- Pronouns
- He/Him
I meant it in the context of the PS4 since that was brought up and that has way higher Fill-rate than Redacted.I don't think it is different here, from a hardware standpoint. If you're buying a GPU, you care about the performance and the dollar amount you're spending for it.
The 5700XT and the 3060 8G for example have the same ROP count and similar enough filtrate and offer similar performance even with different FP32 and FP16 and the latter having lower memory bandwidth. This being for RDNA based cards.
But for a GCN based card like the Vega 64, which has 64 ROPs like the 3060 8G and it is 98.94 vs 113.7 in fillrate, being 87.01% of the 3060 8G. The Vega 64 has similar FP32 performance but has much higher FP16 performance, not counting TC as those are separate in this and games won’t use those for FP16 unless made with it in mind, it’s only use in games is DLSS. Anyway, the 64 has way higher Memory Bandwidth than the 3060 8G. But, the 3060 outperforms the 64 by about 21%.
It’s possible the high memory b/w is doing a lot of the heavy lifting here to close the gap between the two cards.
Speaking of the cache situation, I think REDACTED might have a System Level Cache like other Orin chips that can help with the small 1MB L2 cache that REDACTED might have on its GPU.
I think it will particularly because to have Point of Coherency you’d need a SLC if it’s the standard ARM, unless they make one that is better than ARM’s like the one that Apple makes in their SoCs.
Other SOCs since 2017/2018 have a SLC to them for the unification and coherency. And Orin while Nvidia could have designed one and had it reside in the L3 cache as that is also possible like with Xavier which houses the PoC+U in the L3, they didn’t on ORIN. And unless Nintendo is perhaps special in which they were able to implement something better than what Arm offers for just T239, I’d expect a SLC even if 4MB.
Maybe that’s why T239 doesn’t have the 4MB of L2, as that would have been redundant for the purpose of the SLC that could be used when needed. Besides the possible reason that @Dakhil mentions in which having a smaller SoC would increase the yield if at SEC, having a 1+4 would still have better yield than a 4+4 and be cheaper too.