Nvidia Post-Volta (Ampere?) Rumor and Speculation Thread

Status
Not open for further replies.
Some unverified rumors about Ampere:
Since the beginning of this month Nvidia has been giving out the technical paper work to tech companies about the upcoming Ampere GPU's, they will be receiving the first reference boards and silicon by the end of this month so that they can start designing and testing the next AIB cards for the arch.

Good News:

-Significant raster performance increases (higher than Turing gains)
-Massive RT performance increases
-Expanded VRAM capacity on the full range of SKU's.
-Dialed down TDP's on many cards in the range (high end remains unchanged at 250w)
-Increased core clocks (healthy 100-200mhz bump over Turing)

Bad News:

-Prices, unchanged on the most popular cards, slight decreases on the unpopular ones. (x80ti and x80's will see a price drop, while x70's and beneath will remain UNCHANGED)
-Overclocking, core clocks stock will be higher than ever, 100-200mhz higher depending on SKU, however this does not mean the 2100-2200mhz ambient cooling upper limit of frequency will be broken when overclocking, which brings me to the next point.
-Voltage limits decreased. We went from 1.2v of Maxwell and previous generation to 1.093v in Pascal (maintained in Turing), we will now see a decrease to sub 1v in Ampere, which will severely limit overclocking frequency, but has a positive impact on power consumption's.

 
-Significant raster performance increases (higher than Turing gains)

I’m waiting for polygon output numbers of 5700 xt and 2080ti. @CarstenS do you have any values of the polygon performance?
 
-Significant raster performance increases (higher than Turing gains)

I’m waiting for polygon output numbers of 5700 xt and 2080ti. @CarstenS do you have any values of the polygon performance?

I can easily see getting bottlenecked by pixel output at high enough resolutions, especially as we head towards next gen consoles next year. Nvidia already had great raster performance, weird to see it be a focus.

Regardless the other things aren't surprising. New process, slightly higher frequencies at worse fmax curve isn't surprising. Pouring more into their specific raytracing is still a bit of a weird gamble to me, really reeeeally hoping RTX cards will be compatibility targets for devs once the new consoles come out. Not a terrible gamble though, as hard as Nvidia has been pushing it.

And of course there's the "more ram!" needed for to keep up with the new consoles. I feel bad for anybody purchasing a high end card this year.
 
I can easily see getting bottlenecked by pixel output at high enough resolutions, especially as we head towards next gen consoles next year. Nvidia already had great raster performance, weird to see it be a focus.

Going to be interesting to see what Nvidia can do when they actually get some heat from AMD and soon Intel.

And of course there's the "more ram!" needed for to keep up with the new consoles. I feel bad for anybody purchasing a high end card this year.

They most likely won't have any trouble, i think high end turing owners probably will get the better multiplatform games even.
 
Pouring more into their specific raytracing is still a bit of a weird gamble to me, really reeeeally hoping RTX cards will be compatibility targets for devs once the new consoles come out. Not a terrible gamble though, as hard as Nvidia has been pushing it.

What's Nvidia specific about the DirectX and Vulkan raytracing apis?
 
Its going to be interesting to see how nVidia spends the transistor budget with a node shrink. While I love the fact that AMD has made such good progress with their Navi design, they are still more or less at parity with Turing in rasterisation. A node shrink "pipe-cleaner" of Turing could potentially surpass Navi even by "simply adding" more RT cores to the package and increasing frequency. And they seriously need to increase the RT side of the equation. A 30-40% performance loss isn't acceptable in gaming scenarios, never mind how advanced and bleeding edge the technology is. If it's going to supplant or heavily displace traditional rasterisation (and it will), then that price/performance to entry doesn't cut it. At least in my mind. A slightly faster 2070 series replacement that doesn't lose more than say 5-10% of frames when you turn RT on for the price of a current 2070 series though? Well now...
 
