1 00:00:00,000 --> 00:00:03,760 Have you ever taken your shiny new graphics card home only to be disappointed at how you 2 00:00:03,760 --> 00:00:07,200 can't turn the settings up as far as you want in some hyped up AAA game? 3 00:00:07,200 --> 00:00:11,920 It turns out that in many cases, this isn't because the GPU is underpowered, instead you 4 00:00:11,920 --> 00:00:15,800 might not have enough video memory, or the amount you do have might just be too slow. 5 00:00:15,800 --> 00:00:20,880 But the next generation of VRAM, GDDR7, promises to help alleviate some of these issues with 6 00:00:20,880 --> 00:00:27,800 never before seen bandwidth. This is a big deal because of the rapidly increasing demands being put on gaming PCs, 7 00:00:27,800 --> 00:00:32,200 where the high resolution textures and complex visual effects of big name titles can overload 8 00:00:32,200 --> 00:00:36,320 all but the highest end hardware, meaning you need enough high speed VRAM to hold all 9 00:00:36,320 --> 00:00:43,040 that data that the GPU is being asked to process. In fact, one of the more common gripes from folks who bought the RTX 3070 on an upper 10 00:00:43,040 --> 00:00:48,000 tier GPU was that it only came with 8GB of VRAM, resulting in performance issues and 11 00:00:48,000 --> 00:00:51,600 graphical glitches in games that the GPU itself could probably have handled. 12 00:00:51,600 --> 00:00:56,240 Enter GDDR7, which is expected to bring us a huge bump in memory bandwidth without sucking 13 00:00:56,240 --> 00:01:00,080 down more power, which is great news for those who just want to turn up the settings but 14 00:01:00,080 --> 00:01:03,960 are already fighting the rising power draws of Neurogen graphics cards. 15 00:01:03,960 --> 00:01:08,600 We're talking 36GB per second per PIN. 16 00:01:08,600 --> 00:01:14,520 Once you translate this into a card with a 2.6-bit wide memory bus, that is a full 1.15 17 00:01:14,520 --> 00:01:21,640 terabytes per second of throughput. To put that into perspective, the RTX 4080 only has around 700GB per second of memory 18 00:01:21,640 --> 00:01:25,600 bandwidth, so all that extra space in the pipe could help demanding titles even without 19 00:01:25,600 --> 00:01:31,960 adding more actual VRAM. But what exactly is the secret sauce that allows GDDR7 to reach these new heights? 20 00:01:31,960 --> 00:01:36,800 GDDR7 uses a little trick called PAM3, and that's not the stuff you spray on your 21 00:01:36,800 --> 00:01:41,000 pan so the tater tots don't stick. It's Pulse Amplitude Modulation. 22 00:01:41,000 --> 00:01:46,800 In PAM3, each clock cycle can have 3 voltage states, positive 1, negative 1, or 0. 23 00:01:46,800 --> 00:01:50,400 Each sequence of 2 cycles encodes 3 bits of data. 24 00:01:50,400 --> 00:01:55,440 For example, a negative 1, followed by a positive 1, corresponds to the bits 0, 1, 0. 25 00:01:55,440 --> 00:02:00,800 In this way, each clock cycle encodes 1.5 bits of data instead of just a single bit. 26 00:02:00,800 --> 00:02:04,520 But you're a savvy viewer, and maybe you already know that the current top-end video 27 00:02:04,520 --> 00:02:07,640 memory, GDDR6X, actually supports PAM4. 28 00:02:07,640 --> 00:02:11,960 I mean, isn't that better? 4 is a bigger number than 3. 29 00:02:11,960 --> 00:02:16,520 I think. Maybe. Well, it's only better to a point. 30 00:02:16,520 --> 00:02:21,600 PAM4's encoding scheme actually delivers 2 bits per cycle, making it faster than PAM3 31 00:02:21,600 --> 00:02:28,520 on paper. But there are some drawbacks. More possible states per cycle means that there's a greater chance for errors in data 32 00:02:28,520 --> 00:02:32,880 transmission. And to compensate for this, you need more error correction functionality as well as 33 00:02:32,880 --> 00:02:36,240 a new physical design and different memory controllers. 34 00:02:36,240 --> 00:02:40,440 All of these things increase both the cost and the power consumption, not to mention 35 00:02:40,440 --> 00:02:46,000 that the GDDR6X can only operate in PAM4 mode when data is being sent in 8-byte bursts 36 00:02:46,000 --> 00:02:51,240 instead of the more standard 16-bytes, meaning GDDR6X can actually have less bandwidth overall 37 00:02:51,240 --> 00:02:56,960 in real-world scenarios. Put all this together, and GDDR7 is supposed to give us more bandwidth while being more 38 00:02:56,960 --> 00:03:01,960 energy efficient than GDDR6X, not to mention it can switch off PAM3 when it's not needed 39 00:03:01,960 --> 00:03:05,120 and instead use the 1-bit per cycle NRZ scheme. 40 00:03:05,120 --> 00:03:11,920 But when are we actually going to see our graphics cards? The answer is probably late 2024, as that's when we're due for another major GPU refresh 41 00:03:11,920 --> 00:03:17,720 from both AMD and NVIDIA. And reports are already out that NVIDIA may be planning to use GDDR7 for its upcoming 42 00:03:17,720 --> 00:03:23,720 RTX 5000 GPU series, which I will definitely be buying as long as I can sell a couple of 43 00:03:23,720 --> 00:03:27,880 my organs first. Thanks for watching guys, like, dislike, check out some of our other videos, and comment 44 00:03:27,880 --> 00:03:30,280 with video suggestions down below, and don't forget to subscribe and follow.