inference/research
verifiedmedium · 0.80updated 2026-10-05

Huawei Atlas 300I Duo inference card (dual 310-series, 48/96 GB)

asic-other da-vinci (310-series, dual) 2022

Fields

id
accelerators/huawei-atlas-300i-duo
Vendor
asic-other
Architecture
da-vinci (310-series, dual)
Release year
2022
Process (nm)
12
Form factors
pcie
VRAM (GB)
96
Memory type
lpddr4x
Memory bandwidth (GB/s)
408
Bandwidth basis
Vendor-published, and the CARD basis is stated on the page itself: 'LPDDR4X 96GB或48GB，总带宽408GB/s，支持ECC' - LPDDR4X, 96 GB or 48 GB, TOTAL bandwidth (entire card) 408 GB/s, ECC. Two bins, one bandwidth figure: the 408 GB/s belongs to the card and the 48 GB bin is an alternative configuration, so a record must pick one capacity bin and say which rather than quoting both. Per-processor bandwidth is not stated but follows as 408/2 = 204 GB/s, consistent with the single-processor 300I Pro's published 204.8 GB/s. Source w5a-huawei-atlas-300i-duo-e-huawei-spec-table.
Peak FLOPS
precisionint8tflops280denseyesvendor_claimyessource_idw5a-huawei-atlas-300i-duo-e-huawei-spec-tablebasisvendor-spec-sheetbasis_detailHuawei's own product page, 技术规格 table, verbatim '| AI算力 | 280 TOPS INT8 <br>140 TFLOPS FP16 |', repeated in 特性 as '单卡最大提供280 TOPS INT8算力'. PRINTED IN TOPS; recorded as 280 in a TFLOPS field since one INT8 op is one operation. CARD-TOTAL, not per processor: this is a dual-310 card and '单卡' means per card, so per processor it is 140 TOPS. The vendor's own 1.86 TOPS/W ratio corroborates the reading against the published 150 W (280/150 = 1.87). No sparsity footnote and no row titled for structured sparsity, so dense:true unhalved. Source w5a-huawei-atlas-300i-duo-e-huawei-spec-table.precisionfp16tflops140denseyesvendor_claimyessource_idw5a-huawei-atlas-300i-duo-e-huawei-spec-tablebasisvendor-spec-sheetbasis_detailSame Huawei 技术规格 row: '280 TOPS INT8 <br>140 TFLOPS FP16'. Card-total like the INT8 figure, and exactly half of it, which is the ordinary FP16:INT8 ratio and confirms this is a dense pair rather than a sparse/dense pair. Dense:true, no footnote to halve.
TDP (W)
150
Unified memory
no
Consumer part
no
notes
flops WAS EMPTY ON THIS RECORD AND THAT WAS OVER-CORRECTED. It rested on cn-huawei-atlas-300i-duo-spec (support.huawei.com), whose own note correctly says 'Huawei does NOT publish an INT8/FP16 TOPS figure for the 310 series anywhere in this document - only core counts.' That is true of that document and was generalised to the vendor. Huawei's own enterprise product page publishes a complete 技术规格 table with 'AI算力 | 280 TOPS INT8 <br>140 TFLOPS FP16' for this exact card. This is the SECOND time in one pass that a negative finding scoped to a single URL was generalised to a vendor who does publish the figure - the first was Moore Threads S4000 - so the pattern is now documented twice and is worth naming as a rule: a negative must always name the URL it tested, never the vendor. THE CARD BASIS MATTERS AND IS NOT OPTIONAL: this is a DUAL-310 card and 280 TOPS INT8 / 140 TFLOPS FP16 are whole-card figures ('单卡最大', per card), with the 408 GB/s bandwidth likewise card-total ('总带宽'). At two processors that implies 140 TOPS INT8 per processor, which matches the 408/2 = 204 bandwidth split the sibling 300I Pro publishes as 204.8. Anyone reading 280 TOPS as a per-processor figure will halve the card twice. SCALE CALIBRATION, added this pass: Huawei's Atlas 300I A2 - a different, newer SKU with no record in data/accelerators - publishes exactly 2x these figures at '560 TOPS INT8，280 TFLOPS FP16', which is the first point in the 310 line where Huawei publishes a per-card INT8 figure at all. So the statement has to be per-SKU, not per-generation: the 300I Duo and the 300I A2 publish compute, while the 300I Pro (24GB) and 300V Pro (48GB) publish NO compute figure on any Huawei page located on this pass. See w5a-huawei-atlas-300i-a2-e-huawei-spec-table. OTHER VENDOR SPECS, verbatim: 功耗 150W; 形态 单槽位全高全长（10.5英寸）; 结构尺寸 266.7mm x 111.15mm x 18.46mm; PCIe x16 Lances, 兼容 x8/x4/x2, PCIe Gen4.0 兼容 3.0/2.0/1.0; 工作环境温度 0℃～55℃. Features: '支持 16 core * 1.9 GHz CPU计算能力', '提供 1.86 TOPS/W 超高能效比，业界领先', '支持256路高清视频实时分析'. Codec detail: H.264/H.265 256x1080P30 decode and 48x1080P30 encode, 'JPEG解码能力4K 1024FPS，编码能力4K 512FPS', max resolution 8192*8192. SPARSITY: no footnote, no '2:4', no sparsity wording. Huawei's convention here matches Intel's on these parts - the figure printed is the figure.

