AMD did manage to improve the power efficiency over RDNA3: At least 28% (100%/78%) more power efficient.
Looks like this is the non-XT 220W TDP version. What's the TDP of this OC card? (power efficiency could be higher if this wasn't OC) Considering this, the Raytracing power efficiency improvement is decent and AMD has caught up to NV in the Raytracing benchmark (about as fast as a 4070 Super (220W TDP) / 5070 (250W TDP, this one looks less power efficient just form the TDP, but idk the raytracing vs raster weighting of the
benchmark).
Wonder about the PT performance,
as AMD's GPUs are much slower in path-tracing:
- Bad result in Indiana Jones And The Great Circle (1440p, Very High, Full RT).
General
issue:
Only HDMI 2.0 support on Linux.
Remaining questions: I wonder if the
power scaling is fixed and need to check out FSR4:
- The first FSR4 upscaling testing looks good.
- In this test (4K Performance upscaling), FSR4 looks better than DLSS CNN model (which already is relieving and now AMD's upscaling is not a dealbreaker anymore), but worse than DLSS transformer model, though the transformer model has its own issues, as is mentioned in the vid.
A higher VRAM consumer version
for AI LLM self-hosting/inferencing (to fit the full SOTA model Qwen2.5-32B-Instruct-Q6_K.gguf, which is 27GB + context, at least 32GB VRAM are required) or for gaming
where 16GB VRAM are not enough, would be nice:
- A 24 GB VRAM (3GB per GDDR6 chip, instead of the current 2GB, if such chips are available (they will be for GDDR7), though I heard of GDDR6W), or preferably
- A 32 GB VRAM (clamshell VRAM design: using 2GB GDDR6 chips, just like now, but on both sides of the PCB), or even more preferably
- A 48 GB VRAM (clamshell VRAM design + using 3GB GDDR6 chips)
I'd prefer the higher VRAM consumer version to be based on the 220W TDP chip, if the performance impact vs the 304W TDP is minimal: When I power limit my GPU to -50% to 100W, the tokens per second speed difference vs the full 200W is only like 3%.
.