An examination of the capital requirements of HDD and silicon-based storage Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 1
2008 Units (M) PB HDD 580 125,400 ipods* 55 357 iphones* 14 100 USB Key** 179 598 Flash Card** 752 1,548 SSD*** 3 48 Other 146 156 Total 1,150 2,782 % HDD PB 100% 0.28% 0.08% 0.48% 1.23% 0.04% 0.12% 2.2% Growing at 35%/year NAND IT Storage * Apple ** Gartner, Oct 2008 ***IDC, Aug 2008 Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 2
2-bit MLC NAND Flash* HDD Wafer Diam. 300mm (12 ) Head Wafer 150mm (6 ) Node 45nm Head Node 60nm Dies/wafer 425 Heads/wafer 30,000 Die capacity 2GB Wafer capacity 850GB Disk Capacity** 375GB Daily output 1,250 wafers Daily output 100,000 disks PB/line/year 390 PB/line/year 14,000 Fab cost $3.4B*** Fab cost $1B * Highest density 2-bit MLC flash. Wafer GB for SLC flash will be lower. ** Highest density SATA disk in 2008. ***Source: IBS Only ~1/2 SATA HDD High capacity HDD ~ $140 Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 3
HDD 2008 2.6 square miles $35B revenue 7% CAGR 125,000PB 35% CAGR* 19% CAGR Silicon 2008 2.2 square miles** $280B revenue*** 16,000PB if all 2.2 sq miles used for 2 bit MLC NAND Fab capital cost increasing at 19%/ Year All WW Si could meet 13% of HDD bytes. That s $4.5B (1.6% of Si). * IDC WW HDD 2008-2012 forecast **Derived from Gartner, Jan 2008. ***Gartner, Nov 2008. Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 4
2008 125,000PB = $35B for HDD market HDD 45nm MLC Per fab 14,000 PB = $4B revenue Capital/Revenue = 0.26 Per fab 400 PB = $2.3B revenue Capital/Revenue = 1.5 The HDD market differs from the flash market. An examination is warranted. That s 3 months At NAND $/GB Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 5
1.8 Ent. 2.5 All HDD Rev ($B) 1.3 5.4 12 35 Flash Cap( $B) 17 53 287 1,080 Cap/Rev 13 10 24 31 HDD Rev ($B) 1.6 5.5 22 45 $1T! Flash Cap ($B) 28 71 729 1,706 Cap/Rev 18 13 33 38 C A G R Rev CAGR 4% 1% 18% 7% Cap CAGR 17% 10% 32% 14% Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 6
2008 2012 $11B per % of HDD PB captured Capital cost = 31% of total HDD revenue per % HDD PB $17B per % of HDD PB captured Capital cost = 38% of total HDD revenue per % HDD PB Actual 2008 SSD share 0.04% = 50PB Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 7
SanDisk HDD 250GB SSD Projects NAND @ $0.24/GB in 2013* 250GB for $60 HDD in 2013: 4TB for $60** 23 hours of 1080p video OK for use in a camera Unbeatable? SSD gives 6% of the GB at the same price Will the market think this SSD is the better value? 4TB HDD 350 hours of 1080p video OK for us in a system 250GB in 2013 is same as 30GB in 2009 - small SSD value proposition involves rationing capacity * E. Harari, 2009 Flash Memory Summit **Assumes 40% CAGR HDD density, while SanDisk assumes NAND prices falling even faster Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 8
$/IOPS Ent HDD SSD Has been asserted as a primary driver in enterprise Chasm analysis shows $/IOPS to be secondary* $5B market, 5% of HDD PB in 2008 Assume 5x over provisioned thus 1% of PB used $11B capital required to produce 1% of PB in 2008 1,250 PB ~40% increase in production Captures $5B in revenue Still 2.2 years of revenue to pay back capital * S. R. Hetzler, IDEMA 2008 This assumes 2-bit MLC. SLC will increase >2x. Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 9
An increase of greater than 20% in cost for doubling of speed of operation is usually considered unwise. J. Presper Eckert (1953) (co-developer ENIAC) Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 10
Keepin g As $/GB declines, data of lower per-unit value can be stored Probability of access to any piece of data is small Aggregate probability of access is very high Value of system increases Deleting Deletion sets future value to $0 Cost of $1TB in 2009: $65 Cost of identifying 1TB to delete > 100x $65 At 10GB/hour, takes 100 hours, ~$10k burdened pro (even more for SW license) Of course, no mistakes will be made Value of system flat but cost increased Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 11
The value of a computing system is proportional to the storage capacity. Growth Performance Storage growth exponential for decades If value < exponential, growth would have slowed (demand pulls supply) If performance dominated value, high end would be all DRAM Enterprise HDD demonstrates the steep market share decline when performance increases costs (Eckert s Law) Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 12
Flash HDD 16 GB $180 $250 Nano 160 GB Classic $600 $750 8 GB HDR-CX100 16GB of flash costs less than 16GB of HDD 160GB of HDD costs less than 160GB of flash 80 GB HDR-XR100 Not a $/GB premium! Flash has lower cost than HDD at small capacities Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 13
Key Trends Capital SSD Budget PB shipped growing @ 35% CAGR Revenue growing @ 7% CAGR Capital is an issue for all Si storage Wafer cost >> HDD cost Silicon very capital intensive relative to HDD Economies of scale already taken at fab level Economic case seems to depend on significant price premium Goes against the key trends above IT budgets are zero sum Making storage faster hasn t increased budgets Sep, 2009 Steven R. Hetzler, Copyright IBM Corp. 2008,2009 14