Post on 15-Feb-2018
transcript
Host Memory Buffer (HMB)based SSD SystemForum J-31: PCIe/NVMe Storage
Jeroen Dorgelo
Mike Chaowei Chen
Agenda
• Market Trends in Client SSD
• Introducing DRAMless architecture for SATA & PCIe
• Host Memory Buffer(HMB) scales up DRAMless PCIe SSD
• HMB DRAMless vs. DRAM-based SSD
• Ecosystem development
13-Aug-15Flash Memory Summit 2015 2
Market Trends
• Main drivers for SSD adoption have been:
– Performance
• PC Mark8 score from HDD to SSD: 1,200 to 4,800 points (SATA)
• Latest PCIe Gen3x4 NVMe drives score over 5,100 points
– Form Factor:
• Ultrabooks require space to fit in battery
• SSD can shrink to module size (M.2)
• BUT all came with a Premium to pay
13-Aug-15Flash Memory Summit 2015 3
170 167 165
145139
Dri
ve
ma
rke
t p
rice
($
)
256GB 128GB HDD
• Now: cost parity of 128GB SSD and 500GB Mobile HDD
Market Trends
13-Aug-15Flash Memory Summit 2015 4
145139
129119
111100
9582
76
103 102 99 9787
7565
5950 46 41 3843 41 40 40 40 40 39 39 39 38 38 38
Dri
ve
ma
rke
t p
rice
($
)
Time
Developing DRAMless architecture
DDR3
4GB
CPU
Southbridge
128GB SSD Mainstream Entry Level
NAND CH 4 2
NAND speed 333MT/s 400MT/s
Capacity 64Gb MLC
16dies
128Gb TLC 8dies
Memory 16bit DRAM I/F DRAMless
• 2015 SSD Architecture: Mainstream vs Entry-level
13-Aug-15Flash Memory Summit 2015 5
Mainstream SSD
SATA
DDR3
PC
NANDNAND
NANDNAND
NANDNAND
NANDNAND
NANDNAND
NANDNAND
NANDNAND
NANDNAND
SSD SOC
CPUs
4CH NAND I/F
DRAM I/FSATA I/F
Memory 16bit DRAM I/F DRAMless
Entry-level SSD
NAND
SSD SOC
CPUs
2CH NAND I/F
NANDNAND
NAND
SATA I/F
NANDNAND
NANDNAND
SATA
What I can get from a $30ish SSD?
• 20X HDD performance in full testing range
• Or 100X HDD IOPS in small 4GB or 8GB range
• And 1/20 space with M.2 2230 or even smaller with BGA
– 1/10 in XY
– 1/2 in Z-height– 1/2 in Z-height
• And <2mW in Devslp mode
13-Aug-15Flash Memory Summit 2015 6
Enabling Host Memory Buffer
• PCIe Gen3x1 NVMe SSD requests Host to allocate
Memory space
– In this case study: 128MB for the Look-up-Table
– But it can claim any custom size of Host DRAM (partial LUT)
• Same BOM cost as Entry Level SATA drive!
