Competitive Balance for Compute Architectures

Since our last post about the Apple Unleashed event, the landscape of consumer-focused compute has achieved a level of competitiveness that hasn’t been experienced since the last time when AMD and Intel were trading blows over a decade ago. While there will be some time to come before the swan song for AMD’s AM4 platform is generally available through the re-spin of Zen3 CPUs with additional cache, Intel’s Alder Lake solutions overcome the extensive malaise of adding plusses to a 14nm platform. The benchmarks available through a myriad of sources paint a mixed picture for the 12th Generation Core series of processors. The support for DDR4 and newer DDR5 memory doesn’t provide a universal win for the outgoing or incoming memory standards. Early leaked benchmarks of the mobile variants of Alder Lake should be taken with a grain of salt due to the benchmark tool utilized and the non-final state of firmware for the applicable platforms. There may also be a need for further optimization of the scheduler within the OS for lower performance core count SKUs.

AMD’s datacenter-focused presentation that highlighted Genoa and Bergamo can provide clues as to how their future-state performance profile and associated downstream processors will fill out and compete against the approach that Intel and Apple are utilizing. The announced design win for the artist formerly known as Facebook involves a very compelling solution that provides a large number of cores (36) in a low-power package (95W TDP). The Xeon Silver lineup and upper tiers of the Xeon-D 2300 series have dominated this space. The advantages that Patrick Kennedy has noted when comparing the AMD offering to a custom Xeon Platinum SKU that Intel created for Facebook are achieved with only a minor uptick in power requirements. The expectation that Intel will be able to fire back and retain some of the business in this space with a custom Sapphire Rapids variant means that both providers of x86 processors cannot take their foot off of the proverbial gas pedal.

As for the M1 Pro and M1 Max, there are still some minor yet not insignificant teething pains within Apple’s platform. The review over at Hardware Unboxed takes a much different look at the platform as a whole when compared to Mac-specific sites. Even with the M1 Pro, which has a myriad of deficits when weighed against the M1 Max, the use case for creative workloads and productivity improvements at a fraction of the power draw incurred in the x86 ecosystem provides a compelling story for consideration. Setting aside the “Apple Tax” concerns for a moment, the performance warrants consideration and may justify acquisition of an appropriately configured MacBook Pro for those still using any of the x86 models. Rosetta 2 performance is still sufficient for running legacy software which has not been ported to a universal binary or properly optimized for M1. Apps which are very well optimized, such as Davinci Resolve 17.4 or newer, provide massive performance gains which are harder to match using traditional desktop solutions. Outputs of edited 720p, 30fps content that has an original run length of ~21 minutes are rendered using the hardware accelerated options complete in one minute and thirty seconds. This is a dramatic uplift over prior efforts on M1 Macs with 8 GB RAM, which in turn was an improvement over a 2020 iMac that contained a Core i9-10910 CPU and Radeon RX 5700 GPU.

Our expectation involves all three companies (Intel, AMD, and Apple) continuing to advance their hardware designs while placing even more effort on the operating system and software elements to enhance their platforms. The inconsistency and reluctance that Intel continues to present for AVX-512 and newer instruction sets in consumer platforms runs counter toward the merits of their OneAPI sales pitch. Based on historical precedent from Apple’s transition from Power PC to Intel x86, the expectation that Rosetta 2 will remain over an extended period of time is unrealistic. The lack of gaming titles due to the costs of porting titles developed for DirectX 12 or Vulkan to Metal does limit the uptake and potential market for Apple Silicon. More work needs to be done on the software side of the house to reduce dependencies on a given platform-specific technology. These types of investments in software and associated cross-platform portability need to gain focus from hardware manufacturers. With the continued stress on silicon supply chains, the investments in software-optimization and enablement of cross-ecosystem capabilities will aid in easing the burdens which will continue to persist over the next few years.