THE NEWS
Nvidia’s 7nm monster GPU arrives
Nvidia’s new monster GPU, Microsoft has a change of heart over open source, and more.
NVIDIA HAS a new GPU architecture called Ampere, and it is big. In fact the first Ampere chip is the world’s biggest 7nm IC, with 54 billion transistors. It took, according to a “keynote” speech made by CEO Jensen Huang from his kitchen at home, the result of the efforts of thousands of engineers over several years. The first card to use the Ampere architecture will be the A100. This is aimed squarely at the data center where it will toil away at cloud services, data analysis, and artificial intelligence. This is the world of high-performance computing, and Nvidia’s press materials are dotted with bar charts showing Ampere’s capabilities above its previous card, the Tesla V100.
The A100 card will have, according to Mr. Huang, “a profound impact on how we architect data centers.” For properly serious work, Nvidia has put together the A100 system board, which carries eight A100 cards, making the world’s largest graphics card. Nvidia claims it has more transistors than have ever been put onto one computer. It’s a beast.
All this is well and good, but what about gaming cards with this wonder chip? They are coming, but here we enter a murky world of leaks and unofficial information gleaned from anonymous sources, so a pinch of salt is required first. Nvidia is notoriously careful with its test cards, including anti-tamper devices and “black box” logging. What we’ve seen so far purports to be a GeForce RTX 3080 Ti, sporting triple fans and two eight-pin power connectors, indicating its increased power efficiency. Inside is a GA102 GPU with 84 streaming processors and 5,376 CUDA cores. There’s expected to be GA103 and GA104 GPUs for GeForce 3000-series cards.
The card carries 12GB of memory with a bandwidth of 864Gb/s, 40 percent more than the RTX 2080. Exact frequency speeds are hard to calculate— the only software that can report it accurately is made by Nvidia—but the sample card apparently runs at over 2.2GHz on boost. The ray-tracing cores (missing on the A100) are much more efficient, by a factor of four. This should be the push ray-tracing gaming
needs; currently you have to tread carefully with the options not to hit the refresh rates or playability too hard. DLSS, or deep learning super sampling–which has offered so much yet somehow never quite managed it—now reaches version 3.0, and while still not a magic performance bullet, it works a lot better. The result of all this is a card that is at worst 40 percent faster than a RTX 2080 Ti, and on some titles 70 percent.
Even though all this is unconfirmed rumor, it is hard not to be excited. Ampere is going to be the card that can give 4K gaming the power it needs. When? If previous launches are anything to go by, Ampere will appear first on Tesla cards, then Quadro. These should appear by the fall. GeForce could possibly be in time for the holidays.
Meanwhile Intel is about to throw its hat into the graphics card ring. It has poured not inconsiderable resources into its own graphics engine, called Xe, and we aren’t too far away from seeing the first fruits, first in Tiger Lake mobile chips, but what is more interesting is its fully discreet graphics card. Intel is starting to tease with pictures and hints of “tens of billions” of transistors. It’s a big jump from capable but unprepossessing integrated graphics to building a fullblown gaming card, but Intel’s engineers seem confident.
Elsewhere, AMD is working on its next-generation GPU, currently known by many names: Big Navi, or Navi 2X, or RDNA 2, etc. It’s expected to bring a gain of about 50 percent per watt and employ up to 5,120 cores. The competition is fierce, and is only getting fiercer; before the end of this year we’ll have new benchmark records, and epic battles for the best graphics card in each sector.
If previous launches are anything to go by, Ampere will appear first on Tesla cards.