- Joined
- Jun 13, 2005
- Messages
- 61,738
- Reaction score
- 26,014
Inside Nvidia's new hardware for Switch 2: what is the T239 processor?
A new, custom Tegra heavy on machine learning and ray tracing.
The Digital Foundry just dropped an article where they conclude a 2-year-old rumor about the Switch 2's core processing unit is almost certainly correct, and the chip in question is destined for Nintendo's next machine. If they're right, the obvious takeaway is that Nintendo sees no reason to change things up. It will once again be going with a custom NVIDIA SoC (system-on-chip) variant of a broader chip they'll manufacture for the industry at large; just as the original Nintendo Switch ran on a "custom" variant of their GM20B graphics chip from its Maxwell generation, better known as the Tegra X1, and then a refinement of that same chip on the improved 16nm fabrication process with a 2019 update.
However, as DF points out, that custom version of the Tegra X1 wasn't actually customized at all. That's why, for example, it was so easy for hackers to run the Nintendo OS on the NVIDIA Shield TV, for example. This is where the new chip will differ-- as it must. The general version of the chip is simply too large.
So DF projects the expected T239 for the Switch 2 to have 1536 CUDA Cores. Among its current generation of Ada Lovelace GPUs, that constitutes half the cores in the RTX 4060. They also project a maximum 102 GB/s memory bandwidth. That would be far less than the RTX 4060 (272 GB/s). In fact, in both those respects, it falls far short of the RTX 4050 Mobile: the least of NVIDIA's current laptop GPU lineup. And, in reality, it's more Ampere (older) than Ada Lovelace (newer).
So ultimately, DF attempts to give an idea of what level of performance to expect. They're forced to do some fuzzy math, but the writer offers the following laptop as a performance analogue. He predicts that DLSS will be critical-- baked in-- to the Switch 2 and the next generation of handheld console gaming for games to run smoothly.
TLDR Cliffs: expect the Nintendo Switch 2 to offer roughly the same raw power as the RTX 2050 Mobile.
*Edit* Oh, I thought I'd add something. As far as comparing to first Switch to this Switch 2, know that the first Switch has 256 CUDA cores (compared to the 1536 that will be in the Switch 2), and the memory bandwidth of the first Switch peaks at 25.6 GB/s (compared to 102.0 GB/s projected for the Switch 2).
A new, custom Tegra heavy on machine learning and ray tracing.
The Digital Foundry just dropped an article where they conclude a 2-year-old rumor about the Switch 2's core processing unit is almost certainly correct, and the chip in question is destined for Nintendo's next machine. If they're right, the obvious takeaway is that Nintendo sees no reason to change things up. It will once again be going with a custom NVIDIA SoC (system-on-chip) variant of a broader chip they'll manufacture for the industry at large; just as the original Nintendo Switch ran on a "custom" variant of their GM20B graphics chip from its Maxwell generation, better known as the Tegra X1, and then a refinement of that same chip on the improved 16nm fabrication process with a 2019 update.
Digital Foundry said:Way back in June 2021, noted technology leaker kopite7kimi posted a detailed picture of Nvidia's T234 processor, revealing for the first time that Nintendo would be receiving a customised variant, dubbed T239. In the two years that followed, a wealth of overwhelming evidence has essentially confirmed that they were right. The T239 is an advanced mobile processor, based on an octo-core ARM A78C CPU cluster, paired with a custom graphics unit based on Nvidia's RTX 30-series Ampere architecture, combined with some backported elements from the latest Ada Lovelace GPUs - and with an all-new file decompression engine for fast engine. It also supports Nvidia's console-specific graphics API, all but confirming that it's destined for the next generation Switch.
However, as DF points out, that custom version of the Tegra X1 wasn't actually customized at all. That's why, for example, it was so easy for hackers to run the Nintendo OS on the NVIDIA Shield TV, for example. This is where the new chip will differ-- as it must. The general version of the chip is simply too large.
Let's be clear here: the existing T234 is a monster of a chip. With a die size of 455mm2, it dwarves the Xbox Series X processor at 360mm2. It's fabricated using the same Samsung 8nm technology used for the RTX 30-series cards, so it's actually a step behind the 7nm and 6nm processes used in current-gen consoles. CPU-size, it features 12 ARM A78AE CPU cores, paired with a GPU based on the 30-series Ampere architecture with 2048 CUDA cores and a 256-bit memory interface. For the automotive and robotics market, T234 requires a lot of machine learning performance, so there's also a deep learning accelerator built in - and its capabilities can be augmented by the tensor cores within the GPU.
So DF projects the expected T239 for the Switch 2 to have 1536 CUDA Cores. Among its current generation of Ada Lovelace GPUs, that constitutes half the cores in the RTX 4060. They also project a maximum 102 GB/s memory bandwidth. That would be far less than the RTX 4060 (272 GB/s). In fact, in both those respects, it falls far short of the RTX 4050 Mobile: the least of NVIDIA's current laptop GPU lineup. And, in reality, it's more Ampere (older) than Ada Lovelace (newer).
Memory bandwidth is crucial for a mobile gaming machine and it's the primary performance bottleneck in the current Switch. The 256-bit interface in T239 is wildly extravagant for a mobile games machine, and Nvidia's Linux update suggests a 128-bit interface, almost certainly paired with LPDDR5 memory. We should expect absolute maximum bandwidth of 102GB/s, but of course, Nintendo can choose to downclock that for improved efficiency.
So ultimately, DF attempts to give an idea of what level of performance to expect. They're forced to do some fuzzy math, but the writer offers the following laptop as a performance analogue. He predicts that DLSS will be critical-- baked in-- to the Switch 2 and the next generation of handheld console gaming for games to run smoothly.
What kind of performance should we expect from T239?
...To get some idea of what a mobile Ampere-class processor can do, I turned my hand to some practical work. There's no real equivalent counterpart for the capabilities of the A78C CPU cluster in the PC space, but when it comes to the GPU, we can get close. I bought in a Dell Vostro 5630 comes equipped with the following specs: a Core i7 1360p CPU, 16 gigs of 4800MHz LPDDR5, a 512GB SSD and an RTX 2050. It's that last component we're looking at more closely here...let's be clear: this is the closest approximation we can get together for the T239 GPU, but more accurately, what you're seeing in the video is an ultra low-spec Ampere GPU running at meagre clock speeds, starved of memory bandwidth. It can only really give us a ballpark idea of what a mobile Ampere chip can deliver, even when I downclocked it to a ridiculously low 750MHz - but even so, the results are intriguing.
I also tried Death Stranding, benchmarking its GPU-heavy intro. At 1080p native, the sequence averages at 34.9fps, which does support the last-gen horsepower narrative. At 720p, frame-rate rises to 52.5fps.
TLDR Cliffs: expect the Nintendo Switch 2 to offer roughly the same raw power as the RTX 2050 Mobile.
*Edit* Oh, I thought I'd add something. As far as comparing to first Switch to this Switch 2, know that the first Switch has 256 CUDA cores (compared to the 1536 that will be in the Switch 2), and the memory bandwidth of the first Switch peaks at 25.6 GB/s (compared to 102.0 GB/s projected for the Switch 2).
Last edited: