Leak: NVIDIA RTX 4090 to Feature 60% Higher Clock Speeds Than RTX 3090, RTX 4080 to Come with 21Gbps Memory & More

As we inch closer and closer to the avalanche of next generation products from AMD, NVIDIA, and Intel, the frequency of leaks keeps getting higher with each passing day. Out of all the hardware set to hit the market later this year, perhaps none is as hyped as NVIDIA’s RTX 40-series.

The successor to the current-gen Turing GPUs, RTX 40-series is expected to bring forth a noticeable boost in performance across the board. So far, we’ve seen numerous reports, leaks, and rumors pertaining to the insane power limits of these cards, along with most of the specs of the lineup.

Today, we add to the long list of rumors for Ada Lovelace and this latest one comes from a familiar source. Notorious leaker, @koptie7kimi, who has regularly provided leaks on the initial launch lineup of RTX 40-series has posted yet another update on Twitter, this time going over the clock speeds of the flagship SKU, along with some other tidbits.

Around a week ago, Kopite provided a major update for RTX 4090, RTX 4080, and RTX 4070. Today’s tweet is a sort of follow-up to that with most of the information carrying on from the previous leak(s) with some added details on top. Most importantly, we get our first taste of what to expect from the top-end in terms of speeds.

RTX 4090 latest updates

Starting with the RTX 4090, the leaker mentions NVIDIA’s next BFGPU is said to feature a base clock speed of 2235MHz with a boost clock of 2520MHz. That is impressive in of itself, but the card goes further and boosts up to 2.8GHz in-game, which represents the GPU’s actual maximum clock speeds. With some aftermarket variants, we could see this number go up to 3GHz.

These kind of figures are the highest we’ve ever seen from NVIDIA. In fact, the base clock speed is around 60% higher than the RTX 3090, while the boost clock is around 50% higher over the RTX 3090. That comparison really puts the improvements in perspective and shows how AMD, who’s generally been the winner when it comes to clock speeds, needs to be worried this time.

The RTX 3090 | NVIDIA

NVIDIA is using TSMC‘s N4 (5nm) process node, which is more advanced than the N5 (5nm) wafers AMD is using for RDNA3. That alone puts the Red Team in a disadvantageous position, but efficiency is still expected to be AMD’s bread and butter, especially with Ada Lovelace’s outrageous TGPs. The RTX 4090 is expected to have a 450W TGP, a number that will be pushed even further with custom models.

Moreover, Kopite reiterates the specs of the 4090 from his prior leaks. As we’ve known before, the RTX 4090 is using a AD102-300 GPU, but this time the leaker has mentioned it as “AD102-300-A1“. This GPU features 16,384 CUDA Cores, aka 128 Stream Processors (SMs) . The full AD102 GPU has 148 SMs which means the 4090 is using a cut-down version, and NVIDIA is saving the full die for a possible 4090 Ti/TITAN card down the line.

Regardless, the 16,384 CUDA Core spec combined with the maximum boost clock of the card (2750MHz), the RTX 4090 has roughly 90 TFLOPs of compute power in single-precision workloads. That’s just slightly lower than the 100 TFLOPs figure we heard about a few weeks ago for the TITAN-class Ada Lovelace card, and almost double of what the RTX 3090 Ti is capable of. Rest assured, the RTX 4090 will be one powerful beast.

That sentiment is further reinstated by the memory specs of the card as it sports 24GB of GDDR6X memory running across a 384-bit memory bus, clocked at 21Gbps. All that comes around to give to over 1TB/s of memory bandwidth. That’s the same as the current BFGPU, the RTX 3090 Ti. Lastly, it’s important to note that the RTX 4090 will use a single 16-pin (PCIe Gen5) power connector, which is capable of delivering up to 600W of power.

The new PCIe Gen5 16-pin power connector | TechSpot

RTX 4080 latest updates

Moving on to the RTX 4080 and RTX 4070 SKUs, both graphics cards are said to feature different GPUs. The RTX 4080 is using a cut-down version of the AD103-300 GPU with 80 Streaming Processors enabled out of the total 84, which makes for 10,240 CUDA Cores. Again, a 4080 Ti model in the future will likely use the full AD103 die. 

Furthermore, Kopite tells us to expect 16GB of GDDR6X memory in the RTX 3080 spread across a 256-bit wide bus interface, also running at 21Gbps like the 4090. However, due to the reduced VRAM capacity and a slower memory bus, the bandwidth is cut down to just 672GB/s. Previously, the memory speed was rumored to be 18Gbps according to the same leaker.

As for the power, we’re looking at a 420W TBP, which is now confirmed from its previously-unsure status in Kopite’s last leak roundup, plus it’s just barely lower than the reference power requirement for the RTX 4090. There is no word on clock speeds for thr 4080 just yet. Lastly, we see the same “A1” being added to the end of the GPU name for the RTX 4080, which means it’s using the “AD103-300-A1” GPU.

Close-up of an NVIDIA GPU | Videocardz

RTX 4070 latest updates

Finally, coming to the RTX 4070, this card is reportedly using a cut-down version of the AD104 GPU, called “AD104-275-Kx-A1” (the ‘x’ is a placeholder for a number). We’ll see the AD104 GPU fully utilized in the RTX 4070 Ti releasing next year. As for this particular GPU, it is rocking 56 Stream Processors, or 7,168 CUDA Cores, around 1000 more than the RTX 3070 Ti, which has 6,144 cores, so this is a nice generational upgrade.

The RTX 4070 comes with 10GB of GDDR6 memory, not GDRR6X—NVIDIA’s special flavor of G6 developed in tandem with Micron exclusively for RTX cards— which is a downgrade from the current-gen RTX 3070 that does feature GDDR6X memory.

That 10GB VRAM is running across an unusual 160-bit memory bus with the speed clocked at 18Gbps, giving us 360GB/s of memory bandwidth, that’s lower than the RTX 3070’s 448GB/s bandwidth. That’s because the RTX 3070 has a 256-bit wide bus interface, despite having 2GB less memory. Moreover, the RTX 4070 is going to feature a 300W TBP (total board power) so expect higher power requirements in third-party variants.

Kopite goes out of their way to remind us that the MSRP of the xx70 class won’t be lower than the RTX 3070 or 3070 Ti, so a price bump may be in order. Let’s just hope this doesn’t give scalpers any opportunity to exploit buyers like last time, and that enough stock is available for everyone.

Bots have been notoriously used to buy GPU stock as soon as it’s available to then sell at scalped prices later | Jim Cooke/Gizmodo

That covers all our bases for the RTX 40-series, or at least the launch lineup of the family. NVIDIA is expected to release one card per month, starting from the lowest-end option, the RTX 4070 in either September or October. From then, expect the RTX 4080 next month, and the 4090 the month after. The RTX 4060, on the other hand, is said to launch around CES 2023 in January.

AMD is expected to launch its RDNA3 GPUs in November, meaning that the Red and Green Team will go head-to-head to claim the crown. Whether NVIDIA is able to defend its throne or AMD reigns supreme in an upset the company has likely been envisioning for a long time, remains to be seen. 

Huzaifa Haroon
Born and raised around computers, Huzaifa is an avid gamer and a Windows enthusiast. When he's not solving the mysteries of technology, you can find him writing about operating systems, striving to inform the curious.