Intel introduced a new line of microprocessor chips for corporate servers and cloud data centers on Tuesday, with a larger claimed boost in performance than expected.
Under siege on all sides in its lucrative server chip business, Intel said its new Xeon Scalable processors were 1.65 times as fast on average as its prior generation at common tasks run by servers. The huge jump included tweaks to software specially designed to speed up tasks with Intel’s chips and will have to be confirmed by outside experts. But Wall Street and the industry had been anticipating a more modest performance increase.
“This represents the best of our 20-year history of data center innovation,” Lisa Spelman, vice president of Intel’s data center group, told Fortune ahead of Tuesday’s presentation, noting that the prior chip generation just came out in March last year. “That’s huge. That’s a lot to continue to deliver on the performance beat rate to our customers.”
Several customers speaking at Intel’s event on Tuesday disclosed more modest improvements. AT&T (t) chief strategy officer John Donovan said the telecommunications giant had seen a 30% performance increase after deploying new Xeon chips.
Still, Intel’s over 90% share of chips sold in computers that run corporate servers and cloud data centers remains under fierce attack. CEO Brian Krzanich has committed to improve server chips more quickly than in the past, though at the expense of profit margins.
But Nvidia is pitching its graphics processing cards as a better solution for running the kinds of artificial intelligence and machine learning tasks that are becoming increasingly popular. And Advanced Micro Devices last month launched a line of Intel-compatible chips called Epyc that it said provided better performance while saving energy and costing less. Some data centers owners like Google are working on making their own custom chips, as well.
Get Data Sheet, Fortune’s technology newsletter.
Investors have been growing increasingly concerned about the challenges, particularly after Intel reported a revenue increase of only 6% for its server chip business in the first quarter. On Monday, Jefferies analyst Mark Lipacis slapped a “sell” rating on Intel’s stock. “As the incumbent with dominant share, we think (Intel) has the most to lose,” Lipacis wrote.
Shares of Intel, which were down 7% so far this year, were about unchanged on Tuesday after Intel unveiled the new Xeons. Shares of competitor Nvidia, which have gained 44% so far this year, were also about unchanged and shares of AMD, up 22% in 2017, rose 1%.
Intel said its new chips increased performance in part by adding more fast connections between different components on each chip, such as the processing cores and the memory storage. The new “mesh architecture” sounded a lot like AMD’s “infinity fabric” that it touted in the Epyc chip along similar lines.
Although Intel was formally introducing the new Xeon chips on Tuesday, it has already sold about 500,000 to some large customers such as Google (googl) and allowed many more to benchmark the performance of the chips. At an event in New York on Tuesday, Intel planned to report some of the customer benchmarks, including that Oracle (orcl) saw 1.3 times faster performance running its 12c database, Baidu (bidu) saw 1.7 times faster search performance, and Ericsson (eric) speeded up video encoding by 1.5 times. All of the improvements included the impact of both the new chips as well as software tweaks related to the chips.
And although Intel did not have access to AMD’s new Epyc chips, the company said comparable versions of its new Xeon Scalable line matched or beat the performance of Epyc chips on the benchmarks tests that AMD revealed last month. AMD has said it will offer single chip packages, known as a one-socket product, that will serve the needs of many customers who have been buying double chip, or two socket products, from Intel, potentially generating big cost savings.
Intel says it doesn’t think the AMD (amd) one-socket chips will be up to the task. “They may be pushing that one-socket conversation so hard not because it’s in the best interests of their customers or because one-socket solutions address all data center needs, but because they’ve achieved a good result on a benchmark,” Spelman said.
The company has been less chatty about the challenge posed by graphics chips from Nvidia (nvda) (and AMD is launching a new line of souped up graphics chips dubbed Vega shortly, as well). The Xeon chip can do some AI tasks well, but Intel’s (intc) more direct answer to the graphics chip challengers is still in the works. The company is integrating technology it acquired buying two-year-old AI chip startup Nervana Systems last year for a new line of chips dubbed Lake Crest.
“Intel is investing in much more optimized silicon,” Spelman says.