Next-Gen CPUs/GPUs have a HUGE problem!

there's a huge problem with a major impact on basically all next-gen CPUs gpus and every other type of computer chip and it has the potential to completely alter the development and design of semiconductors as a whole it's affecting every company from AMD and Nvidia to Apple and even Intel No One Is Safe we just got confirmation that as time cells stop scaling if you're now wondering what the heck astrum is and why it's so important I've got you covered in this video we will take a look at what SRAM does and why it's approaching death has massive implications for the entire chip industry and the future of PC Hardware SRAM short for static random access memory is a type of memory cell not unlike dram which is used in your computer's main system memory and what most people would simply call Ram but ashram is much faster than dram and the preferred memory type for any kind of low level cache on computer chips any modern CPU or GPU uses a cache system where the most important data is stored on the chip itself right next to the CPU or GPU cores so they have blazing fast access to the information needed for upcoming calculations anything that doesn't need to be accessed that quickly is stored on the much larger much cheaper but also much slower system memory and because SRAM cache is located directly on the chip itself it's also manufactured in the same process node as the chip Apple's latest a610 chip for example is produced in tsmcs 4 nanometer node and so is the cache on the a16 nvidia's new RTX 40 gpus are also produced using tcmc's four nanometer node and so has the huge amount of cash on those gpus amd's new ryzen 7000 CPUs are again produced by tsmc on a 5 nanometer process and they also include a large amount of 5 nanometer SRAM cells and while Intel still uses their own manufacturing process they also include a lot of sram-based cash on their CPUs but why is the scaling of SRAM so important ever since the invention of silicon-based semiconductor chips production process nodes have gotten smaller and smaller that way modern chip can use more and more transistors while keeping their small size which allows for faster chips and is important to keep costs down but over the past years it has become apparent that not all parts of the computer chip continue getting smaller at the same rate while the logic of the CPU and GPU still scales very well even to this day as stem cells have long started to lag behind the scaling of the rest of the chip at first logic density increased by a factor of 2x and asram scaled only by 1.9 x then the disparity started to grow and with recent process notes it wasn't uncommon to see only a 1.2 to 1.35 x as I'm scaling while logic downstairs still improved at 1.7 x or higher this has led to a point where aspirin-based cash is consuming a disproportionately large area of any modern computer chip just because it has lag behind in scaling for so long and now for the first time ever it looks like tsmc's new three nanometer node will offer zero density Improvement compared to current 5 nanometer nodes that means the size of an Aston cell won't shrink when using tcmc's upcoming and 3E process node this news comes at basically the worst time ever because just recently all modern high performance chips have started to implement increasingly larger astron-based cache systems into the design amd's version CPUs in 2017 had a maximum of 4 megabytes L2 and 60 megabytes L3 cache a new ryzen 7000 CPU has up to 60 megabytes L2 and 64 megabytes L3 cache Intel's most recent Raptor leg CPUs are another great example the main difference between the new Raptor leg design and last year's older leg is the amount of cash Raptor Lake more than double Steel 2 cache and increases dl3 cash by another 20 it gets even crazier when we take a look at nvidia's rtx40 Loveless gpus while a RTX 390 TI has only 6 megabytes of L2 cache the RTX 4090 has a whopping 72 megabytes that's a 12s increase in L2 cache this trend is visible in basically every CPU or GPU design Apple's A and M socs are also using large amounts of cash and with the exception of intloom all these companies rely on tsmc to manufacture their chips nvidia's next-gen Blackwell gpus will most likely use tcmc's m3e process node with literally zero astrum scaling and the need for even more cash the chips will become much larger and thus a lot more expensive to produce same with send 5 on a possible 3 nanometer node Apple's a17 and M3 socs which are rumored to use a N3 variant also face the same problem with little to no astrum scaling Apple either has to stop increasing the amount of cash or swallow the bitter pill and produce even larger silicon dies and if you think Intel is in the clear because they are not using tsmc as a manufacturer I've got bad news for you first Intel's own process nodes are suffering the same fate regarding the scaling of Asian cells it's not a tsmc problem it's a physics problem SM cells are very delicate and their building blocks are often designed by hand they just can't get any smaller without risking their functionality and second not only are intels or tpus produced by tsmc but their upcoming meteor Lake CPUs will also use parts manufacturer by tsmc it really seems like a dead end modern shapes need more and more cash but not only has ashram scaling slowed down over the past years it seems to have come to a complete stop but isn't there anything that can be done about it great question so glad you asked there's a reason I didn't mention amb's latest Arden A3 gpus and the ryzen 7 5800 x3d or in these upcoming send 4 variants that use 3D recash because chiplets can be one of the solutions to combat the stagnation of ASM scaling to be honest it doesn't really solve the problem it's more of a smart workaround to counter the number one problem Rising manufacturing costs all the process nodes are not only more mature and thus easier to manufacture they are also much cheaper and if new bleeding edge nodes don't offer any benefits regarding the size of ashram cells it would be super clever to Just Produce all the cash in an older node wouldn't it if only there wasn't a problem that there are parts of a CPU or GPU that you want to manufacturer in the best nodes available because it's full of logic transistors that still scale well and benefit from higher clock speeds and lower power consumption the solution is to split off the cache and other parts of the chip that don't scale well with new nodes produce them in an older and cheaper process node and then add the parts together with fancy packaging that's exactly what AMD is doing with the new RX 7900 XCX and XT gpus these graphics cards consist of one large chip that contains all the GPU cores and other logic units and up to six tiny ships that only hold cash and a memory interface which is an analog design and scales equally back as SRAM the smart part is that the important main chiplet the so-called Graphics compute time is manufactured in a bleeding edge 5 nanometer process while the small memory cache dies are produced in an older and cheaper six nanometer node and since cache and the analog interface don't scale well but are cheaper to produce on the older node it's a cost Advantage it's a similar concept type with amd's 3D cache first seen on send 3D and the ryzen 7 5800x3d here AMD is using the same 790 process for both the CPU triplet and the cache chiplet but that was just a proof of concept you can bet that once AMD CPUs are on more modern node with zero SRAM scaling like tcmcs and 3E the cache chiplet for the 3D V cache will be produced in an older process node and AMD isn't the only company working on chiplets Intel's next-gen meteor Lake will split up the CPU into different chiplets it might take until another generation could go there but it's already being developed and with the knowledge we just gain I'm sure you can see why it's not only likely but almost certain that Nvidia and apple are also working on implementing triplet design into their future products with the impending death of SM scaling it's not really a choice but the only feasible way to produce competitive chips in the future there are other possible solutions being worked on most different types of memory cells that might still allow further scaling but there isn't a clear successor to assume inside it's do or die and the future clearly belongs to chiplets now you know why let me know in the comments how fast you think we will see chiplet designs from Nvidia and apple and who will come out on top when confronted with the new reality of semiconductor manufacturing the cards are being mixed again and I'd like to get to know your opinion you know what to do if you found this video interesting and see you in the next one

Channel: High Yield

Views: 198,235

Keywords: SRAM, sram, sram cache, l2 cache, l3 cache, TSMC N3, N3E, 3nm tsmc, node shrink, amd, intel, nvidia, apple, chiplets, chiplet design, chiplet cpu, chiplet gpu, future semiconductors, death of sram

Id: vQ5JPqeFitM

Length: 8min 58sec (538 seconds)

Published: Thu Dec 22 2022