The graphics card, the H20, is the most powerful of three
Nvidia has been developing for the Chinese market after the U.S. expanded bans
on high-end chip exports, aiming to hamper China's ability to develop AI and
sophisticated computers for its military.
The H20 will naturally deliver less computing power than
Nvidia's flagship H100 AI chip and the H800 - the later China-specific card
that was also banned in October.
But specifications for the H20 also appear to indicate it is
less powerful than Huawei's (HWT.UL) Ascend 910B in some key areas, according
to three sources, who were not authorised to speak to media and declined to be
identified.
Nvidia has in recent weeks been pricing orders for H20
distributors in China in a range of $12,000 to $15,000 per card, according to
two sources.
Some distributors have started advertising the chips with a
significant markup to the lower end of that range at about 110,000 yuan
($15,320), one of the sources said. By comparison, Huawei's 910B is being sold
for around 120,000 yuan, two of the sources said.
Nvidia declined to comment.
One source said distributors are offering H20 servers, which
are pre-configured with 8 of the AI chips, for 1.4 million yuan. By comparison,
servers that used 8 of the H800 chips were sold at around 2 million yuan when
they were launched a year ago.
Distributors have told clients they will be able to start
delivering the H20 products in small batches in the first quarter of 2024 and
in larger quantities from the second quarter, the source added.
Before the U.S. curbs, Nvidia dominated China's AI chip
market with more than 90% share. However, it currently faces increasing
competition from domestic rivals, chief among them being Huawei.
Huawei's 910B chip is widely considered the most competitive
AI offering now available within China and has become more popular amid concern
that buyers could be faced with further restricted access to Nvidia's products
resulting from U.S. sanctions.
In terms of specifications, one example of where the H20
appears to lag the 910B in its FP32 performance - a critical metric that
measures how quickly a chip can process common tasks and which is rated at less
than half of its rival's capability, one source said.
However, the H20 appears to have an advantage over the 910B
in terms of interconnect speed, which measures how quickly data can transfer
between chips, according to the source.
That means the H20 remains competitive with the 910B in
applications that require linking a large number of chips together to work as a
system, he said.
Reuters reported last month that Nvidia plans to begin mass
production of the H20 in the second quarter of this year.
It was originally scheduled for launch last November but
that plan was delayed, with sources saying at the time that the delay was due
to issues that server manufacturers were having in integrating the chip.
Nvidia also plans to roll out two other China-specific
chips, the L20 and the L2. Reuters was not able to determine the current status
of the rollout for those two chips. Neither the H20, the L20 or the L2 are
currently listed on Nvidia's website.
Underscoring the importance of the China market to Nvidia,
CEO Jensen Huang visited the company's offices in Shenzhen, Shanghai and
Beijing last month for the company's annual parties before the Lunar New Year
holiday, local media reported.
0 comments:
Post a Comment