As we inch closer to the next round of major GPU releases in the form of a proper new generation from NVIDIA and AMD, along with Intel who will be launching its first-ever discrete GPUs very soon, leaks and rumors are starting to pile up. Today, we got perhaps the year’s biggest leak dump so far with two popular (and credible) leakers choosing to share information about AMD, NVIDIA, and Intel’s upcoming GPUs all on the same day.
Kopite7kimi and the infamous Greymon55 both took to Twitter to provide updates on the upcoming NVIDIA Ada Lovelace and AMD RDNA 3 GPUs, both expected to hit the shelves later this year. Moreover, even though Intel’s first-ever discrete GPU generation isn’t here yet, we already have rumors about next-gen Intel Arc B-Series (Battlemage) GPUs. We’ll touch on all the rumors, starting first with NVIDIA.
NVIDIA Ada Lovelace “AD102” GPU
Firstly, I should get this rumor out of the way: some sources are now reporting that NVIDIA’s next-gen GeForce RTX series won’t actually be called RTX 40xx, instead the company might be skipping a generation and call Ada Lovelace RTX 50 series. Whatever it may end up being called officially, what’s more important are the underlying specs. That’s why in the tweet attached below, the leaker is referring to Ada Lovelace by the GPU class (90, 80, 70 etc.) and not calling it RTX 40.
With that out of the way, leaker Kopite7kimi shared the specs of three Ada Lovelace GPUs—presumably the ones that would launch first—and all three of them pack in quite the requirements. The lowest-end RTX 4070 class GPU in that list has a TDP of 400W, just for context the RTX 3090 has a 350W TDP. We’re also looking at 24Gbps GDDR6X memory across the board with varying memory capacities depending on the GPU class, as you can see in the tweet below.
**90 PG137/139-SKU330, AD102-300, 24G 21Gbps, 600W
— kopite7kimi (@kopite7kimi) April 11, 2022
However, these specs weren’t shared today, in fact they were shared on April 1st and despite the obvious underlying April Fool’s doubt, the specs are very much supposed to be real. What was shared today was an update to Kopite’s original rumored specs in which he corrects the original tweet by sharing updated specs of the RTX 4090 class GPU. Now, the GPU has been given the SKU number “330” whereas it was previously SKU “310“.
Essentially, this implies that the RTX 4080 class GPU might be the one used to power the RTX 4090 as well, which means that RTX 4090 is likely using the full die whereas the RTX 3080 is using a cut-down version of it. Moreover, that leaves the RTX 4070 class GPU as the second GPU in the lineup, named “AD104“, the RTX 4080/4090 GPU, on the other hand, is called “AD102“.
Another notable change in Kopite’s new revised specs is the reduced memory speed. In the original tweet, we saw the RTX 4090 class GPU featuring a 24Gbps speed but in the updated specs, it now says 21Gbps which now means that the entire Ada Lovelace launch lineup does not share the same memory speed. Last but not least, Kopite also “PG137” before the SKU number of the 4090 class GPU but with an oblique indicating that he isn’t sure whether the SKU is PG139 or PG137.
By now, you’ve probably come to heard that Ada Lovelace will be on the extreme end of the spectrum and that certainly looks to be true. The top-end AD102 GPU is set to bring massive improvements over the current generation with an expected 600W TDP (facilitated by 16-pin PCIe Gen5 power connectors), 24GB of memory across a 384-bit bus, a nearly 40% increase in theoretical TFLOPs and 71% increase in Streaming Processors over the RTX 3090 Ti.
AMD RDNA 3 “Navi 31” GPU
While all that enticing information about NVIDIA’s next-gen BFGPU might lead you to believe that the Green Team has this one in the bag, AMD (and Intel) are going to be more competitive than ever this time around. While reports already exist in the wild claiming that Navi 33, one of the RDNA 3 GPUs, is even faster than the Radeon RX 6900XT, today’s rumor from infamous leaker Greymon55 is specifically about the Navi 31 flagship GPU.
two 5nm GCD
four 6nm MCD
one interconnected controller
7 Chiplets in total
— Greymon55 (@greymon55) April 11, 2022
Greymon55 took to Twitter to reveal the possible specs of Navi 31 that will compete directly against NVIDIA’s aforementioned AD102 GPU. According to the leaker, Navi 31 will feature 7 chiplets in total resulting in unprecedented performance gains over the current generation. As you know, GPU designs right now are mostly monolithic which means there is only one chip on the die.
MCM or Multi-Chip-Module designs in GPUs that combine multiple chips and fuse them on a single die have been done in the past, most notably by AMD too, but none of them have really yielded successful results. This time around, though, AMD is looking to take MCM seriously with a big shift in design philosophy. AMD has already seen massive success with chiplets in their CPUs so it only makes sense that this technology would eventually make its way over to their GPUs as well.
