Nvidia decided to host a series of pre-recorded videos each dedicated to a specific topic regarding the latest microarchitecture, Ampere, for the newest line of desktop and server graphics processors. During the event GPU accelerated data center processing, DLSS, RTX, GPU accelerated Scientific computing, Conversational AI and the A100 Ampere GPU were displayed and discussed. Pre-show Nvidia CEO Jensen Huang confirmed, prior to, the GTC 2020 Keynote that the latest Ampere microarchitecture will be used in all their upcoming graphics cards. Previous generations of Nvidia’s graphics processors used tiered microarchitecture, an example being Turing for the GeForce and Quadro line and Tesla. For newcomers, the aforementioned titles probably make you feel like an immigrant at the DMV, GeForce is the GTX/RTX gaming cards, Quadro is for consumer workstations and Tesla is their enterprise data center cards. Data-Center-Scale Accelerated Computing The Keynote started off with Jensen discussing data-center-scale accelerated computing and how medical professionals are using Nvidia’s current Turing architecture in large server-farm / data centers to advance cure’s and vaccines for diseases. It was quite fascinating as they explained how the raw performance of GPU’s (able to crunch gigabytes per second) were accelerating the way in which research fields can collect and compile large data sets when moving from CPU (can crunch megabytes per second) based data centers. DLSS & RTX After the data center opening presentation, Jensen moved onto discussing DLSS 2.0. Deep Learning Super Sampling (DLSS) uses machine learning to make a lower resolution image appear as a higher resolution image, 720p to 1080p, without using a lot of video card overhead. Super Sampling has been used for several years but was quite an expensive feature to implement even on high-end desktop cards. Nvidia compared the difference between the two versions of the deep learning technology. While DLSS 1.0 did not work very well, making the high-resolution images quite blurry, 2.0 works much better and they showed off a 1080p image that had the visual quality of a native 16k image, quite impressive. Conversational AI Jensen moved onto scientific computing and conversational AI, which I found to be extremely fascinating. While it will probably take another decade before Artificial Intelligence can perfectly mimic human interaction, the current level of progress Nvidia has made is far beyond what I would have originally thought. As Jensen points out in the conference; Conversation AI must be able to, in real-time, process speech, natural language understanding, text-to-speech while also rendering 3D models, facial animation and lip-syncing animations in fractions of a second. Conversation AI is a technology I am going to keep my eye on for the foreseeable future. A100 Ampere Nvidia eventually unveiled the A100 Ampere GPU, a data center video card. The A100 is designed to be used in large-scale server farms for AI deep learning and computing as well as robotics and autonomous vehicles. While it was disappointing not to get the consumer cards revealed, we can take a lot away from the A100 GPU as it will be scaled down for the 3000 series cards. Official Specifications of the A100 Ampere GPU: FP32 CUDA Cores – 6912 Boost Clock – ~1.41GHz Memory Clock – 2.4Gbps HBM2 Memory Bus Width – 5120-bit Memory Bandwidth – 1.6TB/sec VRAM – 40GB Single Precision – 19.5 TFLOPs Double Precision – 9.7 TFLOPs (1/2 FP32 rate) INT8 Tensor – 624 TOPs FP16 Tensor – 312 TFLOPs TF32 Tensor – 156 TFLOPs Interconnect – NVLink 3 12 Links (600GB/sec) GPU – A100 (826mm2) Transistor Count – 54.2B TDP – 400W Manufacturing Process – TSMC 7N Interface – SXM4 Architecture – Ampere Consumer Cards While the next-gen GeForce Cards were not shown off during GTC, we could see announcements of Quadro and GeForce at SIGGRAPH or Gamescom, both having digital events in August. The final venue for possible disclosures is Computex in September, which will fall in line with CD Projekt’s blockbuster Cyberpunk 2077 which Nvidia has been aggressively supporting. The good news, despite a late in the year, reveals, the first set of Add-in-Board partners (EVGA, ASUS, MSI) 3000 series GPU’s will be available at retail just mere weeks (if not launched simultaneously with the Founders) after the Founders (direct from Nvidia) launch. Buy or Wait? An interesting question for consumers now is whether to buy or wait for the new GPUs. If I had an older GPU that was not keeping up with modern titles or if I were in the process of building a PC and need a budget GPU, then the current offerings are more than capable of accommodating current needs. With the launch of the new cards there is no guarantee that they will release at the most desired price points or SKU’s. When Turing was introduced the only available cards were the 2070 ($500), 2080 ($800), and 2080Ti ($1200), with the more affordable mid-range GPUs coming in around four to six months later. I am lucky to be in the camp of being able to wait as I am currently rocking a 2070, which is a great card, but can’t adequately keep up with my monitors (120Hz) refresh rate at 3440×1440 resolution without turning some quality settings down or run RTX at a decent framerate. Conclusion Overall Nvidia’s GTC 2020 Keynote was an interesting conference as it focused on AI deep learning, data center and scientific accelerated computing, RTX and DLSS, and conversational AI. While Nvidia showed off the A100 Ampere GPU, it, unfortunately, is a data center GPU. I was hoping Nvidia would also unveil the next generation of consumer RTX cards, rumored to be the 3000 series of cards. Luckily Nvidia is going to have other conferences throughout the summer, I am keeping my fingers crossed they unveil their consumer lineup sooner rather than later, as current rumors are painting some impressive hardware.