Is there a BIGGER RDNA3 GPU coming?

if there's one thing that has accompanied amd's GPU launches in recent years is the hopium that the red team is holding something back sandbagging a Halo product that will finally Dethrone Nvidia time and again AMD has failed to match its fans expectations but perhaps the introduction of chiplets in audio 3 changes things this time around perhaps this time there really is a bigger Navi maybe we're too cool die chip that's coming that will beat nvidia's Halo gpus well unfortunately things are not that simple and today we will take a deep dive into amd's GPU strategy for the next few generations and look at what the limitations of chiplets means for us PC enthusiasts today's video is sponsored by urcdkeys.com if you buy a retail Windows 10 key you'll probably spend a hundred dollars or more but if you buy an OEM key using a service like you are CD Keys it will cost you less than 15 after you use the coupon code c25 the keys work globally and if you want Windows 11 what I recommend you do is just buy Windows 10 and get a free upgrade while you can personally I use Windows 10 Pro once you've made your purchase using one of the many different payment options available you will instantly find your key in your purchased orders on the UR CD Keys website click get keys and copy the key then after you've installed Windows of course Press Start and search for activate and click on activation settings then click change product key pasting the key and click next that's it quick and easy and cheap and your Windows 10 is now activated if you need office 2021 professional you can use the same 25 discount code c25 and get it for just 59.98 since you're over there check out the Black Friday sale for some great discounts on peripherals like lightweight gaming mice and controllers and some very affordable 60 mechanical keyboards thank you to u r CD keys for sponsoring today's video check the links in the video description to get your OEM Windows keys today the move to chiplets in the consumer GPU space is a critical inflection point for the PC industry and AMD should be praised for taking the risk yet again of changing the Paradigm like they did with the Zen micro architecture in CPUs but triplets don't come without complications firstly let's look at what amd's overall vision for the future of computation is and how they are arranging their product lines around it and betting hard on triplets in the process while we all want AMD to deliver a 4090 killer it's important to understand that amd's primary focus going forward is Energy Efficiency not Halo performance at any cost here's why if we look at the world's computational demand we see that we're on a non-sustainable path if we continue with the current path of energy required per bit we will reach a market dynamics limited scenario where the profitability of chips won't be matched by the cost of manufacturing them along with cost of running them and there won't be enough energy produced in the world anyway to keep them running so something needs to be done there are several dot horse technologies that could completely change the landscape like Quantum or photonics and others but the industry can't wait around and with the current technology we have the best solution to the challenges we are facing is specialization meaning domain-specific Hardware or accelerators that are designed to do one task very efficiently looking at the performance per watt curves we can see that CPUs cover a broad range of applications because of their general purpose nature but are not very energy efficient conversely semi-custom socs like Asics and fpgas will cover only a very specific set of applications but be very energy efficient at doing those workloads a good example would be a Bitcoin Asic Miner or a TPU or a video accelerator in the middle are gpus which have been a great choice for many Data Center workloads because they are fairly efficient at highly parallel operations the problem here is that it's close to impossible to have that many design teams working on every kind of accelerator that the market is demanding to cover the growing broad range of applications so amd's vision for the future is that of a modular design so a variety of chiplets some general purpose some domain specific and some configurable so that performance per watts can be improved massively as you know AMD has been heavily invested in chiplets for this reason first with CPUs and now with gpus the problem with hiplets is the overhead which chiplets doesn't need for additional area for interfaces there's replicated logic like we see in the rdna3 memory chiplets higher packaging costs larger overall die sizes compared to the same device in monolithic form and an increasing overall complexity not to mention a higher energy cost in moving data around this is important the segment of the market that is interested in the a large monolithic eye and is willing to pay for it he's very small and AMD realizes this this used to be the case in the server Market only but since the introduction of the RTX 2000 series from Nvidia this Paradox has now transitioned to the consumer Market the majority of consumers and server customers want the cheaper ties not the Halo product Nvidia has identified this and capitalized on it so you have products like the RTX 4090 which is a Halo product that very few people will buy in relative terms and is priced accordingly the only way for this sort of product to exist is in monolithic form to create a gpus powerful as the 4090 inch Hitler's form AMD would need a massive GPU package that would be much larger than the radical limit of 800 millimeters square that we're seeing with Nvidia and that's assuming you could solve the GPU bandwidth challenges more on that in a second there's at least a 10 area overhead when using chiplets compared to monolithic take for instance the first generation of amd's Epic the package consists of four dies each having 213 millimeters squared so that's