|
|
|
|
|
|
by jlawer
1375 days ago
|
|
|
High performance datacenter cores will likely remain a niche for the best chip design companies. Architecture mostly matters in that your workload / ecosystem is supported. I can't see this being a place where an open source RISC-V core wins. Even if you had the complete plans to a top of the range EPYC processor, you would struggle to Fab it, and by the time you were able to do so, it would likely not be valuable. The problem is DC CPUs are driven by power efficiency. Building the most efficient processors requires the latest manufacturing nodes and incredible optimisation. Generally your chips must be within the last 1-3 generations, the inefficiencies are very noticeable. Power efficiency impacts cooling and density, additionally driving the economics of DCs. The existing manufacturers have the teams to churn these out, for someone else to join the club it would be a massive investment (just look at home many failed Server Arm companies there have been). You can also see that even companies that have ARM server chips are mostly just tweaking cores provided from ARM, not developing their own from scratch. Given this I think your unlikely to see this unless the economics change. If one of the cloud giants decided that they aren't getting value with ARM, they might end up developing their own, potentially even using RISC-V as the architecture, but I can't imagine them wanting to share it. Unless you end up with a multi-cloud consortium deciding that its in all their interests to build a common processor in order to turn it into a commodity. In that case you might get an open source core out of it, but the fab problems still remain. |
|
|
The product is still software. If you double the flops, but software is only half efficient in the end you move nowhere. You need performant compilers to go along chip/architecture. You also need user software. A "performant" CPU that does not run upstream versions of postgres/python/java/nginx/whatever is mostly useless for a DC.
A chip design/architecture being a commodity is a good way to have user software and compiler tuning going.