Skip to main content

Year-End Comparison Article - Data Collection Progress Tracker

Created: 2026-06-28 Last updated: 2026-06-28 Goal: collect complete data for the "Domestic AI Chips vs International AI Chips: 2026 Year-End Comparison" article

📊 Collected Data Summary​

1. Ascend 910C (Huawei)​

Data source: Lennart Heim analysis (republished on Zhihu)

MetricValueConfidenceNotes
FP16~800 TFLOPSMediumHeim estimate, not officially published
Memory bandwidth~3.2 TB/sMediumHeim estimate
Performance benchmark~80% of H100MediumHeim assessment
Architecturedual 910B die stackingHighconfirmed
Logic area60% larger than H100MediumHeim analysis
FP32~320 TFLOPS (derived)LowDerived from architecture, unverified

To be added:

  • Official FP32 data (or reliable third-party tests)
  • INT8 compute
  • FP8 support status
  • Actual training performance test data

2. B200/GB200 (NVIDIA)​

Data source: Zhihu technical breakdowns, EET China

MetricValueConfidenceNotes
Transistor count208 billionHighOfficial data
FP162,250 TFLOPSHighEET China report
FP418 PFLOPS (peak)HighOfficial data
Memory192 GB HBM3eHighOfficial data
Memory bandwidth8 TB/sHighOfficial data
TDP1000W (per GPU)HighHGX B200 configuration
GB200 TDP2700W (2 GPU + 1 CPU)HighOfficial data
NV-HBI10 TB/sHighDual-die interconnect bandwidth

To be added:

  • Actual FP32 performance tests
  • Actual deployment cases and performance benchmarks
  • Pricing information

3. Suisi 2.0 / Yunsui T21 (Enflame)​

Data source: Baidu Baike, Enflame official

MetricValueConfidenceNotes
FP3240 TFLOPSHighOfficial data
TF32160 TFLOPSHighOfficial data
INT8256 TOPSHighOfficial data
Memory64 GB HBM2EHighOfficial data
Memory bandwidth1.8 TB/sHighOfficial data
TDP~300W (estimated)MediumOAM module
ArchitectureGCU-CARAHighEnflame in-house
Process12nm (GlobalFoundries)HighOfficial data
InterconnectGCU-LARE, 300GB/sHighOfficial data

Product positioning: cloud training

To be added:

  • T21 exact TDP
  • Actual training performance tests
  • Next-generation product (Suisi 3.0) specs

4. Kunlun R200 (Baidu)​

Data source: Zhihu technical spec analysis

MetricValueConfidenceNotes
FP3232 TFLOPSHighOfficial data
FP16128 TFLOPSHighOfficial data
INT8256 TOPSHighOfficial data
Memory16/32 GB GDDR6HighOfficial data
Memory bandwidth512 GB/sHighOfficial data
TDP150WHighOfficial data
ArchitectureXPU-RHighBaidu in-house
InterfacePCIe Gen4 x16HighOfficial data

Product positioning: inference-focused, with some training capability

To be added:

  • R300 (next generation) specs
  • Actual performance test data

5. Tiangai 150 (BI-V150) (Iluvatar CoreX)​

Data source: SMZDM unboxing review (2026-01-08)

MetricValueConfidenceNotes
FP3248 TFLOPSMediumUnboxing review data
FP16~192 TFLOPS (derived)LowDerived from FP32 × 4
Memory64 GB HBM2eHighOfficial data
Memory bandwidth~2 TB/s (derived)LowDerived from HBM2e
TDP350WHighOfficial data
Process7nmHighOfficial data
Architectureivcore11HighIluvatar CoreX in-house
InterfacePCIe 4.0 x16HighOfficial data

Product positioning: general-purpose compute GPU (training + inference)

To be added:

  • Official FP16 data (currently a derived value)
  • Official INT8 data
  • Official memory bandwidth data
  • Actual performance test data

Update log (2026-06-28):

  • ✅ Added FP32 = 48 TFLOPS (source: SMZDM unboxing review)
  • ⚠️ FP16 is a derived value and needs official confirmation