a total die package of 852 millimeters squared if AMD had created the same epic chip in monolithic form it would have been only 777 millimeters squared so a multi-chip module needs roughly 10 percent more area than a monolithic equivalent that means that the 7900 XTX would be 10 smaller had AMD made it monolithic in the second generation of Epic AMD further desegregated the package moving the i o and memory interfaces to its own separate chiplet the package was then populated with the variable number of ccds or CPU chiplet dies around this new i o die the challenge here was to make this whole package behave like a monolithic die to make this happen AMD had to free up package layers to do signal routing you can see the under CCD routing in this image there's a lot of wiring required to get the ccds talking to each other and to the i o die and all this wiring has to meet power delivery requirements there's a strict limit to how much current you can use in such a device in the transition to the consumer products AMD was able to not just maintain the CCD exactly the same but also reuse the i o die by chopping off a quarter of the server i o die and repurposing it for the zantoo chips it's also important to understand what makes sense to put into a chippers package and what doesn't if we are to understand amd's vision for the rdna gpus in the future for instance why didn't AMD also add dram to the Epic package there's actually little benefit to doing that the flight time through to DDR is actually not that different than on package so unless you are looking to specifically reduce the size of the board which is for instance the Apple approach with the M1 there's not significant latency savings within package dram versus going off package there's a gaining efficiency but the latency doesn't improve that much another thing to consider is that the organic package so that's this green bit that the CPU or GPU are placed on is much more expensive than motherboard or GPU board area that means it's a lot cheaper to keep dram outside of the package furthermore what if oems want to provide systems with varying sizes of dram if AMD added the ramp to the attic package then oems would be limited to whatever size was on there and then there's the margins problem anything that AMD adds to a package that comes from an outside provider will need to be margined up otherwise the market would kill AMD and their shares would drop AMD needs to keep increasing margins that's just the way things are so that's yet another reason why adding things to an MCM package needs to meet a careful balance of performance energy cost and profit margins the exception here is in compute only devices and I covered that in my last video where hbm needs to be on package but that's a very specific segment and we won't be seeing hbm back in consumer devices anytime soon so that's just a simple example but you need to understand that not everything makes sense to be added into an MCM device be it epic wuzanne or rdna3 and each of these will have individual limitations and objectives that will dictate what makes sense to be in package and what doesn't one big difference you'll notice from amd's epic and Zen lines compared to rdna3 is that the CPUs package is on an organic substrate whereas rdna3 is on an interposer so simply doubling the 7900 XTX graphic style is not possible because the interposer would exceed the radical limit for silicon so could AMD put two Graphics Ties on the same package if Andy was planning day-to-day communication for a larger rdna3 device it's unlikely that the 7900 XTX would be built on an interposer like Zan AMD would have used an organic substrate if they were planning such a device it's unlikely that you can have fast enough die to die routing to make that possible without an interposer the modularity that we see in Zen 2 and Zen 3 only makes sense in an organic substrate both technically and economically now the third option would be for AMD to create a whole new larger Graphics called chiplet would that be possible of course by even assuming AMD with throw modularity out of the window and adds say two more core clusters we'd be looking at an interposer based GPU at around 600 millimeters square or even larger and with an extra two classes be enough to beat the RTX 1490 I doubt it what about more Graphics clusters at that point it starts to make no sense to use a chiplets approach it would just be more cost effective to make a monolithic die the same way Nvidia did with the 4090 you can't make into poses that large without running into a ton of issues and don't forget that there's a 10 area overhead in a chiplets device compared to a monolithic one like I explained earlier you might out that the rumored rdna 3 navi-33 die will be monolithic so why wouldn't AMD make a monolithic Navi 31 the point at which makes sense to move to chiplets is the 200 millimeter square die so if Navi 33 is 200 millimeter or less it makes more sense economically for it to be monolithic anything larger and AMD will make it in chiplet form because we're using interposes here there's no modularity like in Zen except for the memory dies which can be reused in addition to that in a recent interview with Gamers Nexus Sam navsuga was questioned about further partitioning the navi 31 die into smaller chippers here's what he had to say the bandwidth requirements are so much higher with GPU because we're Distributing all of this work terabytes of data and textures and vertices so if you try to Route all that information across the CPU interfaces it would be just gigantic this goes back to the routing schematic and issues I showed in Zen 2 currently there's no way for AMD to maximize the chiplet's CPU strategy in the GPU space so we can't have multiple Graphics core dice because we currently don't have a way to do that in an organic substrate and it's unlikely AMD will do a larger die than Navi 31 because it's built on an interposer and the cost to manufacture go exponentially