMD EPYC Family World Record Performance Summary Mar 2022
byAMD
 
AMD EPYC Family of Processors World Record
byAMD
 
AMD EPYC Family of Processors World Record
byAMD
 
AMD EPYC World Records
byAMD
 
AMD EPYC 7002 World Records
byAMD
 
AMD EPYC 7002 World Records
byAMD
 
AMD EPYC 100 World Records and Counting
byAMD
 
AMD EPYC 7002 Launch World Records
byAMD
 
AMD Next Horizon
byAMD
 
AMD Next Horizon
byAMD
 
AMD Next Horizon
byAMD
 
Race to Reality: The Next Billion-People Market Opportunity
byAMD
 
GPU Compute in Medical and Print Imaging
byAMD
 
Enabling ARM® Server Technology for the Datacenter
byAMD
 
Lessons From MineCraft: Building the Right SMB Network
byAMD
 

Recently uploaded

PDF
Spec-Driven Development with Kiro: Elevating Software Quality, Traceability, ...
byHaim Michael
 
PPTX
Introducing VisualSim 2610 The Next Leap in System Level Modeling
byDeepak Shankar
 
PDF
Security-Fails-in-the-Gaps-Between-Systems.pptx.pdf
byOhhproJunction
 
PDF
final~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~.pdf
byomarbishtawi04
 
PDF
Explaining the flow of purpose-specific BOM
byakipii ogaoga
 
PDF
Traditional-Security-Models-No-Longer-Work.pptx (1).pdf
byOhhproJunction
 
PPTX
Do You Control the AI, or Does the AI Control You?
bymedhateltoukhy
 
PPTX
Celonis Optimizing your future with process
bydevjeremyappleyard
 
PDF
Poročilo odbora CIS (CH08873) za leto 2025 na letni skupščini IEEE Slovenija ...
byUniversity of Maribor
 
PDF
From Artificial Intelligence to African Intelligence: Transforming Research &...
byMoses Kemibaro
 
PDF
Towards a Vibrant AI Hardware Accelerator Ecosystem, invited talk at the 4th ...
byMichiaki Tatsubori
 
PDF
Odoo Implementation Checklist: A Strategic ERP Blueprint for Business-Ready D...
byBANIBRO IT SOLUTION
 
PDF
NTG -Company Profile_Software_Vendor_Company_in_MENA
byMustafa Kuğu
 
PDF
Constraint Collapse and Fidelity Decay in Scaled Language Models
bySemantic Fidelity Lab | A. Jacobs
 
PDF
February 2026 Patch Tuesday hosted by Chris Goettl and Todd Schell
byIvanti
 
PDF
Session 1/5: Enhancing Automation with Screenplay & Business Rules
bysuhanisingh58689
 
PPTX
PPTX game guess the logo with a twistppt
byAdrianAnig
 
PDF
GTM-and-Sales-Plan for a cyber security product
byAshish Jangir
 
PPTX
Automation Without Apprentices: How AI Challenges the Open Source Way
byIcinga
 
PDF
Oracle Cloud Infrastructure 2025 Architect Professional (1Z0-1127-25) Master ...
byExamCert App
 
Spec-Driven Development with Kiro: Elevating Software Quality, Traceability, ...
byHaim Michael
 
Introducing VisualSim 2610 The Next Leap in System Level Modeling
byDeepak Shankar
 
Security-Fails-in-the-Gaps-Between-Systems.pptx.pdf
byOhhproJunction
 
final~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~.pdf
byomarbishtawi04
 
Explaining the flow of purpose-specific BOM
byakipii ogaoga
 
Traditional-Security-Models-No-Longer-Work.pptx (1).pdf
byOhhproJunction
 
Do You Control the AI, or Does the AI Control You?
bymedhateltoukhy
 
Celonis Optimizing your future with process
bydevjeremyappleyard
 
Poročilo odbora CIS (CH08873) za leto 2025 na letni skupščini IEEE Slovenija ...
byUniversity of Maribor
 
From Artificial Intelligence to African Intelligence: Transforming Research &...
byMoses Kemibaro
 
Towards a Vibrant AI Hardware Accelerator Ecosystem, invited talk at the 4th ...
byMichiaki Tatsubori
 
Odoo Implementation Checklist: A Strategic ERP Blueprint for Business-Ready D...
byBANIBRO IT SOLUTION
 
NTG -Company Profile_Software_Vendor_Company_in_MENA
byMustafa Kuğu
 
Constraint Collapse and Fidelity Decay in Scaled Language Models
bySemantic Fidelity Lab | A. Jacobs
 
February 2026 Patch Tuesday hosted by Chris Goettl and Todd Schell
byIvanti
 
Session 1/5: Enhancing Automation with Screenplay & Business Rules
bysuhanisingh58689
 