AMD’s move to a multi-chip design with (certain) RDNA 3 GPUs will allows them to pack in a lot more power in a smaller space while consuming less power, something that Moore’s Law Is Dead attests to. Moving back to the tweet, Greymon mentions that there will be 2 GCDs, 4 MCDs, and 1 interconnected controller, making 7 chiplets in total on one die, though we are not sure if the MCDs are actually stacked on top of the GCDs or located elsewhere on the die.
The GCD or Graphics Complex Die, which this GPU has two of, is the actual GPU, whereas the MCD or Memory Controller Die (rumored) will reportedly house the Infinity Cache and/or the memory controllers. The GCDs are manufactured using TSMC‘s 5nm process node while the MCDs are using TSMC‘s 6nm process nodes, according to Greymonn. The leaker, however, does not provide any word on the interconnected controller’s fabrication.
This will make AMD the first manufacturer to incorporate a MCM design in a GPU, a USP that could lend them a massive competitive edge. Reports suggest that the top-end Navi 31 GPU is more than 2x faster than the current Radeon RX 6900XT. If AMD does succeed here, they will have a real shot at winning this graphics generation and cement their name in the industry as an innovator forever.
Intel Arc B-Series (Battlemage)
Last but definitely not the least, we have Intel and its forthcoming next-generation of discrete GPUs. I should preface all of what I’m about to say with the fact that this information wasn’t actually revealed today. In fact, the tweet containing rumors about Intel’s second generation GPU outing is about 2 weeks old but was only discovered now by Videocardz. Regardless, let’s begin.
Some time ago, Intel accidentally published a test driver online for the public to see. This test driver was meant for internal testing of Intel’s next generation Arc GPUs, codenamed “Battlemage“. Thanks to that leak, we learned that the company was already working on the next generation of Arc Graphics and that it would still be based on the same Xe-HPG (High Performance Gaming) architecture that Intel will debut with Arc A-Series this year.
That driver leak contained the name of a bunch of GPU configurations, one of which mentioned “ELG_X4“, implying that this GPU would be using a 4-tile config of the next-gen Intel Battlemage GPU. While right now, we don’t know the specs of Intel Arc Battlemage and even leaks are scarce since the generation is not due until 2023/2024, the following tweet from Redfire75369 sheds some light on the potential spec-sheet by analyzing that driver leak furthermore.
In each configuration below, it would have the corresponding amount of FP32 ALUs.
1T: 5120 (+25% ACM SoC1/G10)
2T: 10240 (x2.5 ACM SoC1/G10)
4T: 20480 (x5 ACM SoC1/G10)
The 2 tile configuration is the most likely as it currently stands.
— Redfire (@Redfire75369) March 30, 2022
Redfire’s analysis suggests that the 4-tile config of Arc Battlemage is likely the top-end SKU and that there will be a two-tile and a one-tile config as well. Focusing on that 4-tile GPU, we could see a total of 40 Xe-Cores and 320 Xe-Vector Engines per each tile, giving us 1280 Vector Engines in total. That means the 4-tile config could feature up to 20480 FP32 cores, which would result in a theoretical 5x performance gain over the Arc A-Series ACM-G10 current-gen flagship GPU from Intel.
Of course, we can’t confirm that that “ELG_X4” really is referring to the tile configuration and all of this is purely based off of speculation. In fact, we can’t even say if a tile configuration like that is planned for Intel’s next generation of Arc Graphics. Intel seems to be going the tile-route with its 14th Gen Meteor Lake CPUs at least, so maybe that is an unofficial indicator of the same design philosophy making its way down to the company’s GPUs.
However, Intel has officially confirmed that Battlemage will target the “Higher Enthusiast” market segment and the subsequent “Celestial” generation would target up to “Ultra Enthusiast” market, so that’s saying something. Since Battlemage will likely release in late 2023 or 2024, there is a chance it could hit the market when the next-gen RTX 4090 and AMD Radeon RX 7900XT GPUs are still on the shelves, making for quite the interesting comparison.
With that, we’ve successfully covered all of the current GPU rumors floating around town. Just kidding, that was only one day’s worth of reports and leaks, and this town never sleeps so expect even more news about all of the things mentioned above in the future. After a dreary last two years, hope is starting to pile up and shine a bright light upon the consumers looking for a new GPU.
AMD and NVIDIA are going to be at their most competitive this season and now with the addition of Intel taking discrete GPUs very seriously, competition is going to be tough and unforgiving. At the end, we can only expect this GPU war to benefit the consumers with more and more GPUs than ever flooding the market as billion-dollar semiconductor giants fight to see who can win our pockets.