Skip to main content

NVIDIA H800 (2023)

Product Overview​

NVIDIA H800 is the first export-compliant AI accelerator NVIDIA launched for the China market, based on the Hopper architecture, officially released in early 2023.

Restricted by US export controls (the October 2022 version), H800's interconnect bandwidth was cut to comply:

  • FP16: 1,979 TFLOPS (sparse, same as the H100)
  • FP8: 3,958 TFLOPS (sparse, same as the H100)
  • Memory: 80GB HBM3 (3.35 TB/s bandwidth, same as the H100)
  • Interconnect: NVLink cut (H100 900 GB/s → H800 ~400-600 GB/s)

H800's positioning: compute unchanged, interconnect bandwidth cut — staying as competitive as possible within compliance.

Core Specifications​

ParameterValue
ArchitectureHopper GH100 (export-compliance cut-down)
ProcessTSMC 4N (4nm)
FP83,958 TFLOPS (sparse)
FP161,979 TFLOPS (sparse)
TF32989 TFLOPS (sparse)
FP3267 TFLOPS
INT83,958 TOPS (sparse)
Memory Capacity80GB HBM3
Memory Bandwidth3.35 TB/s
InterconnectNVLink (downgraded, ~400-600 GB/s)
TDP350W (SXM5)
ReleaseEarly 2023
DiscontinuedOctober 2023 (new US export control rules)

Export Compliance Background​

DateEvent
2022-10US BIS issues initial export controls restricting A100/H100 exports to China
2023-03NVIDIA launches H800 (NVLink bandwidth cut for compliance)
2023-10US issues stricter new export control rules
2023-10H800 also falls under the controls, shipments to China halted

H800 vs H100 Compliance-Cut Comparison​

ParameterH100H800Cut
FP16 compute1,979 TFLOPS1,979 TFLOPSNone
FP8 compute3,958 TFLOPS3,958 TFLOPSNone
Memory80GB HBM380GB HBM3None
Memory bandwidth3.35 TB/s3.35 TB/sNone
NVLink bandwidth900 GB/s~400-600 GB/sCut ~33-55%
TDP700W350WReduced 50%

Strategy: NVIDIA chose to keep compute and memory (important for single-card inference) while cutting interconnect bandwidth and power (impacting large-scale training clusters).

H800 vs H20​

MetricH800H20Difference
FP161,979 TFLOPS148 TFLOPSH800 13.4x
FP83,958 TFLOPS296 TFLOPSH800 13.4x
Memory80GB HBM396GB HBM3H20 +20%
Bandwidth3.35 TB/s4.0 TB/sH20 +19%
NVLink~400-600 GB/s (downgraded)900 GB/s (downgraded)Similar
TDP350W400WH20 +14%
Lifecycle2023.03 - 2023.10 (7 months only)2024.02 - 2025.07H20 longer

Conclusion: H800 compute far exceeds H20, but its lifecycle was extremely short (only 7 months). H20 is NVIDIA's second-generation compliant product after the October 2023 rules.

Comparison with Domestic Chips​

H800 vs Hygon DCU K100​

MetricNVIDIA H800Hygon DCU K100Difference
FP161,979 TFLOPS192 TFLOPSH800 10.3x
Memory80GB HBM364GB HBM3H800 +25%
Bandwidth3.35 TB/s3.2 TB/sH800 +5%
EcosystemCUDA (complete)ROCm (compatible)H800 advantage
SupplyDiscontinued October 2023Stable supplyK100 advantage

Note: H800 is discontinued, so this comparison is largely moot. The current battleground for domestic substitution is H20 vs domestic chips.

Use Cases (historical review)​

  • ✅ Short-term procurement in 2023 (no longer procurable)
  • ✅ Single-card inference (compute uncut, 80GB large memory)
  • ❌ Large-scale training clusters (NVLink bandwidth cut, multi-card interconnect performance drops)
  • ❌ Procurement after 2024 (discontinued, no supply)

Discontinuation Impact (October 2023)​

In October 2023, the US issued stricter new export control rules, and H800 also fell under the controls:

Impact on China's AI industry:

  • Customers who bought H800: no short-term impact, but no expansion possible
  • Customers who did not: forced to move to H20 (heavily cut compute) or domestic chips
  • Domestic substitution timeline: brought forward 1-2 years

Vendor Information​

ParameterValue
CompanyNVIDIA Corporation
ArchitectureHopper (first-generation export-compliance cut-down)
ReleaseEarly 2023
DiscontinuedOctober 2023 (new US export control rules)
China market positioningFirst compliant replacement for the H100
LifecycleOnly ~7 months (shortest ever)