NVIDIA has allegedly gone on an enormous spending spree to accumulate a few of TSMC’s next-gen 5nm wafer provide for its next-gen GeForce RTX 40 ‘Ada Lovelace’ GPUs.
NVIDIA Goes on TSMC 5nm Spending Spree, Billions of {Dollars} Paid To Purchase Wafers For GeForce RTX 40 ‘Ada Lovelace’ GPUs
NVIDIA’s Ada Lovelace GPUs powering the next-generation GeForce RTX 40 sequence graphics card lineup are anticipated to make the most of TSMC’s 5nm course of node. Each AMD and NVIDIA are anticipated to make the most of the node for its next-gen lineup but it surely appears to be like like NVIDIA may be very critical in ensuring they get sufficient wafer provide for its lineup and as such, they’ve paid the Taiwanese semiconductor producer a number of Billions of {dollars} as a sophisticated fee for 5nm wafers.
In accordance with business sources, TSMC’s necessities for Apple, MediaTek, AMD and different three main prospects are comparatively low. They don’t must pay an excessive amount of deposit upfront to stabilize manufacturing capability. Clients like NVIDIA must pay big advance funds upfront in the event that they wish to acquire 5nm manufacturing orders.
MyDrivers reports that NVIDIA has pay as you go TSMC round $1.64 Billion US in Q3 2023and can pay $1.79 Billion US in Q1 2023. The entire long-term ‘Multi-Billion’ greenback deal is about to price NVIDIA an insane $6.9 Billion US which is far larger than what they paid final yr. NVIDIA won’t simply use this cash to acquire wafer provide from TSMC but in addition from Samsung but it surely appears to be like like that almost all of the quantity can be spent on TSMC’s 5nm expertise.
NVIDIA GeForce RTX 4090 Graphics Card – Ada Lovelace Powered AD102 Flagship GPU
Primarily based upon earlier rumors, there have been whispers that NVIDIA would make the most of TSMC’s N5 (5nm) course of node for its Ada Lovelace GPUs. This contains the AD102 SKU too which can be a completely monolithic design. Speaking about particular GPU configurations, the flagship AD102 GPU is alleged to characteristic a clock pace as excessive as 2.5 GHz (2.3 GHz common enhance). The precise tweet states that the GPU clock for Ada Lovelace ‘AD102’ could possibly be 2.3 GHz or higher so let’s take that as a baseline and beforehand leaked specs to determine the place the efficiency ought to land.
The NVIDIA AD102 “ADA GPU” seems to have 18432 CUDA Cores primarily based on the preliminary specs (which may change), housed inside 144 SM items. That is virtually twice the cores current in Ampere which was already a large step up from Turing. A 2.3-2.5 GHz clock pace would give us as much as 85 to 92 TFLOPs of compute efficiency (FP32). That is greater than twice the FP32 efficiency of the prevailing RTX 3090 which packs 36 TFLOPs of FP32 compute energy.
The 150% efficiency soar appears to be like big however one ought to do not forget that NVIDIA already gave an enormous soar in FP32 numbers this technology with Ampere. The Ampere GA102 GPU (RTX 3090) provides 36 TFLOPs whereas the Turing TU102 GPU (RTX 2080 Ti) supplied 13 TFLOPs. That is over a 150% improve in FP32 Flops however the real-world gaming efficiency improve for the RTX 3090 averaged at round 50-60% quicker over the RTX 2080 Ti. So one factor we should not neglect is that Flops do not equal GPU gaming efficiency as of late. Moreover, we do not know if 2.3-2.5 GHz is the typical enhance or the height enhance with the previous which means that there could possibly be even larger compute potential for AD102.
Except for that, leaks have additionally acknowledged that the NVIDIA GeForce RTX 40 flagship would retain a 384-bit bus interface, just like the RTX 3090. What’s attention-grabbing is although that the leaker mentions G6X which implies that NVIDIA will not be transferring to a brand new reminiscence normal till after Ada Lovelace and make the most of the upper pin-speeds of G6X of 21 Gbps for its next-generation playing cards earlier than we see a more moderen normal (e.g. GDDR7). The cardboard will characteristic 24 GB of reminiscence so we are able to both anticipate single-sided 16Gb DRAM or dual-sided 8Gb DRAM modules.
NVIDIA CUDA GPU (RUMORED) Preliminary:
GPU | TU102 | GA102 | AD102 |
---|---|---|---|
Structure | Turing | Ampere | Ada Lovelace |
Course of | TSMC 12nm NFF | Samsung 8nm | 5nm |
Graphics Processing Clusters (GPC) | 6 | 7 | 12 |
Texture Processing Clusters (TPC) | 36 | 42 | 72 |
Streaming Multiprocessors (SM) | 72 | 84 | 144 |
CUDA Cores | 4608 | 10752 | 18432 |
Theoretical TFLOPs | 16.1 | 37.6 | ~90 TFLOPs? |
Reminiscence Sort | GDDR6 | GDDR6X | GDDR6X |
Reminiscence Bus | 384-bit | 384-bit | 384-bit |
Reminiscence Capability | 11 GB (2080 Ti) | 24 GB (3090) | 24 GB (4090?) |
Flagship SKU | RTX 2080 Ti | RTX 3090 | RTX 4090? |
TGP | 250W | 350W | 450-650W? |
Launch | Sep. 2018 | Sept. 20 | 2023 (TBC) |
The NVIDIA Ada Lovelace GPUs will energy the next-generation GeForce RTX 40 graphics playing cards that can go head-on with AMD’s RDNA 3 primarily based Radeon RX 7000 sequence graphics playing cards. There’s nonetheless some hypothesis relating to the usage of MCM by NVIDIA. The Hopper GPU, which is primarily aimed on the Datacenter & AI section, is allegedly taping out quickly and can characteristic an MCM structure. NVIDIA will not be utilizing an MCM design on its Ada Lovelace GPUs so they may hold the normal monolithic design.
Information Supply: HardwareTimes