The GeForce RTX 4090 is an absolute monster of a graphics card, however the battle for the next generation of GPUs is just getting began. On Thursday, AMD revealed its personal 4K gaming champion, and the Radeon RX 7900 XTX and 7900 XT goal to topple Nvidia’s goliath with assist from a stash of sensible methods that might make David himself blush.

In addition to improved ray tracing capabilities, the addition of AI cores, and reminiscence galore, these first RDNA 3-architecture GPUs are additionally the primary graphics playing cards that includes a multi-die “chiplet” design, swiping inspiration from AMD’s epic Ryzen success.

Adam Patrick Murray / IDG

Better but? Just like when Ryzen disrupted Intel’s stranglehold on CPUs, transferring to chiplets helps AMD drastically undercut Nvidia’s pricing. While the RTX 4090 prices a chest-clutching $1,599, and the approaching GeForce RTX 4080 prices $1,199, the brand new 24GB Radeon RX 7900 XTX prices $999 and the 20GB RX 7900 will price $899 once they launch on December 13.

Even if AMD’s flagship doesn’t handle to go fairly toe-to-toe with Nvidia’s beast when the benchmark mud lastly settles, it ought to nonetheless be insanely quick—and sport a way more palatable price ticket.

Let’s dig in.

Meet AMD’s RDNA 3 chiplets

Comparing AMD and Nvidia’s methods to David and Goliath appears particularly salient right here. Nvidia slammed the pedal to the steel with its new “Ada Lovelace” structure within the RTX 4090. The 4090 packs a big 608mm² GPU, filled with 76.3 billion transistors. Those envelop every thing you’d usually count on in a GPU die, from media and show engines to the reminiscence controls. It actually is a goliath of a GPU.

AMD

Adam Patrick Murray / IDG

With RDNA 3 and the Radeon RX 7900 XTX, AMD selected a unique path. It nonetheless incorporates a central GPU die, after all—you want one—however it measures a comparably paltry 300mm², constructed utilizing a 5nm TSMC manufacturing node just like Nvidia’s. That’s as a result of AMD cut up its GDDR6 reminiscence interface and second-generation Infinity Cache into separate dies—six in complete—made utilizing the extra mature and cost-efficient 6nm node. It’s a radical change for GPU design, and one which helps AMD maintain prices down. All advised, the Radeon RX 7900 XTX wields 58 billion transistors.

AMD

The Radeon RX 7900 XTX and 7900 XT every come outfitted with 96MB of that Infinity Cache (which is principally on-die L3 cache for sooner reminiscence transfers and gaming speeds), with every of the six dies additionally sporting 64-bit reminiscence controllers. That’s lower than what the RDNA 2-based Radeon RX 6900 XT and 6800-series packed—these had 128MB—however AMD product know-how architect Sam Naffziger defined that this model of Infinity Cache has been configured to do extra with much less, particularly paired to the 24GB of GDDR6 reminiscence over a broad 384-bit bus (or 20GB/320-bit bus within the 7900 XT’s case).

“We’ve tuned this Infinity Cache to enable better data reuse with less capacity to accomplish several things for us,” Naffziger stated. “With the high hit rates, it directly sources most of the memory requests. This delivers data to the engine in less time and uses less power than going out to DRAM.” In sum, he stated, RDNA 3’s Infinity Cache presents 2.7x the height bandwidth offered by its successor, delivering as much as 5.3 terabytes per second for directions that keep inside the GPU and Infinity Cache.

Bottom line: The Radeon RX 7900 XTX and 7900 XT received’t depart you wanting for reminiscence.

Radeon RX 7900 XTX: Inside RDNA 3’s GPU upgrades

But don’t let the introduction of discrete reminiscence dies idiot you: The core RDNA 3 GPU additionally acquired some severe upgrades.

AMD

Architecture tweaks, the migration to TSMC’s superior 5nm node, and splitting out the reminiscence to its personal trailing-node chiplet helped AMD exceed its effectivity targets, with CEO Lisa Su claiming that RDNA 3 presents 54 p.c larger performance-per-watt than RDNA 2. That let AMD stick to a conventional 2x 8-pin energy connector design, neatly side-stepping the melting 12VHPWR adapter controversy Nvidia finds itself embroiled in with the RTX 4090.

With RDNA 3, AMD managed to squeeze in 54 p.c extra transistors than with RDNA 2, in a smaller space, for an total density enchancment of 165 p.c.

Adam Patrick Murray / IDG

That’s all properly and good, however avid gamers are extra excited about what these fancy new RDNA 3 dies do. A deeper dive will little question come earlier than the Radeon RX 7900 XTX’s December 13 launch date, however AMD offered some tantalizing morsels of knowledge at the moment.

