Huawei Ascend 910: The Complete Guide

Huawei’s then-rotating-chairman Eric Xu described it in blunt terms at launch: “Ascend 910 performs much better than we expected. Without a doubt, it has more computing power than any other AI processor in the world” 2 — a claim made against the AI training chips available in 2019, not against anything shipping today. This page treats it as a launch-day marketing claim, not an independently verified benchmark.

Independent die-teardown analysis by Tom’s Hardware found the original Ascend 910’s compute die was manufactured by TSMC — the world’s most advanced contract chipmaker at the time 3. Every Ascend training chip that followed, starting with the 910B (covered on its own page on this site), was built on a different foundry once US restrictions cut off Huawei’s access to TSMC’s most advanced processes — making the original 910 the last Ascend training chip built on leading-edge, non-Chinese silicon.

Tom’s Hardware’s die analysis (not a Huawei-published figure) describes the compute die, which it names “Virtuvian,” as 456.25 mm², with a 1024-bit mesh network-on-chip running at 2GHz and roughly 128 GB/s of bandwidth per AI core 3. Huawei itself has not published a die size, transistor count, or detailed microarchitecture diagram for Ascend 910 — every architectural figure on this page beyond Huawei’s own launch-day compute claims comes from this third-party teardown, clearly labelled as such.

Launch-day compute and power figures per contemporaneous press coverage 12; die-level figures per independent teardown analysis, not a Huawei publication 3.

This is a press/analyst attribution, not a Huawei statement — Huawei has never named a foundry for any Ascend chip in any source this site has found, on this page or any of its siblings. Tom’s Hardware’s teardown-based conclusion that the original 910 was TSMC-made, using some EUV process layers consistent with N7+, is the basis for every foundry claim on this page 3. It matters specifically because the next generation, Ascend 910B, shows a different die signature that the same analysis attributes to SMIC — the foundry switch this page’s “hidden in plain sight” section below explains.

Ascend 910 was positioned from launch as the compute engine behind Huawei’s own AI cloud and training infrastructure, paired specifically with MindSpore, Huawei’s own machine-learning framework built to run on it 4. This site found no evidence Ascend 910 was ever sold to, or deployed by, a customer outside Huawei’s own systems and partners.

Consistent with every other Ascend chip on this site, Huawei has never published a standalone chip price. Any commercial arrangement for Ascend 910-based systems ran through Huawei’s own cloud and enterprise sales channels, not a published price list.

This site found no independent third-party benchmark of Ascend 910 against contemporary 2019-era training chips such as NVIDIA’s V100 (covered on its own page on this site). Treat Huawei’s launch-day framing as a marketing claim from the chip’s maker, in the same way this site treats every vendor’s own performance claims elsewhere.

Ascend 910 and Ascend 310 were announced together as opposite ends of one family in October 2018; Ascend 910 itself did not ship until August 2019 1. Its direct successor, Ascend 910B (covered with 910C on its own page on this site), is described by independent analysis as “massively different” — a different foundry, a larger die, and fewer AI cores despite being the next generation, not simply a faster version of the same chip 3.

Huawei was added to the US Commerce Department’s Entity List in May 2019 5 — three months before Ascend 910 launched. The restriction did not immediately block the chip that was already in production, but it reshaped every chip that came after it: Tom’s Hardware’s independent die comparison finds Ascend 910B’s compute die is larger (665.61 mm² vs. 456.25 mm²) with fewer AI cores (25 vs. 32) than the original 910. Tom’s Hardware’s analysis leaves the cause an open either/or rather than a verdict: either SMIC’s process has “significantly lower transistor density compared to TSMC’s N7+,” or Huawei substantially changed the core design itself 3 — either way, the clearest technical fingerprint, in silicon, of what losing access to the world’s most advanced foundry actually cost Huawei.