PPTX game guess the logo with a twistppt
byAdrianAnig
 
GTM-and-Sales-Plan for a cyber security product
byAshish Jangir
 
Automation Without Apprentices: How AI Challenges the Open Source Way
byIcinga
 
Oracle Cloud Infrastructure 2025 Architect Professional (1Z0-1127-25) Master ...
byExamCert App
 

Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core

  • 1.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 1 of 33 Zen 2: The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core T. Singh1, S. Rangarajan1, D. John1, R. Schreiber1, S. Oliver1, R. Seahra2, A. Schaefer1 1AMD, Austin, TX, 2AMD, Markham, ON, Canada Presented at ISSCC 2020
  • 2.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 2 of 33 Outline • Motivation • Market Segments • Architecture • Core Complex • Technology • Implementation • SRAMs • Power • Silicon Results • Conclusion
  • 3.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 3 of 33 Motivation • Zen was a huge lift • Zen2 compelling successor to Zen • Goals – Give above industry trend generational performance improvement – Enable 2x cores same socket – Improve single thread (1T) performance • How can we do this? – Technology port – Architectural changes – Physical design and methodology changes • AMD was aggressive and we did all of the above to achieve the goals!!
  • 4.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 4 of 33 Zen 2 Market Segments
  • 5.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 5 of 33 Zen 2 Architecture • Changes from Zen – New TAGE Branch Predictor – Optimized L1 Instruction Cache: 32K/8-way vs. 64K/4-way – 2X Op Cache Capacity: 4K vs. 2K ops – 2X Floating Point Data Path Width: 256b vs 128b – 3rd Address Generation Unit – Larger Physical Structures: Integer Scheduler, PRF, ROB, Store Queue, L2DTLB – 2X L1 Data Cache Read/Write Bandwidth – 2X L3 Cache: 16MB vs. 8MB per Core Complex (CCX) • +15%1 single thread (1T) IPC over Zen • ~9% switching capacitance (CAC) improvement over previous generation, technology neutral 1 AMD "Zen 2" CPU-based system scored an estimated 15% higher than previous generation AMD “Zen” based system using estimated SPECint®_base2006 results. SPEC and SPECint are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org.
  • 6.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 6 of 33 Core Functional Units • 32KB IC • 32KB DC • ~20 blocks, ~400K avg instances • ROM for uCODE • 5 L1 RAM variants • Chip Pervasive Logic (CPL) – clock/test block Floating Point Data Cache Load/ Store ALU Scheduler Branch Prediction I-CacheDecode L2 Cache uCode CPL
  • 7.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 7 of 33 L2/L3 Cache Hierarchy • Only 3 unique custom macros – Down from 8 on Zen • Each 4M slice is identical • Multi-stage clock gating in L3 to keep clock distribution power the same as 8M L3 from Zen • LDOs incorporated into the L3 to supply VDDM to L2 and L3 arrays – Loss of package distribution of VDDM meant LDOs had to be moved closer – Must reduce current on VDDM CTLL3Tags L3Data 4M Slice L2 Data L2 Tags L2 Status Shadow tag macros for serving external probes 512K L2 LDOs
  • 8.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 8 of 33 Zen 2 Core Complex (CCX) • 4 core complex • L3 size increases to 16MB • Design for flexibility • Maximize # cores for server case
  • 9.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 9 of 33 Zen 2 CCX Configs HEDT/Server 4 Core, 16MB L3 CCX APU 4 Core, 4MB L3 CCX Value 2 Core, 4MB L3 CCX • Zen 2 Core can be used in various configs covering a wide power range • Multiple CCX can be placed to achieved desired core count Cores Market TDP 8 Notebook 15W 6 Desktop 65 W 8 Desktop/Server 65-120 W 12 Desktop/Server 105-120 W 16 Desktop/Server 105-155 W 24 HEDT/Server 155-280 W 32 HEDT/Server 155-280 W 48 Server 200-225 W 64 HEDT/Server 200-280 W
  • 10.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 10 of 33 Zen vs. Zen 2 Technology Comparison Zen Zen 2 Tech 14nm FinFET 7nm FinFET Cores/CCX 4 Cores, 8 Threads 4 Cores, 8 Threads Area/CCX 44 mm2 31.3 mm2 L2/core 512KB 512KB L3/CCX 8MB 16MB CPP 78 nm 57 nm Fin Pitch 48 nm 30 nm 1x Metal Pitch 64 nm 57 nm Stdcell Track Library 10.5 track 6 track Cu Metal Layers 11 w/ MiM 13 w/ MiM
  • 11.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 11 of 33 Zen vs. Zen 2 Technology Comparison (cont) Zen (14nm) Zen 2 (7nm) Layer Name Pitch Layer Name Pitch M0 StdCell Internal n/a M0 StdCell Internal 1.0x M1 StdCell Internal 1.0x M1 Stdcell & BEOL 1.425x M2-M3 1.0x M2-M3 1.0x-1.1x M4-M7 1.25x M4-M7 2.0x M8-M9 2.0x M8-M9 2.0x --- --- M10-M11 3.15x M10-M11 (RDL) 11.25x M12-M13 (RDL) 18.0x
  • 12.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 12 of 33 Place and Route Design Optimization • 7nm FinFET presents unique route challenges – Lower layer jogs forbidden – Denser standard cells with reduction in track height – Increased lower level metal resistance • Deep collaboration between AMD CAD, foundry, and EDA partners – Cell density management – Advanced legalization techniques – Improved pre-route timing estimates – Wire Engineering and Via Ladders Same-Layer Jogs Forbidden Inter-Layer Jumpers Required