6. Hanbo SV100 series (Hanbo Semiconductor)​

Data source: Zhihu article (published 2022)

MetricValueConfidenceNotes
INT8>200 TOPSMediumOfficial data (2022)
Supported precisionsFP16, BF16, INT8HighOfficial data
TDP75W (VA1 accelerator card)HighOfficial data
Memory32 GBHighVA1 accelerator card
Video decoding64+ channels of 1080pHighOfficial data
Architecturegeneral-purpose DSAHighHanbo in-house

Product positioning: cloud AI inference

To be added:

  • SG100 (GPU) specs
  • Latest product specs (2026)
  • Actual performance test data

7. 2026 Q1 market data​

Data source: Songguo Market Insights (Baidu Baijiahao)

MetricValueYoY Change
Domestic AI chip market share52.3%+32.3pp (from 20% to 52.3%)
NVIDIA market share42.7%-52.3pp (from 95% to 42.7%)
MilestoneFirst time above 50%A historic first

Key insights:

  • ✅ Policy-driven procurement dominates (localization rate >70% in strategic sectors such as government, finance, and energy)
  • ✅ Ecosystem migration begins: DeepSeek V4 completed its migration from CUDA to CANN
  • ⚠️ After the policy dividend fades in 2027-2028, market-based competition will be the real test

To be added:

  • 2026 Q2 data (available in July)
  • Specific shipment data by vendor
  • Market share broken down by use case (training vs inference)

8. TsingMicro (reconfigurable AI chip RPU)​

Data source: Sohu news (2025-07-27)

MetricValueConfidenceNotes
Architecturereconfigurable (RPU)HighTsinghua technology transfer
ProductTX81 (cloud)MediumMass production in 2025
Orders~20,000 cardsMediumReported figure
Shipments>20 million units (cumulative, including edge)MediumReported figure
Key featuredynamic reconfiguration, adapts to different tasksHighOfficial claim

Product positioning: differentiated route (non-GPU architecture)

To be added:

  • TX81 detailed compute data
  • Cloud product specs
  • Actual performance test data

🔍 Data To Be Collected​

High priority (within this week)​

Domestic chips​

  • Tiangai 150 (BI-V150): FP32/FP16/INT8 compute, memory bandwidth ← urgently needed
  • Enflame Yunsui T21: full TDP power figure
  • Kunlun R300: full specs
  • Hanbo SG100: full specs (GPU product)
  • TsingMicro TX81: detailed compute data

International chips​

  • MI300X/MI325X: actual performance test data, pricing
  • TPU v6 (Trillium): spec data
  • L40S/L4: inference performance data

Market data​

  • H1 2026 shipment data (IDC, CCID)
  • Revenue data by vendor (Huawei, Cambricon, Enflame, etc.)
  • Major customer procurement announcements (ByteDance, Alibaba, Baidu, etc.)

Medium priority (within July)​

  • Latest versions and features of each chip's software stack
  • Actual model migration cases and timelines
  • Developer community activity data

Low priority (August-September)​

  • Detailed energy-efficiency tests of each chip
  • Actual deployment case studies
  • User satisfaction and feedback

📅 Data Collection Schedule​

Week 1-2 (2026-06-28 ~ 2026-07-11)​

Goal: complete high-priority chip spec collection

  • Create the data collection tracker file
  • Collect Ascend 910C specs (FP16 ~800 TFLOPS)
  • Collect full B200/GB200 specs
  • Collect 2026 Q1 market share data
  • Collect Suisi 2.0/T21 specs
  • Collect Kunlun R200 specs
  • Collect Tiangai 150 detailed specs ← in progress
  • Update the Ascend 910C page (FP16 updated to 800 TFLOPS)

Week 3-4 (2026-07-12 ~ 2026-07-25)​

Goal: complete market data and ecosystem comparison

  • Search for H1 2026 shipment data
  • Collect software stack and ecosystem comparison material
  • Create draft comparison tables for each chip

Week 5-8 (2026-07-26 ~ 2026-08-22)​

Goal: fill in missing data, start article writing

  • Collect major customer procurement cases
  • Collect actual deployment performance test data
  • Follow vendor product launches and technical announcements
  • Complete the draft of Chapter 1 (Introduction)

