Hits harder but then Bruce lands a more efficient 5.5 GB/s
KICK and gains the edge...
See here:
"Fewer CUs at higher clock rates are better than more CUs at lower clock rates in terms of performance. Because of several factors, the performance never scales 100% linearly with more CUs, but with the frequenz (as in Ghz) if there are no other bottlenecks.
E.g. The Xbox Series X needs more threads to use the machine and not every work instruction is completely independent and scales perfectly.
Then there are factors outside of the CUs, such as the *work allocator* for the CUs, that was a problem especially for GCN and will probably be a bit for RDNA as well. There is a shader processor input block, that initializes registers and tasks for a shader array, but it can only supply a compute unit every four cycles.
The more CUs per SPI exist, the more this HW block limits, because it does not manage to supply all CUs in time and there are more and more waiting cycles.
Based on the maximum clock difference of 22.2%, the PS5 wins simply because it has the same number of special computing units. The front and back end seem to be the same, i.e. 4 rasterizers, 4 prim units and 64 ROPs can be found in both. This means that, in the best case, the PS5 has a 22% higher pixel fill rate and can calculate 22% faster triangles on the part of the fixed function HW.
Another factor is e.g. Cache ratios for the L1 and L2 $ compared to the compute units.
There is 128KB L1 $ per shader array, both probably have four shader arrays, with the XSX then 13 CUs 128KB share, with the PS5 only 9 (or if only scaled via WGPs, two times 8 and two times 10 ). The same applies to the L2 $, which has to supply relatively fewer mouths. (not sure what what the OG meant to say, maybe a google trans. error)
All of this changes the cache-hit-rate in favor of the PS5.
Overall, however, it depends on how much the configuration of the PS5 brings advantages compared to the XSX. Especially if the PS5 "only" clocks with 2.1 or 2 GHz at high load, then 22.2% -> 15.1% or only 9.6% clock advantage, that could hardly or even in practice no longer play a role, at least as far as the fixed-function hardware is concerned, and of course the whole GPU performance scales down.
As it currently looks, there will be one intersection engine per TMU, i.e. four per CU.
36 x 4 = 144 on the PS5 and 52 x 4 = 208 on the XSX.
Similar to the other things, this scales with the beat, so:
PS5 -> 144 x 2.23 = 321.12 billion intersection calculations per second (in the best case, with the maximum clock).
XSX -> 208 x 1,825 = 379.6 billion intersection calculations per second (+ 18.2% compared to the PS5)
A certain bonus for the PS5 arises from the relatively higher L1 / L2 $ capacity, which will probably help with ray tracing. But without benchmarks you will not be able to specify exactly and since the clock of the PS5 can still go down, more caution must be taken about the practical difference of the machines."
posted by Locuza, 3D forum.com