System Architecture
Alex Bensch, Pete SevichCMPE 550, Spring 2014
Outline
• PS4 high level overview and history• PS4 Main Board• Jaguar Microarchitecture• APU/GPU specifics• Conclusion
History of the PlayStation
PlayStation (1994)PS2/PSOne (2000)PS3 (2007)PS4 (2013)
PS4 System Overview
PS4 Features
• “Play as you download”• Remote play support w/ PS Vita• Hardware supported streaming
– Integrated streaming software in overlay– Live or previous gameplay
• DualShock 4 controller support
DualShock 4
• 6 axis IMU (3+3 DOF)• Lightbar• MicroUSB extension port• LED Touchpad display• Bluetooth 2.1 communication• 3.7V LIPO battery
– 1000 mAh capacity
PS4 Hardware
• 8 CPU Cores @ 1.6GHz (x86-64)• 8GB GDDR5 Memory, 176GB/s• Radeon 7870-Based GPU• Blu-Ray/DVD Drive• 500GB HDD• 802.11 b/g/n, Gigabit Ethernet• USB 3.0, Bluetooth 2.1
Main PCB
Accelerated Processing Unit (APU)
• Combines CPU and GPU modules in SoC• Advantages
– Compact– Cost effective– Low power
• Disadvantages– Lower spec ceiling– Shared memory between GPU and CPU could limit
performance potential• PS4 uses a custom APU w/ Jaguar modules
Jaguar Implementations• Netbooks, tablets• Mini‐PCs, notebooks• Microservers• Thin clients• XB1 and PS4
Jaguar General Information
• Mid 2013• 28nm lithography • Low‐power (6‐25W TDP*)• Big improvements over Bobcat • Allows APU implementations• 2 or 4 cores, up to 2.05GHz*• Split 64kB L1, up to 2MB shared L2• DDR3/DDR3L 1066‐1600* + ECC*Based on current implementations
PS4 APU Details
• Two quad-core Jaguar CPU modules– 1.6GHz clock – Each core has split 64kB L1 cache– Shared 2MB L2 cache (16 way associative)
• Modified Radeon 7870 GPU– 800MHz clock– 18 compute units (“cores”)– 1152 shader cores– 1.84 TFLOPS shader performance
Jaguar Pipeline
Pipeline Highlights
• Two issue, six dispatch– Six instructions fetched at once– Two decoded per cycle
• Appears to use Tomasulo– Out of order scheduling– Multiple issue
Comparison to Bobcat
• 15% increased OOS window– 56 entries for Bobcat– 64 entries for Jaguar
• Expanded FP op bus and vector units– 64-bit for Bobcat– 128-bit for Jaguar
• Additional iDec and RegRead stage in pipeline
Performance
• FP Ops– Multiply
• Single precision - 2 cycles• Double precision - 4 cycles
– 5 for X87 and only every third cycle– Division - 14-22 cycles– Square root - 16 - 35 cycles
• Some reciprocal square root ops - 2 cycles
Summary
• PS history• PS4 basics and hardware• Jaguar Microarchitecture• PS4 APU/GPU specifics
Questions?
Differences with Xbox One
• 1.75GHz CPU• DDR3 RAM• 47MB eSRAM*• 853MHz GPU• 12/14 CUs• 768 shader cores• 1.23 TFLOPS