Next-Gen RTX 4000 GPUs From NVIDIA to Use TSMC’s 5nm Process Node

It feels like just yesterday we saw the disastrous launch of current-gen GPUs and many of us are still struggling to get our hands on one. Even though stock availability has improved considerably since pandemic highs, prices are at an all time worse. The (sub) $200 price bracket for graphic cards seems to have been completely abandoned by manufacturers and mid-range is apparently now the point of entry into this new generation of GPUs.

But NVIDIA and AMD seem completely oblivious to all this and are already preparing to launch the next generation of their graphic cards in Q3/Q4 2022. This is technically not unusual as with recent GPU launches, we’ve seen both companies move to a 2-year refresh period where a new generation of graphic cards comes out every two years. RTX 30-series saw a September 2020 release date so, Q3/Q4 2022 will mark a full two-year gap and NVIDIA will launch the RTX 40-series accordingly.

RTX “Ada Lovelace” 4000

The RTX 4000 series of GPUs will be based on the “Ada Lovelace” architecture. Where NVIDIA’s datacenter GPUs (Hopper) are moving away from standard monolithic designs to MCM, the consumer-level Ada Lovelace cards will retain the monolithic architecture. Moreover, Ada Lovelace is set to be similar to Ampere (RTX 3000) in several other ways. All RTX 4000 graphic cards will feature GDDR6X memory, just like RTX 3000, and 21Gbps will be the maximum clock speed seen in the flagship AD-102 SKU. The upcoming RTX 3090 Ti is also set to feature 21Gbps GDDR6X memory from Micron.

Moreover, the memory bus inside the RTX 40-series flagship will be 384-bit wide, same as the RTX 3090 and 3090 Ti. We’re unsure if the AD-102 GPU will use 1GB (8Gb) or 2GB (16Gb) memory modules as GDDR6X memory was only available in 1GB modules up until recently. The RTX 3090 Ti will be the first GPU to utilize 2GB GDDR6X modules, halving the total amount of modules needed to reach the 24GB VRAM capacity. Speaking of which, the RTX 4000 flagship will also carry the same 24GB memory configuration.

Hello, TSMC

As we all know, NVIDIA tapped Samsung to manufacture its RTX 30-series GPUs. All RTX 3000 graphic cards are fabricated using Samsung‘s 8nm process, which is significantly inferior to TSMC‘s 7nm process seen in AMD RX 6000 cards. Adding insult to injury, Samsung’s yields have proven to be notoriously bad but NVIDIA chose the Korean giant because of its quicker delivery times, something that TSMC simply cannot even think about considering how much they have on their plate.

That’s all about to change with Ada Lovelace. According to a recent DigiTimes report, NVIDIA will be using TSMC’s 5nm process node to manufacture its RTX 40-series GPUs. This 5nm node is set to bring massive improvements over Samsung’s 8nm node. Specifically speaking, a 1.8x increase in transistor density and ~30% decrement in power consumption. Ada Lovelace is still going to be a power leech, though, as the RTX 4000 flagship is rumored to have a TDP of 500W.

Thank you @chiakokhua, very cool!

The actual DigiTimes report is hidden behind a paywall, but Twitter-user RetiredEngineer was able to supply us with a little snippet of the information inside. As you can see in the tweet below, plans are in full-motion in Taiwan for the production of next-gen GPUs. Plus, we also learn that apart from the Ada Lovelace RTX 40-series cards, NVIDIA’s datacenter GPUs will also be manufactured using TSMC’s 5nm process node.

"Nvidia's biennial GPU refresh coming in 2022, riding on metaverse and gaming. Following H100, based on Hopper architecture, using TSMC's 5nm + CoWoS, aimed at datacenter/AI, gaming GPU RTX40 series, based on Ada Lovelace architecture, will also tap TSMC's 5nm…."

— RetiredEngineer® (@chiakokhua) November 30, 2021

NVIDIA going back to TSMC shouldn’t come as a surprise, however, as the company has a rich and close history with the Taiwanese semiconductor giant. TSMC has manufactured GPUs for NVIDIA’s for ages and even the previous-gen RTX 20-series cards were manufactured by TSMC. NVIDIA turned down TSMC for Ampere because it simply was the most logical option. Samsung not only offered cheaper (albeit worse) silicone but they actually had stock ready to be delivered quickly. But, the inventory problem seems to be sorted now as NVIDIA is returning back to TSMC.

On top of that, this switch in manufacturers was hinted at from August of 2021. News of NVIDIA going back to TSMC leaked out earlier this year and several reports followed the initial leak, only further confirming the likelihood of this. And, now, it’s all but officially confirmed that the RTX 40-series will be manufactured using TSMC’s 5nm process. That being said, it must be awkward for NVIDIA going back to the friend it just ditched.