NVIDIA "Ampere" Successor Reportedly Codenamed "Hopper"

https://www.techpowerup.com/261164/nvidia-ampere-successor-reportedly-codenamed-hopper

"Hopper" is also rumored to introduce MCM (multi-chip module) GPU packages, which are packages with multiple GPU dies. Currently, GPU MCMs are packages that have one GPU die surrounded by memory dies or stacks. This combination of GPU dies could make up "giant cores," at least in the higher end of the performance spectrum. NVIDIA reserves MCMs for only its most expensive Tesla family of compute accelerators, or Quadro professional graphics cards, and seldom offers client-segment GeForce products.
 
Last edited:
What is techpowerup considering a current MCM here? GP100 and GV100 that use HBM2?
 
What is techpowerup considering a current MCM here? GP100 and GV100 that use HBM2?
Seems like it. They're confusing MCM with on-package memory, which of course are very different. And by extension, they're stating that a new multi-GPU MCM would be limited to the professional cards.
 
Seems like it. They're confusing MCM with on-package memory, which of course are very different. And by extension, they're stating that a new multi-GPU MCM would be limited to the professional cards.
Technically they are "MCM's" too, as "multi-chip module" doesn't really specify what kind of chips we're talking about
 
Isn't what they are really talking about chiplets?
For the supposed followup to supposed Ampere, yes, but they're also calling the current GPUs with HBM2 MCM's in the bolded part of del42sa's message. And it's technically right, too.
 
NVIDIA Next Generation Hopper GPU Leaked – Based On MCM Design, Launching After Ampere
Not only that but you are looking at massive yield gains from just shifting to an MCM based approach instead of a monolithic die. A single huge die has abysmal yields, is expensive to produce and usually has high wastage. Multiple chips totaling the same die size would offer yield increases straight of the bat. This is a great argument in favor of the NVIDIA Hopper GPU.
AMD-Navi-GPU-Launching-in-2018-Could-Be-MCM-Based-1480x912.png

Let's see. According to the approximation, a 300mm wafer should be able to produce 114 monolithic dies (22x22) or 491 smaller dies (11x11). Since we need 4 smaller dies to equal 1 monolithic part, we end up with 122 484mm² MCM dies. That's a yield gain of 7.6% right there.

The yield gains are even larger for bigger chips. The upper limit of lithographic techniques (with reasonable yields) is roughly 815mm². On a single 300mm wafer, we can get about 64 of these (28.55x28.55) or 285 smaller dies (14.27x14.27). That gives us a total of 71 MCM based dies for a yield increase of roughly 11%. Now full disclosure, this is a very rough approximation and does not take into account several factors such as packaging yields, rectangular die and other shape-based optimization of the wafer etc but the basic idea holds well. Conversely, it also does not take into account increased gains by lowered wastage - a faulty 815mm² monolithic die is much more wasteful than a single 203mm² one! This means this approach has the added benefit of minimizing the impact of defective dies - which will add on to these yield numbers once you factor in unusable dies.
https://wccftech.com/nvidia-hopper-gpu-mcm-leaked/
 
Today at Supercomputing 2019 Jensen Huang will unveil Magnum IO. It’s a new interconnect technology aimed at supercomputers. We do not have any other details than provided below.

Quote from NVIDIA:

GPU-Accelerated I/O and Storage Software to Eliminate Data Transfer Bottlenecks for AI, Data Science, and HPC Workloads
High-Bandwidth, Low-Latency Massive Storage Access with Lower CPU Utilization
Delivers up to 20x faster data throughput on multi-server, multi-GPU computing nodes.
Compatible with a wide range of communications interconnects and APIs — including NVIDIA NVLink™ and NCCL, as well as OpenMPI and UCX — GPUDirect is composed of peer-to-peer and RDMA elements.
https://www.nvidia.com/en-us/data-center/magnum-io/
https://nvidianews.nvidia.com/news/...ks-for-data-scientists-and-ai-hpc-researchers

will Nvidia use it for Hopper MCM ?
 
Status
Not open for further replies.
Back
Top