At Intel’s Architecture Day back in 2018, the company spoke about design changes it was making in future products. Going forward, we were told, Intel would design CPUs in a way that allowed it to deploy them flexibly, rather than strictly locking a given CPU design to a process node. While the company didn’t say it publicly, the idea that Intel might take a 10nm CPU and backport it to 14nm was already being floated as a solution to its woes.
During an AMA for the Rocket Lake platform, Intel revealed that the CPU backport project that created Cypress Cove was finalized in Q1 2019. The reason Intel targeted Ice Lake’s Sunny Cove processor as opposed to the more advanced Willow Cove chip inside Tiger Lake is that the TGL design wasn’t finalized yet. Rocket Lake shipped at the tail end of Q1 2021, which indicates it took Intel roughly two years to build the new floorplan and backport the core.
This timeline offers a window into how chip design cycles work. In Q1 2019, Intel’s top-end chip was the 9900K and the company still had its own solid stack of performance wins at the top of the CPU market, especially in gaming. Intel greenlit Cypress Cove and began working to backport Ice Lake to 14nm while simultaneously teeing up the Core i9-9900KS (launched October 2019) and the 10th Gen Core i9-10900K (launched May 2020).
This is how semiconductor manufacturing tends to work. AMD is currently shipping Zen 3, finalizing Zen 4, and working on Zen 5. Intel just launched Rocket Lake with Alder Lake coming later in 2021 and a 7nm successor reportedly intended for 2023. This means Intel and AMD are both picking design features and targets based on what they think the competitive situation will be 1-2 years later.
The AMA states that an eight-core die represented the largest die + UHD graphics that Intel could manufacture without clarifying if this is related to the design of the LGA1200 socket or some other product limitation. The years since Ryzen’s launch have, we think, illustrated factual differences in how AMD approaches product design versus Intel.
While not every AM4 motherboard supports every Ryzen CPU, AMD has demonstrated that it designs its platforms for a longer life cycle and better upgrade paths than Intel does, despite the vast differences in their respective incomes and resources. If AMD was able to engineer AM4 to support both a switch from monolithic dies to chiplets with a central I/O die and a doubling of CPU core counts, then there’s no reason for Intel to perpetually discover that its own products run out of headroom so easily — unless they’re intended to. We continue to suspect that the limiting factors on Rocket Lake were thermal and power-related. Intel is absolutely capable of manufacturing larger chips than an eight-core RKL + integrated Xe GPU. Claiming that it can’t, with no reference to what the limiting factor is, sends an erroneous message about Intel’s manufacturing capability — unless you think the same company that manages to build a 28-core monolithic Xeon can’t manage a 10-core desktop CPU with an IGP.
Pat Gelsinger has pledged to bring Alder Lake to desktop first before the end of 2021. If Intel keeps to this schedule, it’ll be the fastest desktop platform phase-out in history. Rocket Lake compares better lower down the stack, but the top chip is competitively fast and non-competitively thirsty where power is concerned.
Part of what we see here is an interesting example of how a company can make one set of reasonable-looking decisions in 2019 and find itself hard-pressed with the final product two years later. But the fact that Alder Lake is coming so quickly after Rocket Lake also speaks to the size and strength of being Intel. Intel has sustained multiple engineering teams working on multiple chips to the point that it can (maybe) launch a new desktop replacement CPU less than a year after launching the first one. That kind of turnaround represents its own insurance to a poor launch, and it’s not something we see happen very often.
The closest analogy I can think of would be in June and August of 2002. In June, AMD launched a 130nm Thoroughbred core Athlon XP, the so-called “T-bred A.” The chip ran hot and didn’t get AMD much of a clock boost. Two months later, AMD unveiled Thoroughbred B, a new spin of the chip with a 9th metal layer and significantly higher frequencies. T-Bred A CPUs topped out at 1.8GHz, but Thoroughbred B could hit 2.25GHz, and it put AMD on a stronger footing against Intel headed into the back half of the year. In any case, Rocket Lake occupies an odd spot in Intel’s pantheon of desktop CPUs and, perhaps, a bit of an object lesson in the difficulties of adapting a CPU design to a process node it wasn’t built for.
Source From Extremetech
Author: Joel Hruska