COMPUTERS OF 2020 AND 2030

Tat IwamotoJanuary 2013

Saturday, January 19, 13

WHERE ARE WE NOW?

•Most popular computer of 2012, iPhone 5•No longer PC. Desktop and laptops are irrelevant.

Saturday, January 19, 13

IPHONE 5 SYSTEM

4Gb LPDDR2

4Gb LPDDR2

Modem

128Gbit NAND

Wifi/BT

Misc Peripherals

CPU

GPU

x21.3GHz

x3325MHz

Saturday, January 19, 13

DIE SIZE

• A6 (32nm) 100mm^2

• LPDDR2 (20nm) 100mm^2 x2

• Flash (20nm) 200mm^2

• = 400mm^2

• Just very rough estimate

Saturday, January 19, 13

MOORE’S LAW

• 2013 = 400mm^2

•Density increase

• 2020 = 16x

• 2030 = 1000x1

10

100

1,000

10,000

100,000

1,000,000

2010 2015 2020 2025 2030

Density converted to 2013 mm^2

Saturday, January 19, 13

SCALE THE SAME CONFIGURATION

• 2020

• 32 core ARM CPU

• 48 core GPU

• 16GB DRAM

• 256GB Flash

• 2030

• 2000 core ARM CPU

• 3000 core GPU

• 1TB DRAM

• 16TB Flash

Saturday, January 19, 13

LIMITERS (1)

• IO Speed

•Do not scale with density

• Assume scale with linear length

• 1.4x for 2x density

• 10Gbit/s limit for PCB

Saturday, January 19, 13

MEMORY IO BANDWIDTH

4Gb LPDDR2

4Gb LPDDR2

128Gbit NAND

1Gbit/s x64 bit=8GB/s

40Mbit/s x4 bit=20MB/s

CPU

GPU

x21.3GHz

x3325MHz

Saturday, January 19, 13

CPU

GPU

x32

x48

MEMORY IO BANDWIDTH 2020

64GB DRAM

256GB Flash

10Gbit/s x100 bit=128GB/s

640Mbit/s x4 bit=320MB/s

1Gbit/s x16 > 10Gbit/sMore pins needed to scale

Saturday, January 19, 13

MEMORY IO BANDWIDTH 2030

1TB DRAM

16TB Flash

10Gbit/s x800 bit=8TB/s

10Gbit/s x16 bit=20GB/s

800 pins is impossible

20Mbit/s x1000 >10Gbit/sMore pins needed to scale

CPU

GPU

x2000

x3000

Saturday, January 19, 13

LIMITERS (2)

• Process difference

• Logic

•DRAM

• Flash

• Use different process

Saturday, January 19, 13

LIMITERS (3)

• Power

•Need to keep battery time

• Frequency and power roughly scales with linear length

• 2x Density = 1.4x Speed, 0.7x Power

•Need to lower activity to keep heat at same level

Saturday, January 19, 13

LIMITERS (4)

• Chips can’t get smaller

• Pad limit

• Package pin limit

Saturday, January 19, 13

INTEGRATE DRAM DIES

• DRAM pins run out first

• Memory is cooler than logic

• Keep flash off chip

• 2x transistors = 1.4x logic + 0.7x additional area

• Use additional area for memory

• DRAM -> SRAM

• DRAM : SRAM 6:1 bit density

Saturday, January 19, 13

2020 RAM ON-CHIP

• 16x Density

•Die -> 7mm^2

• x 4 for same power= 28mm^2

• 72mm^2 for memory

• 1GB DRAM -> 12mm^2

•DRAM vs Logic density gap 2:1

•DRAM:SRAM = 1:6

• 1GB SRAM = 144mm^2

• 72mm^2 = 512MB SRAM

Saturday, January 19, 13

CPU

GPU

x8

x12

2020 SOC256GB Flash

640Mbit/s x4 bit=320MB/s

No external DRAMNo pinsNo IO power

512MBSRAM

CPU 1.3 x 4 = 5.2GHzGPU 325 x 4 = 1.3GHzEnough memory bandwidth

Saturday, January 19, 13

2030 ONE CHIP• 1000x Density

•Die -> 0.1mm^2

• x 32 for same power= 3mm^2

• 97mm^2 for memory

• Logic only 3%

•Not dominant area

• 1GB DRAM -> 0.2mm^2

• 16GB Flash -> 0.2mm^2

• 1 GB SRAM = 2.4mm^2

• 97mm^2 = 40GB

• Build chip on Flash process

• SRAM + Flash

•Maybe lower frequencySaturday, January 19, 13

CPU

GPU

x64

x96

2030 SOC

No external DRAM/FlashNo pinsNo IO power40GB

SRAMor Flash CPU 1.3 x 32 = 40GHz

GPU 325 x 32 = 10GHzConsider SRAM / Flash mix

Saturday, January 19, 13

TWEAKS

• Use better CPU/GPU

• Less cores / more memory

• Smaller overall die?

Saturday, January 19, 13

MORE OPTIONS

• Integrate other IO peripherals

• Should be enough room for that, too

•OK as long as pin count is low

• Special IO may require lower density

Saturday, January 19, 13

OTHER DEVELOPMENTS

•Mixed process technology: logic + DRAM + Flash

• Package technology: more pins, more interconnect speed

• Interconnect technology: 10Gbit/s+

Saturday, January 19, 13