GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.

nVidia always had more efficient internal infrastructure. It is not just bw.
They've also always had a better memory architecture. AMD chips, including their CPUs, are always far below their rated bandwidth compared to NVIDIA or even Intel or Power. They've usually had more memory channels and L3 cache than Xeon as of late, though, so they still win.

Nvidia's reputation for reliability is overblown and outdated.
They're digging HBM chips out of the gutter these days to make H100 & H200 shipments. Just absolute garbage-quality hardware is being shipped to datacenters these days; you're lucky if your DGX doesn't explode into flames after 18 months.
 
No, these benchmarks are misleading. Typical laptop battery use doesn't involve spinning all the compute units as hot as you can with the most demanding game you can run and waiting for the battery to drain. It's extremely bursty with lots of idling. IME AMD CPUs tend to run at higher frequencies and suck a lot more power than Intel CPUs at idle. Just a broad trend.

The Laptop Mag web surfing test is much more representative of typical energy consumption while on battery and thus a lot closer to the results you will see in the real world.

However, I can make my ASUS laptop last longer with better software, I'm all ears. I have a Ryzen 7 6800U that refuses to run below 2 GHz most of the time unless I limit it in the ASUS software. Even then, I've never gotten over 90 minutes.
I have a 7940hs its pretty efficient using at most 5-6w while browsing the web. I was fixing a Lenovo laptop with a 13500h and even heavily limited it it wouldn't go lower than 7-9w. Unfortunately there arent many live amd vs intel comparisons like we have for smartphones or even the m chips vs intel. So I cant prove 100% what I'm saying as I don't have an intel laptop at the ready.
 
What gets me is that the 5050 (GB207) is 149mm^2 for 2,560 CUDA cores, while the 5060 Ti (GB206) is 181mm^2 for 4,608 cores. That's 80% more cores for 21.5% more die area.

The memory controller and cache takes up a lot of space, and you can't really scale it down without reducing buswidth. I think that is what is going on here. But appears to be a shitty card
 
The memory controller and cache takes up a lot of space, and you can't really scale it down without reducing buswidth. I think that is what is going on here. But appears to be a shitty card
That's what I gathered. It's not just a shitty card (generally slower than 4060 while using more power), the die it's based on is imbalanced and that drives up the cost.

Maybe they can afford to sell the 5050 for $200, but the limits of the low-end are in sight if the price floor is too high. An even cheaper 64-bit Blackwell spiritual successor to the 6500 XT could be impractical, with poor price/perf, and panned even if it had 8 GB. And 12 GB GDDR7 on 64-bit would defeat the purpose.

But it does have DLSS multi-frame gen support so everything I said was wrong and we should all run out and buy one. It's as good as a 4070 Super.
 
How would you put 12 GB gddr7 on a 64 width bus?
You can put 4 chips on 64-bit in clamshell configuration (like the rare 8 GB version of the 6500 XT) and 8 chips on 128-bit (9060 XT 16 GB or 5060 Ti 16 GB).

GDDR6 maxes out at 2 GB chips, while GDDR7 has 3 GB (standard allows 4 GB, 6 GB, 8 GB in the future, but only if memory manufacturers make it).

So 12 GB is technically possible on 64-bit, 24 GB on 128-bit, 36 GB on 192-bit, 48 GB on 256-bit. But those aren't going to be seen outside of pro cards because gamers don't need that much.

We will see:
5070 Super 18 GB (192-bit)
5070 Ti Super 24 GB and 5080 Super 24 GB (256-bit)

The RTX PRO 6000 Blackwell Workstation Edition uses a 512-bit bus, 3 GB GDDR7 chips, and clamshell to reach 96 GB.
 
Back