Graphics card specs - bus width matters, bandwidth.
Posted by guyza123@reddit | hardware | View on Reddit | 10 comments
I've been doing testing with Forza Horizon 4 on various GPUs. A GTX 1660 super runs it perfectly, so does a RX 6700, but both a RX 6600 and RTX 4060 stutter sometimes.
The only reason that makes sense is limited bandwidth. RX 6600 and RTX 4060 both have a 128-bit bus and GDDR6 ram.
I don't think it's their pci-e 4.0 x 8 either, as I am running them on a pci-e 4.0 system.
Logical-Database4510@reddit
Forza uses insane levels of memory bandwidth because it uses MSAA
guyza123@reddit (OP)
Very useful. I was cranking the MSAA. Thanks.
Logical-Database4510@reddit
Yeah you can easily repeat the same type of tests in a benchmark using furmark and crank the MSAA. If you're bandwidth constrained you'll see your GPU sitting at max usage but performance will be mediocre and it'll be using much less power than it should be because the memory controller is getting the shit kicked out of it trying to super sample targeted portions of the screen to completely insane resolutions. It's a big reason why the industry moved away from MSAA in general; it's just not all that efficient anymore given modern renderers.
Strazdas1@reddit
That actually explains why i had full utilization with extremely low power usage in watch dogs. Heavy use of MSAA in there too. I guess the card ended up bandwidth starved, but at 144fps i didnt have incentive to look for a cause.
BFGsuno@reddit
But actually doesn't matter.
512width bus on older GPU will not usually be better than 128bus on new card. And that's just pure bandwidth without account for efficiency like texture compression decompression etc.
I never understood this fixation some people have on the memory bus. It doesn't matter. Real efficiency matters, then bandwidth.
BFBooger@reddit
People act like the bus width is some sort of key feature. "XYZ is not really a 60 series product because it is a 128 bit bus, it is a 50 series product".
Yet NVidia could have increased the bus to 192 bit, and decreased the L2 cache significantly to recover the die space. In the end, that would be a slower product despite more bandwidth, yet somehow 60 series worthy due to the bus width.
An RX 580 had a 256 bit bus and 256GB/sec bandwidth. A 7600XT has a 128 bit bus and 288GB/sec bandwidth, but is more than 2x as fast.
A 1080Ti has a wider bus and more bandwidth than either an RX 2080 or RX 3070, but is quite a bit slower than both. A 5060Ti 16GB has a 128 bit bus, the same bandwidth as an RX 2080 or RX 3070, yet is 50% faster than a 1080Ti with less bandwidth and a much narrower bus.
We aren't buying bus width or bandwidth, we're buying performance. It doesn't matter if they get that performance from bandwidth or cache or core optimizations or whatever.
The only thing that matters is price/performance within your budget and power constraints.
* estimate based on pre-launch leak info
As we can see, a 9060XT 16GB will likely be almost 3x as fast as an RX 580, even with half the bus width and only 30% more bandwidth.
-Purrfection-@reddit
Yep, performance is performance, doesn't matter if it's achieved by nuking the bus and adding cache or the opposite.
I understand though that bus width is a recognizable metric that stays relatively similar through generations unlike frequency, bandwidth, or core counts, so people latched onto it as the only thing that could 'mark' a card's market position. I'd liken it to the number of cylinders in ICE cars. They don't change that much through the years so you could think it means something yet a 4 cylinder from today can easily beat a V12 from 1975.
crab_quiche@reddit
When DDR5 came out the same people were throwing fits because x32 channels weren’t “real” channels.
SherbertExisting3509@reddit
Bus width is only ONE part of GPU design
The R9 290x had a 512bit memory bus, but its overall bandwidth was constrained by the slow 5.5GB/s GDDR5 speeds.
The 780ti still beat it with a 384bit bus + 7gbps GDDR5
Sopel97@reddit
bus width doesn't matter directly, but yes, 4060 has lower memory bandwidth