Steam Machine takes on Sony PS5, Xbox Series X and Steam Deck in specs comparison

[Andrea]

I'm sorry...but looking at your 7600M page there is a theoretical peak of 90W and ~17TFLOP with 32bit FP. In this article you state that the Steam Machine has 8.9TFLOP and it's roughly 5x the Steam Deck when:

1. Valve Talks about roughly 6x in "real world" performance (whatever it means)
2. These estimates can't be applied to the TFlops of one device and just do 5x...it's more nuanced than that.

Another last bit...at 1080p and with competitive games the CPU is very important...in many games this machine will perform better than consoles, unless we suddenly change the way of reading data from one day to the other. Considering that Valve has no released any TFLOP data...either report the 7600M data or not make that comparison at all, it feels like a fabricated information at this point...

[Dual Issue]

...it's more nuanced than that.

The doubling of TFLOPs is if you use the dual issue capabilities of the rdna3 compute units, which most games don't use.

Same thing happened when we went from 680M -> 780M. We went from 3.4 to over 8 TFLOPs but performance difference between them in actual games was within ~10% margin. So it behaved more like a 4 TFLOP or half it's rated flops.

So yeah, this 7600M in steam machine might as well be 8.9 TFLOPs as that would be more accurate. It's slightly less powerful than the PS5 GPU, which is rated for 11 TFLOPs. The only thing is it has less vram which may lead to lower quality textures in games and is why some people gauge it to be between Xbox series s and ps5 (although it's a lot closer to the later).

...unless we suddenly change the way of reading data from one day to the other.

But that's exactly what they did on the ps5? NXgamer did a video about this, how the zen2 cores in that console perform closer to a windows pc in games with zen 3/4 cpu.

Now granted this steam machine won't be running windows which is super bloated but it's still going through proton translation layers and not going to be as efficient as bare metal native optimization.