itsmydamnation
Veteran
but GB203 is also a larger die, and at the same size AMD probably would have around the same shader count.
The AMD 9070 / 9070XT Reviews and Discussion Thread
- Thread starter Albuquerque
- Start date
Arch. Same node as nvidia
Bondrewd
Veteran
It's a product strategy, same as their CPUs.
itsmydamnation
Veteran
Pushing hard to get as close to the 5070ti as possible, if you lock to vsync and aren't 100% loading the gpu 5070ti and 9070tx are equal power draw.
I think it the right call performance and performance per dollar are about 18 orders of magnitude more important.
9070XT is pushed too hard to get as close to the 5070ti as possible. The 9070 is fine, but there are some strange discrepancies between different reviewers.
Interestingly, with a 144FPS limit, Computerbase shows that these are the outright most efficient cards, by a massive margin at 1440p:
https://www.computerbase.de/artikel...-9#abschnitt_energieeffizienz_in_fps_pro_watt
Yes GB203 is 6% larger and about 20% faster in raster, more in RT while using less power.
RDNA 4 down clocks and powers down quite nicely when capped. Per TPU it drops to 1450Mhz @ 0.68v capped to 60fps in Cyberpunk. For comparison the 5080 only drops to 1700Mhz @ 0.86v. AMD has a very nice voltage curve there.
Man from Atlantis
Veteran
Yeah this is huge improvement over RDNA3, even better than Ada GPUs. Should be great on laptops.
4K@144Hz
4K@144Hz
These metrics are not correct. Nvidia calculates these metrics from the lowest guaranteed clock while AMD calculates from a clock much closer to max. At actual clockspeeds the 5070ti is ahead in every theoretical metric aside from fillrate. 5070ti has roughly 10% more compute and texture throughput in addition to the huge bandwidth surplus.
You’re right. I used TPU numbers and they use advertised boost clock to calculate metrics for both AMD and Nvidia. Nvidia’s advertised boost is way more conservative than AMD’s. Actual clocks are ~3125 for the 9070 XT vs ~2775 for the 5070 Ti.
That gives the 9070 XT a 3% advantage in compute. How did you arrive at a 10% deficit? The fillrate advantage for the 9070 XT is 50% at those clocks. I think my point stands.
If you want to discuss uarch efficiency you can’t discount the GDDR7 advantage and ignore everything else.
3125 is a max OC clock. Average clock is going to be slightly above 2800.
What are you basing that on? HUB got 3125 on the Sapphire PURE. TPU got around 3200 on the Nitro. Advertised boost is 2970.
On reference clocks. You can't compare AIB boards with a factory OC vs reference clocks on an Nvidia GPU.
What’s your source for 9070 XT reference clocks boosting to only 2800? HUB has very different numbers.
All of the models HUB tested have a factory OC. They also only showed speeds running TLOU. Given how poor that port is, it's possible lower utilization/power results in abnormally high clocks. Computerbase has a larger sample. Their card too has a small factory OC though. The 5070ti can also be overclocked to have 3000+ sustained in games. Both of these GPUs clock very similarly.
Ok not sure how you got 2800 from those numbers. TPU also shows the clock ranges across 25 games and the average.
They reported an average of high 2800s. Minus the small OC and low 2800s seems like a very fair estimate.
It shouldn’t be surprising that AMD has a more graceful architecture for traditional workloads. Nvidia hasn’t bothered to improve this aspect much for 3 generations now.