  • 13.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 13 of 33 Placement Restricted by Large Cells • Multi-row cells benefit power and area, but create placement challenges • Clustering of flops has many benefits but can cause placement issues • Resulting small gaps are challenging to use and required innovation to exploit – New algorithms – Flexible power grid choices
  • 14.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 14 of 33 Design RC Miscorrelation • Pre-route vs Post-route miscorrelation caused by length and layer assumptions • Pre-route miscorrelations for resistance and capacitance have differing root causes – Layer assignment for resistance – Length estimates for resistance and capacitance • Based on previously modeled trends, EDA tools may have challenges estimating delay • Required innovation to tackle Layer Normalized Resistance Normalized Capacitance M1 1.00 1.00 M2 3.17 0.96 M3 2.31 0.96 M4 0.72 0.75 M5 0.55 0.83 M6 0.52 0.83 M7 0.55 0.83 M8 0.52 0.83 M9 0.55 0.92 M10 0.16 0.96 M11 0.16 0.92
  • 15.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 15 of 33 Pre-Route Correlation Improvements • Plots showing ClockTreeSynthesis vs Route timing • Large variance in initial results – Large number of paths have overly- pessimistic delay during pre-route steps. Tools waste resources trying to fix – Significant number of paths have optimistic delay estimates. These paths are under- optimized • Employed timing with targeted capacitance scaling and global route- based layer estimation – Standard deviation dramatically improved while keeping a slightly pessimistic mean cts_vs_route.slack.corr cts_vs_route.slack_delta.hist cts_vs_route.slack.corr cts_vs_route.slack_delta.hist Timing Slack Correlation Timing Slack Delta Initial Results Improved Results Pessimistic Optimistic
  • 16.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 16 of 33 Wire Engineering Challenges • Lower layers getting more resistive with latest technology nodes – Very short routes in tight data paths need a buffer – Routes longer than Steiner due to complex rules – Challenging for optimization tools to comprehend • Critical signals need to get to higher layers quickly
  • 17.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 17 of 33 Wire Engineering and Via Ladders • Team used selective layer optimization, buffering, pre-routes, and via ladders to exploit the fast layers for critical signals • Two types of via ladders – High Performance: for large buffers driving long wires – EM: for high-activity gates (e.g., clock drivers) – Mitigated EM issues on large fanout nodes with high activity Top Via Ladder View Side Via Ladder View
  • 18.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 18 of 33 L2/L3 Cache Changes • Zen had an on-die LDO to generate VDDM supply for use by cache arrays • Zen 2’s package choices make using package layers for VDDM distribution impossible • Moved the bitline precharge from VDDM to VDD to reduce current VDDM VDDM BLT[] BLC[] WRCS[] RDCSX[] SAPCX SAEN WDT_X WDC_X XCENX SAT SAC SAC_INT SAT_INT BLPCX WL[N:0] BLT[] BLC[] WRCS[] RDCSX[] SAPCX SAEN XCENX SAT SAC SAT SAC BLPCX WL[N:0] NegBL Write DriverWDT_X WDC_X
  • 19.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 19 of 33 • Moving bitline precharge to VDD creates both bitcell stability and writeability challenges • High level of configurability allows for silicon flexibility VDD Precharge Challenges SRAMSRAMSRAMSRAMs WLUdEn NegBlEn Assist configurations Fuses Assist controller System Management Programming details superVminEn superVmaxEn Voltage thresholds superVmaxEn=1superVminEn=1 VDDM VDDmax VDDmin VDD Controller pauses voltage increase and unsets superVminEn register before continuing to raise voltage Controller pauses voltage increase and sets superVmaxEn register before continuing to raise voltage VDD where VDDM-VDD=superVminThreshold VDD where VDD-VDDM=superVmaxThreshold
  • 20.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 20 of 33 • Moving precharge to VDD reduced our current enough to allow on die-distribution but presents other challenges • Read before write timing challenges at low VDD, high VDDM VDD Precharge Timing Challenges WL@ constant VDDM BLPCX @ high VDD BLPCX @ low VDD WL on before Bitline precharge turns off at low VDD! Bitline precharge turns on before WL turns off at high VDD! Power races with WL Read before write challenges
  • 21.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 21 of 33 • Solving these multiple voltage timing challenges required a number of techniques – Dual voltage clock shapers to average two voltage domains • Can alter the number of these buffers on VDD or VDDM or remove them entirely to make timing more or less dependent on either supply – False read before write problem can be mitigated by compressing the front end of the WL during a write operation Solving Timing Challenges ISOX@VDDM Input@VDD shapedFallInput @VDD VDDM LS LS VDD Psuedo-dynamic level shifter WREN WLCLK WLCLK_shape WLCLK WL during read WL during write
