WEBVTT 00:00:00.000 --> 00:00:05.160 One of the things people got outright angry about when the now-canceled RTX 408012 gig 00:00:05.160 --> 00:00:12.160 from NVIDIA was announced was the memory bandwidth. A spec you might also see described as bus width, or something similar. 00:00:12.160 --> 00:00:19.080 But why exactly is it so important? Is it some ultra-wide wheeled vehicle that carries your data to and fro, assuming they 00:00:19.080 --> 00:00:23.640 have a permission slip? Well, actually, that questionable analogy isn't far from the truth. 00:00:23.640 --> 00:00:28.760 You see, a bus in computing is a general term for a data link between two components. 00:00:28.760 --> 00:00:33.920 For example, USB stands for universal serial bus, because it's supposed to be a do-everything 00:00:33.920 --> 00:00:41.200 connection for transmitting data. But in graphics card land, the memory bus connects the GPU itself to your card's video 00:00:41.200 --> 00:00:47.520 memory or VRAM. The basic way a graphics card works is that the GPU, which you can see here, acts upon 00:00:47.520 --> 00:00:52.480 data stored in the VRAM, which are these little modules, to generate the images you see on 00:00:52.480 --> 00:00:57.680 your screen. Among other things, VRAM contains the textures that make the objects on your screen look like 00:00:57.720 --> 00:01:01.160 they're supposed to, as well as data used to create visual effects. 00:01:01.160 --> 00:01:05.960 The higher the resolution, texture quality settings, or VFX settings, the more VRAM 00:01:05.960 --> 00:01:10.800 your GPU will use, which is why you sometimes see people criticize NVIDIA or AMD when 00:01:10.800 --> 00:01:15.920 they pair a powerful GPU that could otherwise run games on high settings with an insufficient 00:01:15.920 --> 00:01:23.640 amount of VRAM. Unacceptable! And the more data that sits in VRAM, the wider memory bus you need to get that data to your 00:01:23.640 --> 00:01:28.880 GPU. But how do you read a spec sheet that gives a number of bits, or gigabytes per second 00:01:28.880 --> 00:01:32.640 for your memory bus, and how do you know if that number is high enough? 00:01:32.640 --> 00:01:37.400 So the bit width is just telling you how many bits the bus can move for each cycle of the 00:01:37.400 --> 00:01:42.320 memory clock. Basically, how much data is being moved per one hertz. 00:01:42.320 --> 00:01:46.560 Of course, seeing as how memory speeds on modern cards are typically well over a thousand 00:01:46.560 --> 00:01:49.800 megahertz, this adds up to quite a lot. 00:01:49.800 --> 00:01:56.000 For example, the top-end RTX 4090 has a bit width of 384 bits, but this translates into 00:01:56.000 --> 00:01:59.840 around 1 terabyte per second of total memory bandwidth. 00:01:59.840 --> 00:02:03.640 Now most of you aren't going to be buying a daddy warbucks-level graphics card, so 00:02:03.640 --> 00:02:07.880 how much does bus width matter as you go down the product stack? 00:02:07.880 --> 00:02:12.640 The answer is, not a whole heck of a lot, without additional context. 00:02:12.640 --> 00:02:16.960 Just like it's unwise to compare raw clock speeds between different generations of CPUs 00:02:16.960 --> 00:02:21.080 because their underlying architectures are so different, you can't simply look at the 00:02:21.080 --> 00:02:26.360 raw bus width for your VRAM and decide that a product is over or underpowered. 00:02:26.360 --> 00:02:32.000 It's not that simple. A good recent instance of this was the reaction to the 12 Gigabyte version of the NVIDIA 00:02:32.000 --> 00:02:37.040 RTX 4080 prior to Team Green pulling the plug on the card completely. 00:02:37.040 --> 00:02:40.760 Although the criticisms of the card being positioned as a 4080, despite the fact that 00:02:40.760 --> 00:02:45.040 it had lower-tier specs, were completely warranted, some of the anger at the memory 00:02:45.040 --> 00:02:53.520 bus specs had us raising an eyebrow. That first it seemed silly that the 408012 Gigabyte version was shipping with a 192-bit 00:02:53.520 --> 00:02:57.480 memory bus, narrower than the previous gen RTX 3070. 00:02:57.480 --> 00:03:03.040 However, with the higher power draws and thermal output of the Ada Lovelace GPU series, the 00:03:03.040 --> 00:03:08.600 408012 gig was to get a massive 48 megabytes of level 2 cache, which is extremely fast 00:03:08.600 --> 00:03:13.400 memory that sits on the GPU die itself and is separate from the main VRAM. 00:03:13.400 --> 00:03:17.520 This might not sound like much, but it's actually a huge leap from the 6 megabytes 00:03:17.520 --> 00:03:20.520 that the previous flagship 3090 Ti came with. 00:03:20.520 --> 00:03:25.520 The idea is that substituting VRAM bandwidth for cache is easier on your card's power 00:03:25.520 --> 00:03:30.040 and thermals, and we might see this become more common in the future as we try and squeeze 00:03:30.040 --> 00:03:34.520 more and more performance out of the GPUs themselves. Okay, so that was a lot. 00:03:34.520 --> 00:03:39.160 Here's our point. A wider bus and more memory bandwidth is better. 00:03:39.160 --> 00:03:44.880 That's true. However, as always, you can't rely just on the spec numbers when buying your next graphics 00:03:44.880 --> 00:03:50.040 card. Instead, read reviews and benchmarks and specifically look to see if there are reported problems 00:03:50.040 --> 00:03:54.200 with textures or visual glitches at higher frame rates and quality settings on cards 00:03:54.200 --> 00:04:01.880 that have relatively narrow memory buses. Prioritize a strong GPU and having enough VRAM and worry about memory bandwidth only 00:04:01.880 --> 00:04:08.160 if something seems off when looking at test results. Besides, there are already enough specs to keep track of anyway, like if the card you 00:04:08.160 --> 00:04:12.840 want will even fit in your case. Thanks, NVIDIA. 00:04:12.840 --> 00:04:19.000 And thank you for watching this video, earnestly. Like the video if you liked it, dislike it if you disliked it, check out our other videos, 00:04:19.000 --> 00:04:22.320 comment below with video suggestions, and there's one more thing I'm gonna ask you 00:04:22.320 --> 00:04:25.680 to do, okay? Can you subscribe and follow to TechWiki? 00:04:25.680 --> 00:04:28.960 Just this one time, okay? Just a little bit. Can we see if we can do that? 00:04:28.960 --> 00:04:29.960 Okay, thanks.