Nvidia RTX 3080 and Ampere: Everything We Know

[harlan4096]

Here's what we know and what we expect from the Nvidia RTX 3080 / Ampere GPUs.

Nvidia's Ampere architecture will be the next major upgrade for GPUs from Team Green, and will find its way into the upcoming GeForce RTX 3080 Ti, RTX 3080, RTX 3070 and RTX 3060 graphics cards. Or perhaps Nvidia will throw us all a curveball again and change the model numbers. Whatever. Either way, we know Ampere is coming and will very likely arrive later this year, and the GPUs should rank high on our GPU hierarchy and list of the best graphics cards. Here's what we know about Ampere, including potential specifications, release date, price, features, and more.

First, it's important to note that all of the rumors and leaks of the past year or so are unconfirmed. Nvidia is normally very tight lipped about what it's working on, and the transition from the Turing to Ampere architecture is going to be particularly big for the company. Looking back at former major launches like RTX 20-series and GTX 10-series gives a good idea of what we can expect. In short, Nvidia won't say anything publicly until the last month before launch, and many of the supposed 'leaks' will turn out to be pure fabrications. Two examples: No one guessed correctly that GTX 1080 Ti would have 11GB of memory, and no one got the RTX 20-series name correct, or its inclusion of dedicated ray tracing hardware, until just a week or two before the unveiling.

We're as excited as anyone about Nvidia's next generation GPU architecture, but we also want to separate fact from fiction. There's precious little of the former that can be proven, and potentially plenty of the latter, so take everything with a grain of salt. Let's also point out that Ampere is critical for Nvidia, on many levels. Recently, in it's Super Spring laptops announcement, Nvidia revealed that "15 million RTX GPUs" have been sold. That sounds nice, but damn if that doesn't seem awfully low for a GPU architecture that's been around for over 18 months.

The problem is that Nvidia doesn't normally provide hard data on the number of units sold. The current Steam Hardware Survey suggests that there are about four times as many GTX 10-series GPUs in the wild as RTX 20-series GPUs, but the statistics behind Steam's survey are opaque at best so we can't be too sure about real figures. Regardless, the attitude of many with RTX 20-series was to "wait and see," with the sage advice being that the first generation of any new technology — ray tracing hardware, in this case — might be interesting, but generation two will be where it really takes off.

Ampere is ray tracing Gen2, in other words, and after a relatively slow start for ray tracing hardware and the RTX 20-series (from our perspective), Ampere has a lot to prove. The RTX 3080, RTX 3070, etc. (which is what we're calling them for now) need to provide not just better performance in games using traditional rendering techniques, but a dramatic increase in ray tracing performance would open the doors to doing more RT effects without tanking performance.

Nvidia GeForce RTX 3080 and Ampere At A Glance:

Up to 128 SMs / 8192 GPU cores (for RTX 3080 Ti)
Nvidia's first 7nm part should be much more efficient than Turing
Release Date: We expect to see Ampere in 2020, probably fall
Price: RTX 3080 likely to cost around $699-$799 (but we hope it's lower)

The Ampere Architecture in GeForce RTX 3080

Let's be real clear on this: We don't know for certain what Nvidia is going to change with the Ampere architecture. The only expectation is that it won't be a simple die shrink of Turing from 12nm to 7nm. Beyond that, there are many things Nvidia can potentially do to make the RTX 3080 and other Ampere GPUs better. Here's the quick roundup of some of the rumors.

First, the fundamental building block of Nvidia GPUs is called a Streaming Multiprocessor (SM). Its AMD analog is the Compute Unit (CU), and at a high level, it's relatively safe and easy to compare the two companies' GPUs based on SMs vs. CUs. The Turing architecture brought plenty of changes to the SM configuration, and it's a safe bet that Ampere will bring additional changes.

Turing added RT cores and Tensor cores, for ray tracing ray/triangle intersection calculations and deep learning FP16 calculations, respectively. Beyond the RT and Tensor cores, the CUDA core is the major GPU hardware in Nvidia graphics cards. For Turing, Nvidia switched from having 128 CUDA cores per SM to 64 CUDA cores. Turing also added a dedicated integer (INT) pipeline to each CUDA core, which allows for concurrent INT and FP (floating-point) calculations. Previously, a shader core would have to switch from doing FP to doing INT, which reduced overall efficiency and throughput. The Turing CUDA cores also added support for rapid packed math (FP16) calculations, which basically double the computational power of FP32 but with reduced precision — FP16 is useful for certain types of calculations.

We're really condensing everything that changed with Turing here, but beyond the above, there were changes to the L1/L2 cache, support for Variable Rate Shading (VRS), mesh shaders, Texture Space Shading (TTS), Multi-View Rendering (MVR), and enhancements to Simultaneous Multi-Projection (SMP). Most of those are now part of the official DirectX 12 Ultimate API, and also have support in VulkanRT. Oh, and the NVENC hardware got a major upgrade that added hardware accelerated encoding and decoding of higher resolutions and more codecs like VP9 and HEVC.

Conventional wisdom is that it's not a great idea to do massive architecture changes at the same time as a major change in lithography. Nvidia is already doing the lithography change going from TSMC 12nm FinFET to TSMC 7nm FinFET, so perhaps there won't be as many architectural changes as in Turing. However, we do expect the ratios of the various core types in the SM to be altered.

Looking at die shots of Nvidia's Pascal and Turing architectures, Turing appears to cram at least 20% more transistors into the SM. Those extra transistors are for the RT and Tensor cores, along with the other updates. It's probably more than 20% as well, since Pascal was 16nm and Turing is an optimized variant that TSMC named 12nm — it's part marketing, but the 16FF and 12FF nodes are definitely not the same. With the shrink to 7nm, Ampere should be able to add even more functionality, like doubling down on RT cores. Maybe Nvidia will also look at adding more Tensor cores for things like improved DLSS, but that seems far less likely.

Nvidia also mentioned with Turing that the concurrent INT and FP pipelines improved overall efficiency and performance by up to 35%, but that means in most scenarios there are now 'too many' INT cores. Nvidia could potentially rework the SM design to either double the FP units, or cut the number of INT pipelines in half, or some other tweak. AMD did something similar to this with its dual-CU design on the Navi 1x / RDNA 1 GPUs, and it wouldn't be surprising to see some adjustments to the Ampere SMs.

For now, this is all conjecture. We expected Nvidia to divulge at least some details related to Ampere at it's GPU Technology Conference (GTC) last month, but GTC was moved online in the wake of COVID-19, and the planned online streaming keynote from Nvidia's CEO ended up being cancelled altogether. At some point, we'll hear from Jensen Huang exactly what's happening with Ampere, hopefully sooner than later, but the coronavirus has thrown everything for a loop.
...

Continue Reading