Physical Verification needs to focus on GPU
Physical Verification could largely benefit from the development of GPU technology. In fact, the advantages of the new technologies could be used in each phase of the process, starting from early design Place and Route phases to the DRC/LVS for tape-out. Compared with most current technologies that are based on huge CPU data centers, the new GPU data centers promise to shift the paradigm and to bring to the market the much-sought ideal of speed and energy savings.
If we analyze Physical Verification we will observe that it is extensively using CPU based solutions, meaning that it could be characterized as costly and time-consuming. As we have explained precedently the main problem in adopting new hardware accelerated technologies is related to the design rules and specifications for manufacturing that are required for each factory. These are making the transition process expensive and complicated. It is expected to see the new technologies adopted first by companies that have agile design methodologies. They will easily shift their physical verification process by adapting the foundry requirements to these new technologies. Later they will be followed by the other Semiconductor industry players who will sense the competitive advantage. We expect to see the same phenomenon as in the AI industry, where a large majority of companies shifted from CPU to GPU.
Why should Physical Verification shift from CPU to GPU Data Centers?
The benefits of shifting from CPU to GPU data centers are strong and derive from various direct and indirect advantages of the new technology.
Data center sizes – The GPU data centers needed for some specific operations are smaller than the CPU ones due to the extremely high power of calculation per unit. Fewer units mean that some buildings could be used for expansion or other economic purposes. Also, smaller space for data centers implies easier administration and better management of the unit. It also implies a smaller maintenance team.
The reality is that an outsider could wonder that the industry does not shift to GPU based solutions for the problems that rely on massive parallelization and still use extensively CPU based software’s that imply huge server data centers.
Level of security – One of the reason EDA software makers and factories refused to shift to cloud processing is that these data servers couldn’t guarantee the level of security they needed for their IPs design. Considering the smaller GPU data center size combined with lower energy consumption, nowadays it is perfectly reasonable and price efficient to built own in-house GPU data centers that would allow the required level of protection.
Speed improvement – It is undeniable that GPUs, through the process of high parallelization, has taken the calculus capacities to unprecedented levels. The speed improvement is so impressive that only by adopting this technology, companies could gain significant competitive advantages. The additional speed in the Physical Verification could shorten the circuit design process significantly.
Electricity consumption – GPUs are known for their very efficient power consumption per unit and seeing that fewer units are needed than in CPU data centers, the overall expenses for electricity are significantly smaller. Their Pflops/s computing power per energy consumption ratio is superior to any known CPU technology. For example, a 1000 CPU Servers Data Center with 16.000 cores costs $5M and consumes 600kW/h. An equivalent GPU Data Center with 18.000 cores can be built with only 3 GPU-Accelerated Servers that cost $30k and consumes 4kW/h.
Saved space – Seeing that the space needed for a GPU farm is usually one order of magnitude smaller than the one needed by any CPU farm, the companies that will shift to the new technology will have the possibility to use their physical facilities for other purposes or to pack huge power calculus capabilities in smaller spaces. Refer to the previous example for a comprehensive image of the saved space.
A faster pace in launching new generations – When speaking of the shift from CPU to GPU technology an important factor that seems to be neglected in most analyses, is the improvement of the new generations. The CPU technology has reached a level of development where new generations bring only minor improvements, GPU is still developing and with each new product generation, we see new functionalities that boost the calculus capacities (for example the newly introduced tensor cores, that amplify hundreds of times the mixed precision calculations).
Competition and the change of paradigm
Today it seems that some of the companies seem not seeing yet the benefits of shifting from CPU to GPU products. This happens because most companies still try to exploit CPU based products for the development of which they made tremendous investments in the past. Also, the specifics of Physical Verification imply the development of new factory rules and procedures for any new verification software. And, as it is well known, each factory has specific rules and procedures, meaning that each new product they will need additional and specific adaptive software. Even though the software producers and the factories seem reluctant to adopt the GPU technology we are confident that the free market competition will determine most companies to make this step and start developing or adopting new products, given the benefits that we discussed above.
AMSIMCEL is ready to bring its contribution to GPU adoption
Here at AMSIMCEL, we are ready to bring our contribution to the industry through our know-how and our GPU based solutions for EDA and ERC. The tests and simulations made in the facilities of AMSIMCEL have results that are more than promising and we are confident that we will be a relevant player in the new age of Physical Verification.
Even if the companies active in Physical Verification like it or not the change is coming and they have to adopt it or risk being excluded from the market. As we have stated before the risks for these companies derive from the following facts:
1. Companies adopting GPU solutions will become more competitive. The factors that increase the competitiveness of the companies are the additional speed, the space savings, the smaller dependency on electricity providers, etc.
2. Companies adopting GPU products will become more cost effective. The fact that the companies will need smaller space, less time and electricity are all cost elements that will make the companies adopting GPU technology more cost-effective. It is clear that the faster they will adopt the technology they will be able to take advantage of the additional competitiveness and benefit of the advantages of early market entrants.
3. Companies adopting GPU products will be faster. It’s needless to say that fast quality services and products delivered on time mean happy clients, and happy clients are a sign of a healthy business. The increase of the companies’ capacity to adapt to the market dynamics would make them better players.
For more on this subject feel free to contact:
Catalin Tugui Gabriel Donici