AMD Radeon RX 7900 XT And GeForce RTX 4090 GPU Leak Tips Beastly Core Counts And Performance

GeForce RTX and Radeon RX

Should you’re studying this website, we need not let you know that AMD and NVIDIA have new GPUs coming—partially as a result of we have already informed you, a lot of instances. Nevertheless, new data from long-time Twitter leakster Greymon55 paints a giant image of each firms’ forthcoming high-end silicon, and we imply *large*.

Like many of the outstanding leakers, Greymon55 is thought for posting fascinating technical particulars of upcoming {hardware} in tiny, cryptic snippets of textual content. This time round, although, they posted a pile of attention-grabbing specs for NVIDIA’s “AD102” halo-tier Ada Lovelace processor in addition to AMD’s “Navi 31” next-generation Radeon. Each tweets appear to summarize and corroborate earlier leaks and rumors.

inline amdtweet
We’ll begin with the Radeon tweet, because it was posted first. On this newest leak, Greymon55 corroborates earlier rumors that the following flagship Radeon half can be an absolute beast with 15,360 shader ALUs throughout 60 workgroup processors. (Greymon55’s preliminary tweet stated 120 WGP, however they later corrected it) Like present high-end Radeons, it’s going to use a comparatively modest 256-bit reminiscence interface. That can join the GPU bundle to a hefty 32GB of GDDR6 reminiscence at a switch price of 18Gbps per pin.

We are saying “GPU bundle” as a result of as rumors have it, Navi 31 can be a multi-chip module (MCM) design like AMD’s just-revealed Intuition MI200 datacenter compute accelerator. Greymon55 signifies that, like AMD’s Zen 2 and Zen 3 CPUs, Navi 31’s a number of cube can be fabricated on separate processes; on this case, TSMC 6nm for the I/O and cache die, whereas TSMC’s refined 5nm course of is employed on the twin power-thirsty compute cube. That third die will purportedly carry as a lot as 512MB of Infinity Cache. That is quadruple the quantity on AMD’s present flagship, the Radeon RX 6900XT.

If the predictions maintain true and AMD is ready to get these large multi-processor quantity munchers as much as “2.4~2.5 GHz,” that may imply that such a theoretical “RX 7900 XT” based mostly on Navi 31 might chew by means of as a lot as 75 teraflops of single-precision (FP32) floating-point math in a second. It makes one misty-eyed to look again on the Radeon HD 7970 and its “meager” 4 TFLOPs efficiency practically a decade in the past.

inline nvidiatweet
On the NVIDIA facet of issues, Greymon55 likewise collates earlier leaks and rumors right into a handy submit stuffed with jargon that we’ll summarize for you. Not like AMD’s design, the Ada Lovelace-based AD102 will supposedly be a single monolithic die, fabricated on TSMC’s 5nm course of. Stated die will comprise 144 shader modules (roughly analogous to AMD’s WGPs), every with 128 shader ALUs, totaling 18,432 “CUDA cores.”

NVIDIA is purportedly planning a wider 384-bit reminiscence interface connecting 24GB of quicker GDDR6X reminiscence at 21 Gbps per pin. If true, that may put the AD102 half at practically 1TB/sec of reminiscence bandwidth—nearly double that of Navi 31. Having a single block of reminiscence linked over a super-fast bus makes NVIDIA’s chip equally serviceable for graphics or bandwidth-hungry compute workloads. Against this, AMD’s RDNA 3 design is meant for graphics primarily, and that is the aim of the massive Infinity Cache, which in concept (and in apply, on extant RDNA 2 elements) serves to cut back the bandwidth necessities of graphics workloads significantly.

Greymon55 theorizes a clock price of “2.3~2.5 GHz” for AD102 as effectively, which might put the chip someplace within the 85 to 92 TFLOPs vary for single-precision compute. Given NVIDIA’s said standing as a “one-architecture firm,” its chips must be the proverbial “lady who can do each.” Fortuitously, Ada Lovelace appears greater than as much as the duty.

inline powertweet
In fact, even with bleeding-edge fabrication processes, humongous GPU efficiency comes at an unlimited energy price. As Greymon55 requested final week, “are you able to settle for it?” The leaker themselves feedback that whereas they don’t have any precise information on Navi 31’s energy utilization, they theorize that the entire board energy should be not less than 350 watts, and certain extra, on as much as as excessive as 550W. Additionally they state that AD102 “should be” not less than 450W, and probably as excessive as 650W.

These sorts of energy numbers are going to make these playing cards a ache to chill, and you are not going to be working them off your old-school PCI Specific energy connectors—it might take 4 eight-pin cables to run such a card! As an alternative, it appears possible that each AMD and NVIDIA can be utilizing the recently-confirmed PCI Specific 5.0 auxiliary energy connector. This connector is analogous (if not similar) to the one which NVIDIA at the moment makes use of on its RTX 3000 collection Ampere GPUs, and is rated to produce as much as 600 watts over a single bundled cable.

Each merchandise are anticipated for late subsequent yr. NVIDIA’s half anticipated in October 2022, and whereas there is not any leaked timeline for AMD’s elements, Greymon55 has implied that the purple workforce is concentrating on an identical date. Probably, earlier than we see any of those elements hit the market, we’ll see that long-rumored RTX 3000 “Tremendous” refresh early subsequent yr, in addition to Intel’s entry into the market with its first discrete GPUs for the reason that ill-fated i740. Assuming the GPU scarcity lets up—and Dr. Lisa Su thinks it’s going to—2022 seems to be an especially thrilling time to be a {hardware} nerd.

Be the first to comment

Leave a Reply

Your email address will not be published.