🔗 Data Source Records​

Sources used​

  1. Lennart Heim's Ascend 910C analysis - republished on Zhihu

  2. NVIDIA B200 key technology breakdown - Zhihu

  3. Domestic AI chip share tops 50% for the first time - Baidu Baijiahao

  4. Suisi 2.0 Baidu Baike entry - Baidu Baike

  5. Kunlun R200 technical spec analysis - Zhihu

  6. Tiangai 150 product documentation - Moark

Data pending verification​

  • ⚠️ Ascend 910C FP16 ~800 TFLOPS: more sources needed for verification
  • ⚠️ Domestic chip market share 52.3%: needs verification against IDC or MIIT official data
  • ⚠️ B200 FP16 2,250 TFLOPS: needs verification against an official whitepaper or reliable reviews
  • ❓ Tiangai 150 FP32/FP16/INT8 compute: not officially published, urgently needed

📝 Data Update Log​

2026-06-28​

New data:

  • ✅ Ascend 910C: FP16 updated to ~800 TFLOPS (previously miswritten as 376 TFLOPS)
  • ✅ Ascend 910C: memory bandwidth updated to ~3.2 TB/s
  • ✅ B200: full specs (transistors, FP16, memory, bandwidth, TDP)
  • ✅ 2026 Q1 market share data: domestic 52.3%, NVIDIA 42.7%
  • ✅ Suisi 2.0/T21: full specs (FP32 40 TFLOPS, TF32 160 TFLOPS, etc.)
  • ✅ Kunlun R200: full specs (FP32 32 TFLOPS, FP16 128 TFLOPS, etc.)
  • ✅ Tiangai 150: basic info (64GB HBM2e, 350W TDP, 7nm)
  • ✅ TsingMicro: company info and product positioning

Pending verification:

  • ❓ Ascend 910C FP32 compute: derived at ~320 TFLOPS, needs confirmation
  • ❓ Tiangai 150 FP32/FP16/INT8 compute: not officially published, urgently needed
  • ❓ Enflame T21 TDP: no official data found

Next steps:

  • Continue collecting specs for domestic chips such as Tiangai 150 and Hanbo SG100
  • Verify the accuracy of collected data
  • Update existing chip pages on the site (e.g. Ascend 910C FP16 data)

📊 Data Completeness Assessment​

Domestic chips (needed for the comparison table)​

ChipFP32FP16INT8MemoryBandwidthTDPCompleteness
Ascend 910C❓ Derived✅ ~800❌✅ 64GB✅ ~3.2✅ 310W60%
MLU690✅ 150✅ 600✅ 1200✅ 64GB❓ ~2TB✅ 280W80%
Tiangai 150❌❌❌✅ 64GB❌✅ 350W30%
Enflame T21✅ 40❓ Estimated ~160✅ 256✅ 64GB✅ 1.8TB❓ ~300W70%
Kunlun R200✅ 32✅ 128✅ 256✅ 16/32GB✅ 512GB✅ 150W90%

International chips (needed for the comparison table)​

ChipFP32FP16INT8MemoryBandwidthTDPCompleteness
H100✅ 60✅ 989✅ 1979✅ 80GB✅ 3.35TB✅ 700W100%
H200✅ 60✅ 989✅ 1979✅ 141GB✅ 4.8TB✅ 700W100%
B200✅ 80✅ 2250✅ 4500✅ 192GB✅ 8TB✅ 1000W100%
MI300X✅ 163✅ 1271✅ 2542✅ 192GB❓ ~5.3TB✅ 750W90%

Summary:

  • ✅ International chip data is highly complete (90-100%)
  • ⚠️ Domestic chip data is moderately complete (30-90%), with Tiangai 150 the most lacking
  • 🔴 Urgently needed: Tiangai 150 compute data, official Ascend 910C FP32 data

Notes:

  • Confidence levels: High = official data or reliable reviews, Medium = cross-validated by multiple sources, Low = single source or derived
  • All data must cite its source and confidence level before use
  • All key data must be re-verified before the article is published