Notes

flops WAS EMPTY ON THIS RECORD AND THAT WAS OVER-CORRECTED. It rested on cn-huawei-atlas-300i-duo-spec (support.huawei.com), whose own note correctly says 'Huawei does NOT publish an INT8/FP16 TOPS figure for the 310 series anywhere in this document - only core counts.' That is true of that document and was generalised to the vendor. Huawei's own enterprise product page publishes a complete 技术规格 table with 'AI算力 | 280 TOPS INT8 <br>140 TFLOPS FP16' for this exact card. This is the SECOND time in one pass that a negative finding scoped to a single URL was generalised to a vendor who does publish the figure - the first was Moore Threads S4000 - so the pattern is now documented twice and is worth naming as a rule: a negative must always name the URL it tested, never the vendor. THE CARD BASIS MATTERS AND IS NOT OPTIONAL: this is a DUAL-310 card and 280 TOPS INT8 / 140 TFLOPS FP16 are whole-card figures ('单卡最大', per card), with the 408 GB/s bandwidth likewise card-total ('总带宽'). At two processors that implies 140 TOPS INT8 per processor, which matches the 408/2 = 204 bandwidth split the sibling 300I Pro publishes as 204.8. Anyone reading 280 TOPS as a per-processor figure will halve the card twice. SCALE CALIBRATION, added this pass: Huawei's Atlas 300I A2 - a different, newer SKU with no record in data/accelerators - publishes exactly 2x these figures at '560 TOPS INT8，280 TFLOPS FP16', which is the first point in the 310 line where Huawei publishes a per-card INT8 figure at all. So the statement has to be per-SKU, not per-generation: the 300I Duo and the 300I A2 publish compute, while the 300I Pro (24GB) and 300V Pro (48GB) publish NO compute figure on any Huawei page located on this pass. See w5a-huawei-atlas-300i-a2-e-huawei-spec-table. OTHER VENDOR SPECS, verbatim: 功耗 150W; 形态 单槽位全高全长（10.5英寸）; 结构尺寸 266.7mm x 111.15mm x 18.46mm; PCIe x16 Lances, 兼容 x8/x4/x2, PCIe Gen4.0 兼容 3.0/2.0/1.0; 工作环境温度 0℃～55℃. Features: '支持 16 core * 1.9 GHz CPU计算能力', '提供 1.86 TOPS/W 超高能效比，业界领先', '支持256路高清视频实时分析'. Codec detail: H.264/H.265 256x1080P30 decode and 48x1080P30 encode, 'JPEG解码能力4K 1024FPS，编码能力4K 512FPS', max resolution 8192*8192. SPARSITY: no footnote, no '2:4', no sparsity wording. Huawei's convention here matches Intel's on these parts - the figure printed is the figure.

Cross-references

Interconnect (1)