RDNA 3 switches to a brand new “unified” compute unit that makes use of dual-issue wave 32 models, so AMD can concern both integer or floating level directions as wanted relying on the workload, which Naffziger says can velocity up gaming body charges and AI duties alike. An enhanced common function register file with 50 p.c extra capability than the RDNA 2 design likewise helps maintain all duties fed.

AMD

Yes, I stated AI duties. Two generations after Nvidia launched tensor AI cores for DLSS and extra in its RTX 20-series GeForce GPUs, AMD is lastly following swimsuit. Each RDNA 3 Compute Unit (CU) packs a pair of AI cores, in addition to an enhanced ray tracing core that helps new devoted directions, 1.5x extra rays in flight, and contemporary ray field sorting and transversal capabilities.

AMD

AMD says RDNA 3 presents as much as 50 p.c extra efficiency per CU than RDNA 2 did, which feels unlikely to catch as much as Nvidia’s blistering RTX 40-series efficiency given the deficit AMD began in, however might make ray-traced video games actually playable on RX 7000-series graphics playing cards. And it’s important to assume these AI cores will probably be put to work upsampling video games earlier than it—particularly since Radeon chief Scott Herkelman teased FSR 3 for a 2023 launch throughout this occasion, with a twofold improve in body charges versus the already-impressive FSR 2.

Speaking of compute models, RDNA 3 might need a a lot smaller die, however it packs in lots of extra CUs than earlier than. The Radeon RX 6900 XT topped out at 80 CUs. The Radeon RX 7900 XTX has 96 working at 2.3GHz, whereas the 7900 XT presents 84 working at 2GHz.

AMD

AMD

From conventional video games to ray tracing to AI duties, AMD’s new flagship ought to provide considerably extra efficiency than earlier than, because the AMD-supplied slides under illustrate. (That stated, as at all times it is best to anticipate impartial benchmarks earlier than judging a brand new graphics card—particularly since these lack any RTX 4090 comparisons that might give us direct perception on how the Radeon RX 7900 XTX compares towards Nvidia’s newest and biggest.)

AMD

AMD

AMD

AMD additionally teased a brand new one-click “Hypr-RX” function that may vastly improve body charges by activating FSR, Radeon Boost, and different useful Radeon applied sciences. Like FSR 3, it’s coming in 2023—particularly the primary half.

Raw efficiency is just a part of the package deal although. AMD additionally upgraded the media and show engines in RDNA 3.

AMD

The Radeon RX 7900 XTX and 7900 XT help DisplayPort 2.1, in contrast to the GeForce RTX 4090. That means it could possibly help 4K shows at as much as a blistering 480Hz, or 8K shows at 165Hz. And 12 shade bits per channel unlocks as much as 68 billion colours, which is welcome, however appears extra like stat porn than one thing anybody aside from content material creators want to fret about.

AMD

Nvidia and Intel maintain essentially the most sway with content material creators, however AMD severely upgraded its media engine with RDNA 3. Not solely does it help AV1 encode and decode, becoming a member of Intel’s Arc and Nvidia’s RTX 40-series, however the GPU can even run simultaneous encode and decode for HEVC or AVC video. Accelerating the media engine’s frequency additionally helps it ship as much as 1.8x the efficiency of RDNA 2, Naffziger stated, lowering export instances. AMD’s new AI cores additionally come to bear right here, with Naffziger promising improved video encoding and convention name high quality because of their assist.

A brewing battle

Adam Patrick Murray / IDG

After a number of disheartening years of shortages and sky-high costs, PC avid gamers lastly have one thing to look ahead to. The GeForce RTX 4090 rocked our socks, however its $1,600 price ticket makes it a luxurious toy for the 0.1 p.c. The Radeon RX 7900 XTX and 7900 XT positively aren’t low-cost, at $999 and $899 respectively, however they’re a complete lot much less than Nvidia’s preliminary choices—and so they seem primed to ship an enthralling 4K+ gaming expertise of their very own on the again of a number of thrilling improvements and beefed-up ray tracing capabilities.

AMD’s in any other case thrilling presentation left us with loads of questions, nevertheless. How will the Radeon RX 7900 XTX and 7900 XT stack up towards the GeForce RTX 4090 and 4080, if not their very own RDNA 2 predecessors? Does switching to a number of dies create any efficiency oddities? Will RDNA 3 lastly ship a compelling ray tracing expertise? We’ll want to attend for impartial benchmarks when these new-breed graphics playing cards hit the streets on December 13 to seek out out the solutions.