Rumor: Nvidia’s Next-Gen GPUs to Top Out at 600W, Possibly with GDDR7 Memory

The latest leaks on the video card front contain a mix of good news, and the usual bad news. The good news is that it doesn’t look like Nvidia’s upcoming Lovelace GPUs will be the 850W nuclear reactors that were previously reported. Instead, the GPU will likely max out at around 600W or so per the latest ATX 3.0 PCIe cable news (more on that below). The bad news is that even at 600W, that’s almost double the amount of power required by a card like the OG RTX 3080, which is a 320W GPU. Nvidia will continue to ship cards for these lower TDP segments, but the company may be preparing to expand the upper range of its hardware by quite a bit. The new Nvidia GPU may also feature GDDR7.

The latest leaks come from Moore’s Law is Dead via Wccftech and we recommend consuming it with a GPU-sized grain of salt. The biggest news is that although Nvidia is going to be pushing the limits of power consumption with its upcoming GPUs, it has apparently decided to reign things in a bit after testing revealed 500W to 600W is the maximum it could do with air-cooling. Although AMD in the past has gone with liquid cooling for some of its flagship GPUs, such as the RX 6900 XT LC, Nvidia has never taken that route, leaving it to its Add-In Board (AIB) partners to offer “hybrid” designs with liquid-and-air cooling or just pure liquid via a pre-installed water block on the GPU.

This power limit is inline with the recent unearthing of the 12VHPWR (12-Volt High Power) PCIe Gen 5 connector for next-gen graphics cards, which comes straight out of Intel’s ATX 3.0 Design Guide and tops out at 600W. The new GPU cable features a 16-pin design with 12 pins for power, and four pins for communication between the PSU and GPU to determine how much power is needed. It’s expected that next-gen GPU will be PCIe 5.0, which will require a compatible power supply, unless someone comes up with an adapter. The RTX 3090 is currently a 350W GPU that takes up three PCIe slots, so just imagine how thick a 500W version might be.

Another bit of info is the possibility of memory speeds landing in GDDR7 territory, which maxes out at 32Gb/s. Nvidia currently uses GDDR6X for its high-end GPUs at 19Gb/s, and GDDR6 for its midrange cards at ~14Gb/s. AMD has shipped GDDR6 at 16Gbps and GDDR6X can theoretically hit 21Gbps, but the RTX 3090 Ti expected to feature that VRAM has yet to appear. We’ve been on GDDR6 for several generations, but haven’t heard much about GDDR7 yet beyond a brief teaser from Samsung late last year.

If Nvidia does offer GDDR7 it will offer a tremendous boost to memory bandwidth even on a 256-bit wide memory bus. The higher-clocked memory could go hand-in-hand with the recently reported boost to L2 cache Nvidia is planning that we reported last week. That rumor alleged Nvidia will increase the amount of L2 cache on the AD102 flagship die by up to 16x. If this all comes together we could be looking at almost a doubling of memory bandwidth on the flagship models.

Though Moore’s Law is Dead lists what he thinks will be the final lineup for the Ada Lovelace line of GPU, and he speculates on how each card will perform, one thing seems clear from the leaks. Nvidia is apparently intent on making the flagship die, the AD102, an absolute beast, while only allowing modest gains on the rest of the product stack below it. For example, he estimates based on leaks specs that RTX 4090 will offer an 80-110 percent boost to rasterization performance compared to the RTX 3090, with double the ray tracing (RT) performance, which would indeed be a massive leap forward. However, the rest of the cards below will do that leap frog thing where a lower-end card is equivalent of a higher-end card from the previous gen, so something like the RTX 4060 will be equivalent of an RTX 3090, both in rasterization and RT performance.

If you’re already thinking, “there’s no way I will be able to afford that,” you are probably right. However, Moore’s Law is Dead reports that Ada Lovelace pricing will be roughly similar to Ampere’s, but with the current market it’s hard to really know what that even means anymore since nobody has been able to buy a GPU at MSRP for almost two years, EVGA queue aside. Also, although GPU prices have indeed been falling as of late, it’s hard to predict what the situation will be when Nvidia’s next-gen launches in September, and Nvidia has also hinted it might sell Ampere side-by-side with Lovelace, so don’t give up hope…yet.

Continue reading

Nvidia Will Mimic AMD’s Smart Access Memory on Ampere: Report

AMD's Smart Access Memory hasn't even shipped yet, but Nvidia claims it can duplicate the feature.

Nvidia Unveils Ampere A100 80GB GPU With 2TB/s of Memory Bandwidth

Nvidia announced an 80GB Ampere A100 GPU this week, for AI software developers who really need some room to stretch their legs.

AMD Will Bring Smart Access Memory Support to Intel, Nvidia Hardware

AMD is reportedly working with Nvidia and Intel to bring hardware support for Smart Access Memory to other GPU and CPU platforms.

Tesla Ordered to Recall 150K+ Vehicles to Repair Memory Failures

Tesla has been asked — or "asked" — to recall some 159,000 vehicles to repair a NAND memory issue that will eventually cause failures on every affected vehicle.