msroadkill612
Posts: 116 +43
The ~unsung heroic AMD elephant is Infinity Fabric.
There are many amd barbs & some luck which get popular credit for this ascendance, but their focus on Fabric has made it the secret sauce that moderates chiplet's innate latency weaknesses, & makes cost effective chiplets, competitive with monolithic perf.
Unlike that other time Amd had a killer edge, Fabric has been a decade plus hard slog. It is not imitable (along w/ a all new range of processors & complete re-validation) in a timely way. IF, by the time intel can legally clone it, they will be enfeebled.
Fabric gives Amd a clear run of at least 5 years.
In the historical context of the article, the philosophical aspects of Fabric & chiplets are pivotal.
AMD faced oblivion at the time. What hope had such a small, weak company against such a strong foe, & they only had a budget for one shot at a product, yet they needed a range of very different processors, & sophisticated servers were a must - they were lucrative & an incubus for lesser products.
How could one product serve all markets?
The answer was of course to build big cpuS using multiple small cpuS as lego blocks. Make one cpu to satisfy multiple markets, but figure how to group them in multiples with coherent caches (via Fabric).
Even they were surprised how many advantages this approach had. Necessity proved very good mum for chiplets.
There are many amd barbs & some luck which get popular credit for this ascendance, but their focus on Fabric has made it the secret sauce that moderates chiplet's innate latency weaknesses, & makes cost effective chiplets, competitive with monolithic perf.
Unlike that other time Amd had a killer edge, Fabric has been a decade plus hard slog. It is not imitable (along w/ a all new range of processors & complete re-validation) in a timely way. IF, by the time intel can legally clone it, they will be enfeebled.
Fabric gives Amd a clear run of at least 5 years.
In the historical context of the article, the philosophical aspects of Fabric & chiplets are pivotal.
AMD faced oblivion at the time. What hope had such a small, weak company against such a strong foe, & they only had a budget for one shot at a product, yet they needed a range of very different processors, & sophisticated servers were a must - they were lucrative & an incubus for lesser products.
How could one product serve all markets?
The answer was of course to build big cpuS using multiple small cpuS as lego blocks. Make one cpu to satisfy multiple markets, but figure how to group them in multiples with coherent caches (via Fabric).
Even they were surprised how many advantages this approach had. Necessity proved very good mum for chiplets.