I've tried to make sense of the PS6 ~9070XT estimate by looking at Navi 48 9070 vs 9070XT. Let me first start with establishing some prereq info (skip ahead if you want results)
Prerequisites for analysis
Both dies are identical in terms of Frontend, RB, L2, LLC, and memory config only differing in shader config and shader and RB clocks. IIRC AMD hasn't disclosed infinity cache clocks so I'll assume they're identical.
Numbers have been sourced from TPU's 9070 review:
https://www.techpowerup.com/review/powercolor-radeon-rx-9070-hellhound/
Results normalized to 9070 = 100
Specs and perf
| v / > | RX 9070 | RX 9070 XT |
| CUs | 56 | 64 |
| SE (RB+3D FF+Frontend) | 4 | 4 |
| Clk avg 25 games | 2700* | 2982 |
| TFLOPs adj_clk | 100 | 126.2 |
| RB+3D FF adj_clk | 100 | 110.4 |
| FPS 4K avg | 100 | 112 |
| FPs 4K RT avg | 100 | 115 |
* = -50mhz due to mild factory OC
9070XT has massive avg gaming compute lead of +26.2% vs 9070. For overall frontend and non-shader FF resources lead shrinks to +10.4%.
Not surprisingly RT which is more compute bound sees larger gains than raster.
Benchmark numbers (4K native)
| v / > | 9070 | 9070 XT |
| Blender* | 100 | 117.6 |
| Cyberpunk 2077 | 100 | 111.6 |
| Stalker 2 | 100 | 110.9 |
| Alan Wake 2 RT | 100 | 114.9 |
| Cyberpunk 2077 RT | 100 | 119.2 |
* = rendering/compute application
Difference confirmed here. Perf scaling is higher with compute bound workloads.
Node PPA
| v / > | clk | pwr |
| N5 -> N4P/N4C | +11% | -22% |
| N5 -> N3B | +10-15% | -25-30% |
| *N4P/N4C -> N3B | -1 - +4% | -4-10% |
| N3B -> N3P | +10% | -20% |
| *N4P/N4C -> N3P | +9-14% | -23-28% |
*Not official number but estimates based on official numbers
With N4C -> N3P a shrink of 9070 matching 9070XT clocks without increasing TDP is easily doable.
PS5 power GPU draw
PS5 is quoted as having 92% efficiency around 200-230W. Source:
https://www.techpowerup.com/review/playstation-5-power-supply-adp-400dr/5.html
Internally that's 184W - 212W. 10W for blue-ray, 10W for misc and 20-30W for CPU seems reasonable. 40-50W which leaves 134W - 172W (including mem PHYs) for PSU GPU.
PS6 rumoured specs
52 CUs (54 CU full config)
3 SEs
2982mhz clock estimate (matching 9070XT)
PS6 raster IPC extrapolation
I assume PS6 only has 3 Shader Engines and assuming no changes to number of RBs and Rasterizer footprint (brute force).
If we normalize to 9070XT clocks it has 33% lead.
If we look at compute throughput 9070XT is 23% ahead.
Whether IPC gain or countering a cachemem bottleneck (L2 + LLC (not in RDNA5) and memory) to catch up with 9070XT in raster requires 23-33% higher performance or assuming a mixed workload likely 25-30%.
Matching 9070 (-10.7% perf 9070XT) only requires 9.8-18.8% higher IPC, or assuming a mixed workload around 12-17%.
PS6 matching ~9070-9070XT raster perf requires anywhere from 12-30% IPC gain vs 9070XT.
9070 -> 9070XT perf scaling is not 100% perfect so the range can prob be lowered a bit. Maybe
10-25% IPC gain should do the trick.
Reversing PS6 power draw and power efficiency
9070 has 220W power draw. I assume PS6 has same power draw and PSU as a PS5. I'll adjust CPU power draw to two scenarios: low power CPU 15-20W or PS5 CPU (20-30W).
GPU needs to hit 144-177W power draw or prev numbers (134-172W).
New node equals out increasing clockspeed so let's use that 220W as a starting point. Removing LLC+96bits from bus + GDDR7 could reduce power draw by 20-30W, maybe more (not sure) = 190-200W.
Perf/watt (low wattage CPU) = +13-32% (9070 perf) or +26-48% (9070 XT perf)
Perf/watt (PS5 CPU wattage) = +16-42% (9070 perf) or +30-59% (9070 XT perf)
I've set up math to find ends of the spectrum. Power draw for non-GPU can fluctuate a lot so the extrapolated perf/watt gain would prob be somewhere in the middle.
Caveat: PS6 will target lower clocks than RDNA5 ATx silicon (power constraints) assuming it has same design goals as RDNA4 (high clocks) which should help reduce power draw. Also suspect the estimate for iso-power with 9070 -> 9070XT clock increase with the node shrink from N4P -> N3P is pessimistic. The new node can prob drive power down a bit also. Both should make it easier to hit perf/watt targets (performance (9070-9070XT raster) within power budget (PS5), even if µarch falls short of my perf/watt estimates ^. That's a result of the node impact watering down the purely µarch perf/watt gains. Fortunately for AMD that lowers the target they needs to hit.
End
All this should help drive down power at iso-clocks (9070XT), or push µarch raster IPC high enough that AMD can get away with lower clocks. Overall this seems realisticand reasonable given how significant many of the changes in RDNA5 are.
Conclusion
~9070XT raster for a significantly smaller GPU config with far lower power budget is very ambitious even when we factor in the node shrink. With that said if RDNA5 is a very good architecture then it should be possible to hit the target and easily achievable if we lower the perf target by 5-10%.
Also remember that this is only the starting point as RDNA5 tailors scheduling and execution to the Work Graphs pipeline. This should help drive further perf gains in addition to any ML and PT gains in the second half of its life (including PS6/PS7 crossgen).
Sorry for the long post. I hope at least some of it makes sense xD
Facebook
Twitter