Word on the vine is that NVIDIA is planning to fabricate the next generation RTX 3000 Ampere GPUs on Samsung’s 8nm process. This would be somewhat unexpected as NVIDIA’s announcement for Ampere has previously stated that they will be using the 7nm TSMC process. The leaker in question, however, has an excellent track record for RTX 3000 series leaks and had the die configurations down pat months before anyone else. This is why even though I am tagging this as a rumor simply because of the sheer contrarian nature of this article; let’s just say I wouldn’t be surprised if it turns out to be true.
NVIDIA planning to tap Samsung’s 8nm process to manufacture for RTX 3090, RTX 3080 and RTX 3070
NVIDIA has already used TSMC’s 7nm process for its higher-end Ampere lineup but if this rumor is true [opinion] and I personally think this is going to turn out to be true [/opinion] then things are about to get very interesting. The TSMC 7nm EUV process is more energy-efficient than the Samsung 8nm process but the latter is likely cheaper and has much more spare production capacity than the former. TSMC’s high-end processes are usually in high demand and AMD seems to have set up shop there and dominating a very large supply of their wafers.
100% Samsung 8nm(develop from 10nm) https://t.co/jYqkSYfku7
— kopite7kimi (@kopite7kimi) July 2, 2020
Going to Samsung might be the logical move here considering 1) they have enough cash reserves to offer the process for cheap and 2) will have ample volume to meet NVIDIA’s demands. According to the tweet, the Samsung 8nm process is based on the 10nm backbone and will be 10% more energy-efficient than 10nm. 8nm will still be a massive jump in performance from NVIDIA’s 12NFF (which is essentially just an optimized version of 16nm) and the company could very easily supplement the lower energy efficiency by feeding it more power (that might explain the high TGP ratings for the RTX 3090 we saw earlier).
yep
— kopite7kimi (@kopite7kimi) July 2, 2020
We had initially heard that NVIDIA might decide to do a staggered launch with some high-performance parts (like the purported RTX 3090) going to TSMC 7nm and lower end parts going with Samsung but according to Kopite, all parts and segments are going to be based on Samsung’s 8nm Process and they are 100% sure about it. This means that RTX 3090, RTX 3080, 3070 and so on – all SKUs are going to be based on the 8nm process. Since the process is less efficient than TSMC’s NVIDIA would need to slightly increase the TGP across the board to still give the massive performance leap we have been seeing in leaked benchmarks (and boy is it something to behold).
The source has also previously stated that NVIDIA’s future Tegra processors are going to be based on the same process as well. NVIDIA felt the brunt of its dependency on TSMC for the first time when AMD booked all of spare 7nm capacity with its CPUs and GPUs. If I were in Jensen’s shoes, I would want to reduce dependency on the foundry as well and diversify away some of that risk. That said, shifting foundries is a *huge* decision that requires tons of R&D and changes in EDA and not something that is done lightly. A lot of capacity at TSMC has also been freed up with Huawei exiting the stage and it remains to be seen whether this information will turn out to be true.
At the end of the day, it really doesn’t matter whether we get the RTX 3000 series on the TSMC 7nm process or Samsung’s 8nm – what matters is performance and we fully expect them to retain their performance lead over AMD.
