As a momentous 2020 fades into the history books, 2021 is expected to be a year of growth and evolution for the semiconductor industry across multiple market segments. Firstly, DDR5 DRAM is slated to enter volume production by the end of 2021, with initial deployments targeting hyperscale data centers. Secondly, AI/ML neural networks – which have achieved an impressive 10X annual increase in the size of training models – could hit sizes of well over a trillion parameters in 2021. Thirdly, 2021 will see an increasing emphasis placed on a hardware-based security paradigm with the proliferation of silicon with security cores specifically designed to protect sensitive cryptographic functions and data.
1. DDR5 DRAM
The recently announced DDR5 DRAM standard supports higher-capacity DRAM devices, allowing servers or system designers to utilize densities of up to 64 Gb DRAMs in a single-die package. DDR4 maxes out at 16 Gb DRAM in a single-die package. So, while DDR4 LRDIMMs could have capacities of up to 64 GB, DDR5 LRDIMMs can achieve 256 GB. Additionally, DDR5 features on-die ECC, error transparency mode, post-package repair, and read and write CRC modes to support higher-capacity DIMMs. Lastly, DDR5 data buffer chips effectively reduce the load on the data bus, enabling higher-capacity DRAMs on the DIMM without degrading latency.
We expect DDR5 to enter volume production at the end of 2021, with the first deployments targeting the hyperscale data centers that are becoming critical hubs of the global data network. As of mid-2020, there were 541 hyperscale data centers worldwide with another 176 in various stages of development. This represents a doubling in the number of operational hyperscale data centers since 2015. We expect construction of hyperscale data centers to continue apace in 2021 to support the zettabytes (1021 bytes) of data generated and consumed by smartphones, PCs, game consoles, IoT devices, advanced driver assistance system (ADAS) enabled vehicles and more.
2. AI/ML Training
Since 2012, AI/ML neural networks have achieved an impressive 10X annual increase in the size of training models. This trend suggests that training models could hit well over a trillion parameters in 2021. Indeed, in late 2020, OpenAI debuted a 175-billion parameter GPT-3 language model – representing a 100X jump over the size of GPT-2’s 1.5 billion parameters (introduced in 2019). It should be noted that even during its heyday, Moore’s Law could not deliver the necessary improvements to keep pace with a 10X annual increase in demand.
Memory capacity and bandwidth – which keeps AI accelerators and processors from being bottlenecked – must remain a central area of focus for the semiconductor industry. In 2021, memory such as GDDR and HBM will continue to evolve as the primary high-performance memory solutions for AI/ML. GDDR offers excellent bandwidth and a ruggedness achieved through over two decades of high-volume manufacturing. These characteristics make it an outstanding choice for high-reliability AI/ML inference applications such as advanced driver-assistance systems (ADAS). For applications like AI/ML training with its insatiable need for bandwidth, the performance of HBM is without rival.
3. Hardware-Based Security
As exploits and breaches continue to multiply, 2021 will see an increasing emphasis placed on a hardware-based security paradigm. To be sure, we can expect a proliferation of silicon with embedded hardware root of trust and network security cores that are specifically designed to protect sensitive cryptographic functions and data. This is the most effective way to secure data when at rest (processed or stored in a device) and when in motion (communicated between connected devices).
For data at rest, a hardware root of trust anchored in silicon can provide a hardened foundation to ensure confidentiality, integrity and authenticity for operating systems, applications, and boot code. For data in motion, security anchored in hardware at the foundational communication layer provides a similar model for trust in communications across the entire network.
In 2021, we will see the increased adoption of a hardware-based security model to protect a wide range of applications and use cases such as networks (MACsec in SoCs and FPGAs), chiplets, anti-tamper technology, and ML models.
Conclusion
2021 is projected to be a year of growth and evolution across multiple market segments. Key semiconductor trends in 2021 will include the rollout and adoption of DDR5 memory, with initial deployments targeting hyperscale data centers. As well, AI/ML neural networks – which have achieved an impressive 10X annual increase in the size of training models – are on track to hit well over a trillion parameters in 2021. Lastly, 2021 will see an increasing emphasis placed on a hardware-based security paradigm, with the proliferation of silicon with dedicated hardware cores specifically designed to protect sensitive cryptographic functions and data.