higher as you increase their interposer size not to mention issues with structural stability in others it could be done but its cost would be prohibitive both for AMD to manufacture and then to the end consumer and it doesn't fit into amd's overall strategy of pursuing Energy Efficiency above delivering excessively expensive Halo products as if that wasn't enough why would AMD call the 7900 XTX and XT a 900 card if there was a whole tier above it last gen we saw a refresh of the 6900 XT an AMD called at the 6950 XT and that represents only about a six percent increase thanks to better binning a factory over o'clock and slightly faster Ram the 6950 XT is not a hold tier above the 6900 XT if Andy had a whole tier above the 7900 XTX it would have called it the 7800 XT now is there a 7950 XTX coming of course it will be a refresh of the 7900 XTX not a whole tier of GPU so there are two areas where I think it's realistic to expect improvements over the 7900 XTX in a future refresh one is the move to a 4 nanometer which is an optimized version of five nanometer and the other is the addition of 3D stack V cache like we saw with the 5800x3d the bigger cache would have a small increase in performance and the move to 4 nanometer would reduce power consumption so AMD could perhaps push clocks perhaps even up to three gigahertz although it's possible the front end can't keep up with that so even with higher frequencies there might not be much of an improvement I think it's realistic to expect maybe a five to ten percent performance in or something along those lines versus the 7900 XTX I think it's time for AMD fans to adjust their expectations like I've said in my reaction video to the 7900 XTX announcement while AMD is doing a phenomenal job on a technical level with these gpus their marketing and pricing decisions are destroying the potential of these cars being a success in the market in my opinion AMD has completely thrown out the window the pricing advantage that chiplets could represent for the consumer the agenda now like we saw with xan 4 is to price things as high as possible and see if consumers will pay for them if they don't then they'll just adjust pricing accordingly to me this is highly damaging to the perception that people will have of these products you only have one opportunity to make a good first impression Zen 4 made a terrible first impression and that damage will be difficult to recover from even with price Cuts I fear AMD is making the same mistake with rtna3 the only saving grace there is that Nvidia also failed hard with the 4080 pricing so people are not going as hard on AMD as they should for their pricing strategy honestly I'm also skeptical that we will see a 3D stacked version of the 7900 XTX I think the price of packaging such a GPU would be too high for a relatively small impact to Performance especially as that wouldn't be enough to surpass the 4090 or a potential 4090 TI the one thing that makes me think AMD will launch a 3D stack version of the 7900 XTX is if people do show an interest in 8k gaming so there are a bunch of 8K monitors and TVs being announced at CES and if there's interest there I think a 3D stack version of the 7900 XTX could have a performance Advantage at 8K 8K would probably benefit from the additional cash maybe that's why AMD tried to position themselves in the rdna3 presentation as an 8K gaming platform maybe they are trying to get ahead of Nvidia I think what's more revealing of amd's strategy is that assuming that the 7900 XTX is indeed fast than the RTX 4080 and it should be we're headed to the point where AMD can comfortably deliver in chipler's form gpus that are faster than nvidia's monolithic mid-range offerings the Halo monolithic gigantic GPU will continue to exist for NVIDIA and it will continue to be the fastest I don't think there's a sound economic path to a Halo product other than through a reticle image GPU in monolithic form an equivalent chiplet space device would have too much overhead and would need to be on an interposer for Gigantic proportions something that's currently not economically viable in the gaming space but AMD will have an advantage from now on when it comes to the mid-range it's a shame they're not pricing their mid-range gpus accordingly so with all that in mind it seems extremely unlikely that AMD will produce a larger GPU this generation not only that for NVIDIA will continue to make sense to produce monolithic gpus as Halo products as long as there are enough able to buy them the Halo product greatly contributes to nvidia's mind share as people will assume the lower tier cards are also faster than the AMD equivalents I think AMD has an uphill battle in changing that mindset without a Halo product it seems to me that AMD has given up on the performance crown and will be heavily invested in conquering the secondary tiers making it very difficult for NVIDIA to continue creating monolithic 80 and 70 class cards that can compete in price this also likely means that we will see Nvidia transition to chipler's gpus but only in the mid-range so the 50 90 and 6090 will be monolithic at three nanometer and one nanometer respectively but it's likely the 5080 and 6080 will be chiplets based to be able to compete with AMD in pricing at least that seems the most likely scenario I'm super excited to test the 7900 XTX and XT in the next coming weeks so be sure to subscribe so you don't miss that this video was made possible by my awesome patrons consider supporting me on patreon for just one dollar per month and get exclusive access to the cortex Discord server thanks for watching and until the next one

Channel: Coreteks

Views: 35,469

Keywords: 7950xtx, rdna3, future gpus, bigger navi, big navi, rtx 4090, 4090 killer, 7900xtx, 7970xtx, 7900xt, 7950xt, gaming gpu, pc gaming gpu, graphics card, best graphics card, graphics card 4k, graphics card 8k, gpu for 4k, gpu for 8k

Id: FG6OiGYexfY

Length: 18min 38sec (1118 seconds)

Published: Fri Nov 25 2022