  • 22.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 22 of 33 CAC Comparison • 3% decrease in flop power allocates more budget for combinational logic
  • 23.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 23 of 33 FLOP Palette Improvements • Rich flop library, balance timing/power needs by driving right flop mix • Up to 8% Fmax benefit from high speed flops in timing critical loop paths Best for Performance Best for Power
  • 24.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 24 of 33 Low Power Gater Latch Energy with AvgApp Activity (fJ) State LP Latch Regular Latch Ratio E=1 0.22 0.18 121% E=0 0.17 1.61 10% Total 0.38 1.79 22% E TE CLKB CLKBB CLKBBCLKB CLKBB CLKB qf_x qf Q Dbar Dbar qf CLK • 90% Power savings in latch for common case of E = 0 through internal self gating • Clock gater latch power contribution from 22% in Zen to 13% in Zen 2 for an average application
  • 25.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 25 of 33 Zen 2 Clock Optimization • Multi mesh plan for the core supported by configurable clock tree construction – FP level mesh gating enabled with minimal timing/area overhead – 15% Mesh power savings in Idle and Average App • Tight clock skew distribution • Relocated clock spines and technology shrink (vs. Zen) achieves similar skew profile while reducing CAC
  • 26.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 26 of 33 Zen vs. Zen 2 CAC Comparison • Primary sources of CAC reduction – 14 nm to 7 nm scaling – 6 track library – Aggressive microarchitectural CAC optimizations
  • 27.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 27 of 33 Generational Leadership Perf/Watt • Performance/Watt driven by a combination of technology and design improvements • Timing – Improved scalability by optimizing at a wider voltage range compared to Zen – Multi-corner optimization • Library choice and optimization – 6 track library enabled additional CAC/leakage savings in addition to default technology entitlement • Design CAC – MBFF, low power clock-gater library optimization – RTL improvements – CAC aware downsizing methodology Zen power @ 100% IPC Zen2 power @ 115% IPC 7nm CAC Savings 7nm Timing Design CAC Savings Library Choice Power Improvements – ISO Frequency
  • 28.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 28 of 33 Frequency/Power Silicon Results • 4 cores active with 2 threads per core • The combined effect of lower Vmin for the same frequency and reduced CAC enabled a 50% reduction in power for a given frequency throughout most of the F(P) curve • This enables 2x cores in the same socket!! 50% power reduction
  • 29.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 29 of 33 Frequency/Voltage Silicon Results • 1 core active with two threads per core, 3 cores idle • F/V curve improved over all voltages • Design worked to improve the low voltage performance for improved linearity • Wide voltage range
  • 30.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 30 of 33 Conclusion • Met Goals – Moved to energy efficient TSMC 7nm finFET – Made huge architectural changes – Improved PD and methodology • Results are clear – Scalable across 15W mobile to 280W Server – 50% reduced power at iso-frequency – Enable 2x cores in same-socket – >15% 1T IPC over previous generation – ~9% CAC improvement over previous generation technology neutral – Enables peak frequencies up to 4.7GHz (+350MHz generationally) • Zen2 delivers generational performance uplift!!
  • 31.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 31 of 33 Acknowledgements • We would like to thank our talented AMD design team across Austin, Fort Collins, Santa Clara, Boston, Markham, and India who contributed to Zen 2 • Please stay for our chiplet paper next
  • 32.
    2.2 : AMDChiplet Architecture for High-Performance Server and Desktop Products © 2020 IEEE International Solid-State Circuits Conference 32 of 33 Disclaimer and Endnotes DISCLAIMER The information contained herein is for informational purposes only, and is subject to change without notice. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein. No license, including implied or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations applicable to the purchase or use of AMD’s products are as set forth in a signed agreement between the parties or in AMD's Standard Terms and Conditions of Sale. GD-18 All rights reserved. AMD, the AMD Arrow logo, EPYC, RYZEN, Threadripper, Radeon, Infinity Fabric, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.
  • 33.
    2.1: Zen 2:The AMD 7nm Energy-Efficient High-Performance x86-64 Microprocessor Core© 2020 IEEE International Solid-State Circuits Conference 33 of 33