Stop the Silent Slowdown Killing Your Performance Models
Ramadan is already a month of focus, discipline, and pushing yourself beyond your limits. These are the same qualities that building a company demands. Being surrounded by like-minded Muslim founders at Alif HQ during this month would compound both dimensions at once. I want to be there the full month, end every day with Taraweeh together, and build in an environment where the spiritual and the professional reinforce each other. That combination of a focused founder community, the barakah of Ramadan, and a shared sense of purpose is something I cannot replicate anywhere else.
The whole point of a performance model is to get fast data back to design teams. But cycle-accurate models silently drift slower with every new feature, every team integration, and every shortcut taken under deadline. Nobody tracks it. Nobody owns it. SimDrift makes this invisible problem visible and fixes it.
Chip companies build cycle-accurate performance models to answer architectural questions fast. But these models are maintained by multiple teams (CPU, GPU, NoC, Modem, Memory), each owning their slice and each under pressure to ship features quickly. Nobody is responsible for the health of the combined model.
The result is silent, compounding drift. Every new feature adds a little bloat. Memory leaks accumulate. Code written fast under deadline is never revisited. The combined model gets slower with every integration cycle, and because there is no single moment of failure, nobody notices. It is just accepted as normal.
At large companies like Qualcomm, you absorb this with headcount. At an AI chip startup with a skeleton architecture team, slow models mean fewer experiments, worse decisions, and slower silicon.
SimDrift is a continuous model health platform for cycle-accurate SystemC/C++ performance models. It lives in the development flow, profiling, flagging, and optimizing simulation code as teams integrate new features, so the combined model never silently drifts.
We run SimDrift on a customer's model for free and show them a drift report: how much slower their model has gotten, by team, over the past 12 months and exactly why. That chart closes the sale. Nobody has ever shown them this before.
The product ships in two natural layers:
Target customer: AI chip startups that do not have the bandwidth to run experiments across every subsystem, including companies like Tenstorrent, Etched, and the dozens of smaller ones that do not make headlines.
This comes directly from three years inside one of the best-resourced SoC teams in the world. At Qualcomm, cycle-accurate models silently bloat with every new feature addition and team integration. Each team prioritizes making their feature work over making it optimal, because there is simply no time to do otherwise. The combined model gets slower with every release cycle, nobody tracks it, and eventually it is just accepted as the cost of progress.
If this drift happens at Qualcomm with all their resources, it happens at every AI chip startup operating with a fraction of the headcount. The pain is universal. The urgency is highest where teams are smallest.
The long-term vision is to become the performance modeling platform for the AI chip era, the infrastructure layer every chip team installs on day one, the way software teams install CI/CD. Starting with model health and expanding into full design space exploration as the customer base matures.
As AI hardware diversifies across custom accelerators, edge inference chips, and datacenter ASICs, the teams building these chips will get smaller and faster, and the pressure to run more experiments with fewer people will only increase. SimDrift is built for that world.