A contemporary new stack of leaks has simply dropped in regards to the subsequent era of AMD Radeon GPUs, they usually level to a particularly highly effective graphics card on the prime finish that is no less than 50% extra highly effective than the Radeon RX 9070 XT and will tackle the longer term RTX 6080 from Nvidia. For the time being, AMD has no product to compete with Nvidia on the subject of high-end GPUs, however it seems to be as if the corporate is planning an all-out assault with its subsequent era of RDNA 5 graphics playing cards.
This is not the primary time we have seen rumors a couple of new high-end AMD GPU, with a earlier leak pointing to the Radeon RX 10090 XT (or no matter it finally ends up being referred to as) doubtlessly with the ability to tackle the RTX 5090. This newest rumor tones down the adrenaline a bit, however there is a good likelihood this new GPU may unseat the RTX 5080 in our information to purchasing the best graphics card if this rumor is true.
This newest leak comes from Kepler_L2 on the Anandtech forums, who has type for AMD GPU leaks, having correctly leaked the variety of compute models, the quantity of reminiscence, and the quantity of Infinity Cache of the top-end RDNA 4 GPU, which we now know because the 9070 XT, again in December 2024. This time, Kepler_L2 has shared some block diagrams that present 4 new AMD GPUs, with the top-end mannequin having some severely highly effective specs.
I’ve shared this diagram beneath, and it reveals a GPU that is full of 96 compute models (CUs), that are the fundamental constructing blocks of AMD GPUs, containing the stream processors, RT cores for ray tracing, and AI cores. As a degree of comparability, the Radeon RX 9070 XT solely has 64 compute models, and nonetheless provides the RTX 5070 Ti a run for its cash, so this GPU seems to be set to be no less than 50% extra highly effective, and possibly extra, as a result of there’s rumored to be a ten% improve in rendering efficiency between like-for-like RDNA 4 and RDNA 5 GPUs.
Every set of six compute models has a render backend (RB), whereas the graphics command processor (GCP), {hardware} scheduler (HWS), and L2 cache sit within the center. The whole thing can also be surrounded by a complete of 16 unified reminiscence controllers (UMCs), and if these are every 32-bit, then we may very well be taking a look at a super-wide 512-bit reminiscence interface in complete, as used on the RTX 5090 in the present day. That simply provides AMD the scope to equip this graphics card with 32GB of reminiscence (and even 48GB utilizing 3GB chips), with very excessive bandwidth, relying on the pace and sort of reminiscence used.
Based on the sooner leak linked above, AMD is splitting its RDNA 5 lineup into a number of tiers, all of which begin with AT, which stands for Alpha Trion. This chip seems to be prefer it may very well be the AT0 mannequin that sits on the prime, whereas the opposite block diagrams leaked by Kepler_L2 may seek advice from the cheaper AT2, AT3, and AT4 fashions.
Assuming that is the case, AT2 would correspond with the subsequent block diagram down, which as an alternative organizes compute models into blocks of 5 per render backend, somewhat than six. This diagram reveals a complete of 40 compute models, together with six reminiscence controllers, implying a 192-bit bus. This spec would put it in between the Radeon RX 9060 XT (32 CUs) and Radeon RX 9070 (56 CUs), suggesting that AMD is trying to present a GPU between the standard price range and mid-range choices with this era.
Under it is a diagram of a GPU that solely has 24 compute models, which may very well be AT3, a GPU that is additionally talked about within the newest video from Moore’s Legislation is Useless, which you’ll see beneath. MLID states that this GPU has 48 compute models, which is precisely double the quantity, suggesting some confusion in regards to the quantity, maybe as a result of they’re dual-issue models – each RDNA 3 and RDNA 4 compute models have the power to execute two directions per clock cycle, as they successfully include two engines.
This diagram reveals eight reminiscence controllers, although, which is greater than the one earlier than it, doubtlessly giving it a 256-bit reminiscence bus, that means it may very well be paired with slower reminiscence and nonetheless obtain respectable bandwidth. MLID has speculated that this GPU may even be paired with low-cost LPDDR5X reminiscence to carry down the associated fee, with its bandwidth boosted by the vast bus.
Lastly, on the finish is what seems to be like a quite simple, super-cheap GPU with simply 12 compute models and 4 reminiscence controllers, which might nonetheless give it a 128-bit bus. Once more, although, that is half the variety of compute models instructed by MLID.
In fact, as with all tales primarily based on rumors and leaks, take this all with a grain of salt for now, however I would like to see AMD come again combating with a real high-end GPU, as sturdy competitors in the end brings down costs. With 96 RDNA 5 compute models, AMD most likely would not be capable of tackle the RTX 6090, however it could doubtlessly have shot at dethroning the RTX 6080, relying on what Nvidia brings us with its personal next-gen GPUs.
Within the meantime, one space the place AMD is main the pack is CPUs, and in case you’re considering of upgrading, you may take a look at our information to purchasing the best gaming CPU, in addition to the very best CPU cooler to maintain its temperatures in test.
Would you wish to see AMD make a real comeback within the high-end GPU market? Talk about the newest GPU tech with us on our group Discord channel.
