Ye, what Intel showed looks good, if true also with regards to power efficiency.
QuoteWhat makes it suck even more is that rdna2 -> rdna3 -> rdna3.5, weren't at all that big of updates to begin with.
They aren't indeed, it's like RDNA3 iGPU is 10% faster than RDNA2 and RDNA3.5 is below 10% faster over RDNA3.
QuoteThe only reason the igpus got any faster was mostly because they were increasing the CUs. So this feels like 5 years of stagnation.
680M: 12 CUs, 768:48:32:12
780M: 12 CUs, 768:48:24:12
880M: 12 CUs, 768:48:24:12
890M: 16 CUs, 1024:64:32:16
(note that 16 CUs are not 33% faster than 12 CUs, but quite a bit less, and 12 CUs are not 50% faster than 8 CUs)
But I agree it feels totally like stagnation. Maybe they are limited by bandwidth, as benchmarks of 5600 MT/s SODIMMs vs soldered LPDDR5X-6400 MT/s show a performance increase:
Youtube:
- "AMD Radeon 680M iGPU: DDR5-4800 vs LPDDR5-6400 Gaming Performance Comparison" <- doing this calc: 60 FPS*(6400/4800) = 80 FPS, but in the vid, only about 70 FPS is measured, which is unfortunate and not a linear scaling, maybe at that bandwidth there are not enough CUs, latency issue, ...
- "DDR5 SODIMM Memory Scaling 4800Mhz 5600Mhz Comparison [..] AMD 6900HX Benchmarks Gaming"
Strix Halo runs at LPDDR5X-8000 and Apple's M5 uses 9600 MT/s (so this memory is also available), but not sure at which MT/s Strix Point runs at. Of course, the disadvantage is that LPDDR5X memory is soldered, but LPCAMM2 exists fix this.
Imagine the Intel+Nvidia APU is good..welp AMD.