You know what? I'm gonna use a real-world example to show you guys that Switch 2 is likely going to punch above what some of these paper specs suggest.
I've been working on a pet project the past few weeks; basically I've been trying to develop a retro-style game console design, but simulating 5th-gen gaming hardware performance. It's actually pretty hard to do because even many of the low-end ARM-based chips available for purchase are way above the high-end spec of gaming systems from the mid to mid-late '90s. I've found some solutions to that though but I'll just focus on one part: the Analog Devices ADSP 2116N.
The reason I've chosen that DSP in the design is both because it's capable of floating-point calculations, and also because the SEGA Model 2 arcade board (specifically 2B-CRX) actually ALSO uses an Analog Devices DSP; 2x 2106N DSPs to be exact. The initial Model 2 board used 6 (!!) Fujitsu DSPs for its geometry processor, but they were clocked at 16 MHz each. Anyway, I'm bringing p the Analog Devices DSPs because I somehow stumbled into picking the 21161N and only later found out the SEGA Model 2 used a very similar SHARC-based DSP from the same device line.
I say similar...but there are some major differences. The 21161N didn't come out too long after the 2106N, but it was priced cheaper and had a lot of advantages. Then again, the 2106N has advantages too, but I'd say the 2116N is overall more efficient despite being weaker in some notable areas. For example, the 2106N has a 64-bit data bus, and SEGA used two of them for the Model 2, so in a way it has a 128-bit (2x 64-bit) data bus. The 21161N only has a 32-bit data bus, by comparison. The 2106N has a 32-bit address bus and can address up to 4 Gigawords (~ 4 GB) of memory, whereas the 21161N only has a 24-bit address bus and can address upwards 256 MB of memory. The 2106N also has 4x link ports (ports allowing communication with other DSPs on a bus with chip selection) at 10-bit width each, while the 2116N only has 2x 8-bit link ports.
But from there, the 2116N has many advantages. Its FIR Filter rate is 8ns vs the 25ns the 2106N has. The 2116N is over 350% faster with inverse square roots than the 2106N; even a setup with 2x 2106Ns cannot overcome that advantage. It has a 61% faster X/Y division rate than the 2016N and, while not being able to address as much off-chip memory in its unified memory space, the 2116N can facilitate up to 256 MB of SDRAM whereas the 2106N doesn't have any off-chip memory support aside from SRAM (granted, since it has a 64-bit data bus you can have a 32-bit peripheral connected as slave device to the external port and still have a 32-bit data space for off-chip SRAM).
The 2116N has other advantages over the 2106N beside those, but the biggest is that it supports higher clock rates. Like Mark Cerny said, when your clocks are faster, everything inside (based on the internal clock) runs that much faster. The max clock speed for the 2106N is 40 MHz, whereas the 2116N can support up to 110 MHz clocks. What's more, there's a linear relationship between some of the benchmarks I mentioned earlier and clock rate, but even if you were to lower a 2116N's clock to match even TWO 2106Ns at the same clock, the 2116N outperforms them by healthy margins in FIR filter, X/Y division, inverse square root calculations etc. Heck, you can probably lower the 2116N's clock to 20 MHz and at worst trade par with 2x 2106Ns at 40 MHz each, at least with those aforementioned calculations.
Why did I bring any of that up? Well, because I wanted to show how even newer lower-tiered options in the same device family can trade blows with or outperform the high-end options of just a generation prior. Now those DSPs are '90s tech and '90s tech advanced much faster than tech does today, I know that much. But it's been a few GPU gens between Series S and Switch 2, and a lifetime's worth between it and PS4 Pro. Never mind, that it's on a very different architecture than Sony/SIE and Microsoft systems and, for gaming, Nvidia's generally been considered the better of the two in terms of performance-to-dollar (in the bracket segments where AMD and Nvidia actually compete) for at least a decade now.
So where are these opinions the Switch 2 needs 4 TFs to be on PS4 Pro or Series S level coming from? That it needs memory bandwidth equivalent to Series S for same performance when docked? It won't need as much bandwidth if it's better at data compression, and wasn't there just a patent leak/reveal weeks ago with Nintendo/Nvidia and data compression? Do people think those tensor cores are going to sit idle on the GPU and do absolutely nothing? Some folks are making the exact same mistakes they made with Series X and PS5 4.5 years ago, ignoring the more intricate features of the PS5 (like its cache scrubbers) either because they don't know how it works or they were actively trying to form a narrative. Look how well that's turned out for Series X all these years later :/
Switch 2 isn't going to get near the performance of Series X or PS5; it was never meant to and I don't know why there are people who thought it would or felt it should. But now some of the same people are trying to imply it won't hit PS4 Pro/Series S levels when docked, without knowing any of the intricacies of the SoC design, or honestly, even understanding intricate aspects of memory, memory controllers, cache, or how CPUs and GPUs work at the foundational level. Whether that's just out of ignorance or them willingly ignoring those things just to push FUD or some agenda, I don't know.
But it's just kinda funny seeing all the doom & glooming, the downplaying, the theatrics. It's pretty good for a laugh.
Buggy Loop
as a minimum performance test considering differences between Nvidia and AMD in that regard