13-Aug-15Flash Memory Summit 2015 7
– But it can claim any custom size of Host DRAM (partial LUT)
PCIe
Gen3x1
PC
DDR3
4GB
CPU
drawing not to scale
Southbridge
128GB SSD
NAND
SSD SOC
CPUs
2CH NAND I/F
NANDNAND
NAND
PCIe I/F
NANDNAND
NANDNAND
HMB
Performance benchmark 128GB TLC SSD
PCIe Gen3x1 NVMe (2CH DRAMless)
PCIe Gen3x1 NVMe HMB (2CH DRAMless)
SATA (4CH DRAM based)
Sequential Read 512KB
+28%
0 100 200 300 400 500 600 700
eMMC 5.1 (1CH DRAMless)
UFS (1CH DRAMless)
SATA (2CH DRAMless)
PCIe Gen3x1 NVMe (2CH DRAMless)
MB/s
13-Aug-15Flash Memory Summit 2015 8
+188%
+28%
Test configuration: Chipset(Z97), Windows 8.1, Intel NVMe driver, CrystalDiskMark
Host Memory Buffer reduces latencies
eMMC 5.1
UFS
SATA (DRAMless)
PCIe Gen3x1 NVMe (DRAMless)
PCIe Gen3x1 NVMe HMB (DRAMless)
SATA (DRAM based)
Application to SSD IO Read Latency (QD=1, 4KB)
• Host Memory Buffer significantly reduces latencies:
– Boosts IOPS with 50% over SATA and 100% over eMMC
0 50 100 150 200 250 300
eMMC 5.1
Latency (us) lower is better
LUT access NAND Read NAND transfer Controller
Link Xfer Platform+adapter Host SW
13-Aug-15Flash Memory Summit 2015 9
Performance benchmark 128GB TLC SSD
PCIe Gen3x1 NVMe (2CH DRAMless)
PCIe Gen3x1 NVMe HMB (2CH DRAMless)
SATA (4CH DRAM based)
Random Read 4KB / QD1
+10%
0 2 4 6 8
eMMC 5.1 (1CH DRAMless)
UFS (1CH DRAMless)
SATA (2CH DRAMless)
PCIe Gen3x1 NVMe (2CH DRAMless)
IOPS
13-Aug-15Flash Memory Summit 2015 10
+100%
+50%
Test configuration: Chipset(Z97), Windows 8.1, Intel NVMe driver, CrystalDiskMark
Performance benchmark 128GB TLC SSD
PCIe Gen3x1 NVMe (2CH DRAMless)
PCIe Gen3x1 NVMe HMB (2CH DRAMless)
SATA (4CH DRAM based)
Random Read 4KB / QD32
-50%
0 20 40 60 80 100
eMMC 5.1 (1CH DRAMless)
UFS (1CH DRAMless)
SATA (2CH DRAMless)
PCIe Gen3x1 NVMe (2CH DRAMless)
IOPS
13-Aug-15Flash Memory Summit 2015 11
+736%
+100%
Test configuration: Chipset(Z97), Windows 8.1, Intel NVMe driver, CrystalDiskMark
Client Workloads are light weight
• Review sites have developed trace based benchmarks
– Capturing user behavior on Windows machine
– Feeding commands to the SSD
• Example of queue depth distribution:
– QD1-4 cover >90% of all cases
13-Aug-15Flash Memory Summit 2015 12
HMB Ecosystem development
• Operating system inbox driver & Bios support
– Windows, Linux, Android, etc
• Chipset to support PCIe as storage interface
– Coming in this year!
• Leading-edge SSD controller to support HMB• Leading-edge SSD controller to support HMB
– Marvell 88NV1140
• Deployment to PC OEMs
– Configuration, integration and qualification
• Marvell is working in all areas to drive HMB
13-Aug-15Flash Memory Summit 2015 14
Summary
• A DRAMless SSD delivers good-enough performance and
capacity for an entry-level client PC system
• Moving from SATA to PCIe Gen3 NVMe increases
sequential read speeds and reduces latencies
• Enabling Host Memory Buffer boosts IOPS performance • Enabling Host Memory Buffer boosts IOPS performance
significantly
– 50-100% compared to other DRAMless solutions
• Marvell’s 88NV1140 enables this without adding BOM cost
• We are working with PC OEM to launch HMB soon
13-Aug-15Flash Memory Summit 2015 15
The Opportunity...
• ...is bigger than you’d expect:
– Tablets/Convertables become productivity devices
– eMMC performance doesn’t scale
– PCIe Gen3x1 is a strong alternative
– Advanced Marvell NVMe SSD
controllers under development!
13-Aug-15Flash Memory Summit 2015 16
BGA SSD M.2230 SSD$0.25 size
• The Future of NVMe is NOW!
– Marvell 88NV1140
– PCIe Gen3x1 NVMe 1.2 with Host Memory Buffer support
13-Aug-15 17Flash Memory Summit 2015