Supercomputers/Analysis/Silicon generation decay
Analysis 07
Refresh cadence and silicon generation decay
Of the accelerator generations in this dataset that have actually left the installed base, the median stayed 7 years. The longest-lived is NVIDIA Tesla K20X at 10 years, a part released in 2012 and still running well into the following decade. The median gap between a part being released and appearing in a large system is 1 year, which is the number most often assumed to be zero.
Phase 1 seed dataset, compiled by hand. These rows were built from public operator, laboratory and vendor sources. A mechanical second-reader pass has since fetched every cited source: 32 of 52 systems have a readable citation that names them, and 4 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
Read the shape, not the spikes. This is the installed base of the 52 systems in this dataset: a curated seed chosen for supply-chain diversity, not a census of the market. At that sample size one machine entering or leaving service moves a share line by tens of points, so year-to-year jumps are composition effects rather than market events. The multi-year trends are the part worth quoting; the individual steps usually have one system's name on them, and the table view under the chart will tell you which.
- NVIDIA
- Intel
- AMD
- Other suppliers
Table view: Accelerator generations in the installed base
| Generation | Supplier | Released | In service | Years | Release→deploy | Peak installed | Peak year | Systems |
|---|---|---|---|---|---|---|---|---|
| IBM PowerXCell 8i | IBM | 2008 | 2008–2013 | 6 | 0 | 1.03 PFlop/s | 2008 | 1 |
| NVIDIA Tesla M2050 | NVIDIA | 2010 | 2010–2010+ | 1 | 0 | 2.57 PFlop/s | 2010 | 1 |
| NVIDIA Tesla K20X | NVIDIA | 2012 | 2012–2021 | 10 | 0 | 17.59 PFlop/s | 2012 | 2 |
| Intel Xeon Phi 31S1P | Intel | 2013 | 2013–2018 | 6 | 0 | 61.44 PFlop/s | 2013 | 1 |
| Intel Xeon Phi 7250 (Knights Landing) | Intel | 2016 | 2016–2023+ | 8 | 0 | 44.85 PFlop/s | 2017 | 3 |
| NVIDIA Tesla P100 | NVIDIA | 2016 | 2017–2024 | 8 | 1 | 29.36 PFlop/s | 2017 | 2 |
| NUDT Matrix-2000 | NUDT | 2017 | 2018–2018+ | 1 | 1 | 61.44 PFlop/s | 2018 | 1 |
| NVIDIA Tesla V100 | NVIDIA | 2017 | 2018–2025+ | 8 | 1 | 263.1 PFlop/s | 2018 | 3 |
| NVIDIA A100 SXM4 40GB | NVIDIA | 2020 | 2020–2025+ | 6 | 0 | 200.5 PFlop/s | 2021 | 5 |
| AMD Instinct MI250X | AMD | 2021 | 2022–2025+ | 4 | 1 | 1.73 EFlop/s | 2022 | 2 |
| NVIDIA A100 SXM4 64GB | NVIDIA | 2020 | 2022–2025+ | 4 | 2 | 241.2 PFlop/s | 2022 | 1 |
| Intel Data Center GPU Max 1550 | Intel | 2023 | 2023–2025+ | 3 | 0 | 1.01 EFlop/s | 2023 | 1 |
| NVIDIA H100 SXM5 | NVIDIA | 2022 | 2023–2025+ | 3 | 1 | 820.8 PFlop/s | 2023 | 5 |
| AMD Instinct MI300A | AMD | 2023 | 2024–2025+ | 2 | 1 | 1.74 EFlop/s | 2024 | 1 |
| NVIDIA GH200 Grace Hopper Superchip | NVIDIA | 2023 | 2024–2025+ | 2 | 1 | 434.9 PFlop/s | 2024 | 4 |
How this was calculated
A generation is present in a year if any system carrying it was in service that year. A system is in service from its first-operational date until it is retired.
Two modelling problems had to be fixed before this chart could be honest, and both are worth stating.
First, component edges originally had no time bounds, so a system that was rebuilt with
different silicon made both generations look present for its whole life. Tianhe-2 is the
case: Intel Xeon Phi from 2013, NUDT Matrix-2000 from 2018. Edges now carry optional
era_start and era_end dates, and without them this page would have
credited a discontinued Intel part with an extra seven years of installed life.
Second, a system recorded as operational with no retirement date means we found no
source saying it was retired, which is an absence of evidence, not evidence of absence.
Systems now carry a status_as_of date recording when the status was last supported
by a source. Tianhe-1A forces this: read naively it would credit a 2010 Fermi part with sixteen
years in service on the strength of a missing citation.
Spans that are still open are excluded from the median, because including a generation that has not finished would bias the figure downward: every currently-deployed part looks short-lived until it is retired. Only the 4 generations with a closed span contribute to the 7-year median.
“Peak installed” is the highest total Rmax of all systems carrying the part in any one year. It is a capability measure, not a unit count.
The read
The headline is how long generations last, and it is longer than the release cadence suggests. Vendors ship a new accelerator every eighteen months to two years; the installed base turns over on something closer to a seven-to-ten-year cycle. Kepler's K20X entered service in 2012 and was still running in 2021. The gap between those two clocks is where most of the confusion about “refresh cycles” lives: a procurement cycle and a product cycle are not the same cycle.
The second pattern is the release-to-deployment lag. Parts do not appear in leadership systems the year they launch; the median here is a year, and the exceptions run longer. That lag compounds with the announcement-to-first-light lag on the schedule page: a system announced around silicon that has not shipped is committing to two uncertainties at once.
The third is displacement, and the clearest case is not a technical one. Intel's Xeon Phi 31S1P has a closed span ending in 2018, and it did not end because something faster arrived: it ended because an export control blocked the upgrade path and NUDT substituted its own part. On this chart that is a generation dying at a policy boundary rather than a performance one, which is exactly the kind of thing a supply-chain view can show and a ranking cannot.
The bars at the right-hand edge are all open, which is the correct rendering and also the limitation. Every current generation looks short-lived because it has not finished yet. Nothing on the right third of this chart should be read as a lifespan.