Summit's speed came from a system of different processors
Oak Ridge's Summit supercomputer paired CPUs with thousands of GPUs. Performance depended on data movement and software readiness alongside chip capability, with different tests producing very different headline numbers.
Oak Ridge National Laboratory unveiled Summit on 8 June 2018 with a headline peak performance of 200 petaflops. The number described an extraordinary machine, but its commercial significance was easier to see one server at a time. Each of Summit's 4,608 compute servers combined two IBM POWER9 processors with six NVIDIA V100 graphics processors. Faster computing depended on coordinating different kinds of processors and moving data between them.
That architecture offered a useful lens for semiconductor and infrastructure startups. An accelerator could be exceptionally capable at its preferred operation and still disappoint in a customer's application. Summit made the surrounding system visible: memory, interconnects, software tools and the work required to prepare scientific codes for a new machine.
The benchmark supplied a different number
The June 2018 TOP500 list subsequently placed Summit first with 122.3 petaflops on High Performance Linpack, the benchmark used to rank the list. Sunway TaihuLight recorded about 93 petaflops and Sierra about 71.6. These results measured a defined numerical workload, whereas the launch announcement described theoretical peak performance. The benchmark used 4,356 Summit nodes; ORNL's launch described the full system of 4,608 servers.
| Period or category | PFLOP/s on HPL |
|---|---|
| Summit | |
| Sunway TaihuLight | |
| Sierra | |
| Titan |
Measured HPL Rmax, rounded to one decimal place. These benchmark results allow comparison across the four systems; theoretical peaks and individual application performance require different measures.
Source2
Keeping those definitions separate avoids a misleading account of performance. The gap between a benchmark and peak capacity is not automatically a product defect; they measure different things under different conditions. Nor does winning Linpack establish the same advantage in every simulation or machine learning task. For an investor assessing a chip company, the corresponding question is which workload was measured, on what configuration, against which alternative.
A GPU needed the rest of the node
Summit extended the hybrid CPU and GPU approach used by Oak Ridge's earlier Titan system. ORNL described more than ten petabytes of aggregate memory, including different memory technologies, connected through fast data paths. At the server level, NVLink connected processors and accelerators; a dual-rail Mellanox EDR InfiniBand network connected servers. These links were part of how the machine could turn arithmetic capability into useful work.
The same constraint appears at a smaller scale when a startup develops a specialised accelerator. If an application repeatedly waits for data, adding arithmetic units does not guarantee a proportionate improvement. A useful evaluation therefore includes data preparation, transfers, memory requirements and result collection. It also examines where the workload changes shape: a fast operation inside an otherwise serial process may save less time than its isolated benchmark suggests.
Precision changed the meaning of a result
The launch announcement also reported a comparative genomics calculation running at 1.88 exaops. ORNL described that as a mixed precision result and said it produced identical results to slower 64-bit calculations previously run on Titan. The result applied to that particular calculation and its arithmetic, rather than to general double precision scientific workloads.
Scroll to view all columns
| Comparison | HPL benchmark | Mixed precision genomics calculation |
|---|---|---|
| What is evaluated | A defined linear algebra benchmark used by TOP500 | A particular comparative genomics calculation reported by ORNL |
| Useful comparison | Other systems' results on the same benchmark | The time and correctness of alternative implementations of the same calculation |
| What it cannot settle | Every application's performance | General double precision supercomputer performance |
The genomics calculation and HPL benchmark use different workloads and arithmetic. Their headline rates cannot form a single performance ranking.
For an accelerator company, lower precision can be a valuable design choice when the customer's accuracy requirements permit it. The commercial case must show that the completed answer remains useful. Inference accuracy, numerical stability or a scientific validation criterion can matter more than the largest figure for operations per second. A buyer needs the relevant quality measure alongside time, resource use and the cost of obtaining the answer.
Software preparation was part of the machine
ORNL's launch described applications prepared ahead of Summit's arrival, spanning astrophysics, materials research, cancer surveillance and systems biology. The facility's technical material describes a Center for Accelerated Application Readiness with thirteen partnership teams preparing codes for the architecture. Application developers, facility staff and hardware suppliers worked together to make those programs use the new machine effectively.
That preparation points to a meaningful source of value for tools companies: helping developers understand bottlenecks, port code, validate results and keep applications maintainable. It also creates a burden for a chip startup. Every proprietary programming requirement asks the customer to invest time and accept a dependency. A performance advantage becomes more commercially persuasive when the supplier can demonstrate how a real development team reaches and sustains it.
A research facility was not a market forecast
At launch, Summit was scheduled to serve select projects during 2018 while ORNL and IBM worked through acceptance, with the bulk of access in 2019 going to teams selected through the INCITE programme. That timetable distinguished an unveiling from broad operational availability. It also reflected the role of a publicly supported scientific facility, whose workloads and procurement process differ from those of an ordinary enterprise customer.
EFLOP/s on HPL
View data
| Date | EFLOP/s on HPL |
|---|---|
| June 2017 | 0.749 |
| November 2017 | 0.845 |
| June 2018 | 1.22 |
Combined HPL performance of the 500 systems in each edition, using TOP500's rounded measurements. The list changes between editions, so these totals cannot isolate performance gains on the same machines or measure commercial computing revenue. The increase cannot be attributed to Summit alone.
Source4
Summit therefore supported an architectural argument, not a revenue forecast for every accelerator supplier. It showed why heterogeneous computing could be powerful when processors, memory, networking and applications were developed together. A startup still had to identify customers with suitable workloads, an affordable deployment path and enough recurring use to justify adoption. The decisive unit of value was a completed task at an acceptable cost and quality.
Investment implications
- Peak arithmetic capacity, benchmark performance and application results answer different questions.
- Accelerators depend on memory movement and software capable of using them effectively.
- A startup's relevant demonstration is a complete customer workload with a credible path into deployment.
Sources
- ORNL Launches Summit Supercomputer (opens in a new tab)
Oak Ridge National Laboratory
- June 2018 TOP500 list (opens in a new tab)
TOP500
- Summit (opens in a new tab)
Oak Ridge Leadership Computing Facility
- Highlights — June 2018 (opens in a new tab)
TOP500
Published
