Roundup Here's your quick roundup of AI news beyond what we've already written about this week.
Crypto-mining is bust for Nvidia: Nvidia reported strong growth in its second quarter financial results, emitted this week, in areas expected like the data center and gaming – but not for cryptocurrency mining. Here is a quick snapshot of its Q2 earnings for fiscal 2019, aka the three months to July 29:
- Revenues reached $3.12bn, up 40 per cent from a year ago.
- Net income was $1.1bn, an increase of 89 per cent from the previous year.
- $1.76 earnings per share up 91 per cent year-on-year.
The demand for GPUs grew 40 per cent from last year to account for $2.66bn in sales, we're told. Popular online titles such as Fortnite and PUBG have helped Nvidia in the gaming department, which grew 52 per cent in terms of revenue to $1.8bn. The boom in deep learning is also accelerating its data center business by 83 per cent, to $760m, where its graphics cards are used as math accelerators. Nvidia’s automotive area is smaller with $161m in revenues, up 13 per cent year-over-year. Its professional visualization arm grew 20 per cent to $281m.
It was weakest in cryptocurrency mining. People just aren't buying Nvidia cards for crafting digital fun bucks any more, relatively speaking, and won't for a while, it seems. So that's good news for folks unable to get hold of an Nvidia card due to hoarding by crypto-coin nerds.
“Our revenue outlook had anticipated cryptocurrency-specific products declining to approximately $100 million, while actual crypto-specific product revenue was $18 million, and we now expect a negligible contribution going forward,” the biz reported during its the earnings call with analysts on Thursday.
A few months back CEO Jensen Huang said a shortage of its chips – particularly the GeForce series – was down to mining Ethereum. The prices skyrocketed for a brief period of time, have been declining, and are going back to normal levels. Huang previously said Nvidia were not targeting the crypto industry, and wanted to reserve GeForce parts for gamers.
When he was asked about how much the gaming industry was impacted by crypto in the earnings call, Huang said: “A lot of gamers at night, they could - while they’re sleeping, they could do some mining. And so, do they buy it for mining or did they buy it for gaming, it’s kind of hard to say.
“And some miners were unable to buy our OEM products, and so they jumped on to the market to buy it from retail, and that probably happened a great deal as well. And that all happened in the last - the previous several quarters, probably starting from late Q3, Q4, Q1, and very little last quarter, and we’re projecting no crypto-mining going forward.”
Nvidia’s stocks are down more than 4.90 per cent at the time of writing.
More Nvidia news... a new GPU architecture: Nvidia also announced a new GPU line: the Quadro RTX with a new architecture at this year’s SIGGRAPH conference.
All Quadro GPUs are the first to have the Turing architecture, and are aimed at intensive computer graphics. “Users can now enjoy powerful capabilities that weren’t expected to be available for at least five more years,” claimed Bob Pette, vice president of professional visualization at Nvidia, earlier this week.
"Designers and artists can interact in real time with their complex designs and visual effects in ray-traced photo-realistic detail. And film studios and production houses can now realize increased throughput with their rendering workloads, leading to significant time and cost savings."
The Turing architecture has up to 4,608 CUDA cores, and can handle up to 16 trillion floating point operations per second in parallel for high-end real-time ray tracing. Nvidia’s NVLink can also be used to connect two GPUs to increase the total memory capacity to 96GB with a 100GB/s link for data transfer.
Prices aren’t cheap, and start at around $2,300 for the Quadro RTX 5000 with 16GB of RAM, the Quadro RTX 6000 rises to $6,300 for 24GB of memory, and finally, the Quadro RTX 8000 is a whopping $10,000 for twice the RAM at 48GB.
Hey Siri scouts out your location to understand you better: Hey Siri, where’s the nearest BART or Tube station? And hey Siri, did you know you're using my location to improve your speech-recognition systems?
Apple's voice-controlled pal is more likely to recognize words like Starbucks or Walgreens, and is less likely to understand what you’re referring to if it’s your corner pizza joint or hairdresser. It’s a known performance bottleneck, and Apple has tried to improve it by using geolocation-based language models (Geo-LM).
“These models enable Siri to better estimate the user’s intended sequence of words by using not only the information provided by the acoustic model and a general language model (like in standard automatic speech recognition) but also information about the POIs in the user’s surroundings,” the Siri Speech Recognition Team said this week.
POIs being point of interests. A user’s location is assigned to a specific Geo-LM based on regions found in the US Census Bureau – Apple said it has most areas of America covered. The Geo-LMs are paired with its normal acoustic model for speech recognition to create.
Both are trained separately: the acoustic model uses 5,000 hours of English speech data, and the Geo-LMs are “sources of collected, privacy-preserved, live usage data.” For one of the test datasets, developers collected the names of POIs from Yelp reviews, and reduced the word error rate by 40 per cent.
You can read more about it right here. ®