NVIDIA Datacenter GPU
Lineage Reference
A consolidated reference for NVIDIA-branded datacenter and compute GPUs from the announced Rubin roadmap back to the Tesla brand era, with ITAD-specific valuation guidance per generation. Built for ITAD operators, secondary market buyers, and procurement teams who need to identify SKUs, verify form-factor compatibility, and flag export-controlled variants in disposition streams.
GPU Resource Configuration Guide
OEM part-number mapping for Dell, HPE, Lenovo, and DGX platforms. Server-first lookup and PN reverse-search.
Open Guide →Every NVIDIA datacenter GPU, with the disposition pitfalls flagged
The page below catalogs NVIDIA-branded datacenter and compute GPUs by architecture generation, ordered newest first: announced Rubin and Feynman roadmap, then Blackwell, Hopper, Ada Lovelace, Ampere, Turing, Volta, Pascal, Maxwell, Kepler, and the Tesla / Fermi early generation. Each section lists the shipped SKUs with year, memory configuration, form factor, and TDP.
What sets this reference apart from a public spec list is the ITAD commentary attached to each generation. Form-factor identification (SXM vs PCIe), HGX baseboard compatibility, China-export-control SKU variants, and firmware-entitlement constraints are the four most common sources of secondary-market pricing error — they are flagged inline at the generation where they apply, and consolidated in the matrices at the bottom of the page.
Out of scope: GeForce consumer cards, Quadro workstation cards, and OEM-only variants are not catalogued here unless they are routinely encountered in enterprise ITAD streams.
Jump to a generation
The “Tesla” name has two unrelated meanings — both appear on this page
The Tesla GPU product brand was retired in May 2020. Datacenter products before that date carried “Tesla” branding regardless of the underlying architecture (e.g., Tesla K80 is Kepler, Tesla V100 is Volta). After May 2020, NVIDIA names datacenter products by architecture letter — A-series for Ampere, L-series for Ada Lovelace, H-series for Hopper, B-series for Blackwell.
Separately, NVIDIA’s first unified-architecture GPU family from 2007 is also called “Tesla” — that is a microarchitecture name, used internally for the original CUDA generation. Neither has any connection to the EV company.
Announced, not yet shipping
For forward-planning purposes only — specifications are provisional until shipping silicon is available.
| Architecture | GPU | Expected | Memory | Notes |
|---|---|---|---|---|
| Rubin | R100 | H2 2026 | HBM4 (8 stacks) | Pairs with Vera ARM CPU |
| Rubin Ultra | GR200 (est.) | 2027 | HBM4 (12 stacks) | — |
| Feynman | TBD | 2028 | TBD | Announced GTC March 2025 |
Blackwell · B-series (2024–present)
Blackwell ships as discrete B100 / B200 / B200A cards and as Grace Blackwell GB200 superchips integrated into NVL72 rack-scale liquid-cooled systems. SXM6 is a new socket and is not backward compatible with HGX H100 / SXM5 servers.
| Model | Year | Memory | Form factor | TDP | Notes |
|---|---|---|---|---|---|
| B100 | 2024 | 192 GB HBM3e | SXM6 | ~700W | 8× HBM stacks |
| B200 | 2024 | 192 GB HBM3e | SXM6 | ~1,000W | Flagship Blackwell |
| B200A | 2025 | 192 GB HBM3e | PCIe | TBD | Air-cooled PCIe variant |
| GB200 Superchip | 2024 | 192 GB HBM3e + 480 GB LPDDR5 | SXM | ~2,700W* | Grace CPU + B200 GPU |
| GB200 NVL72 | 2024 | 72× B200 (13.8 TB total) | Rack system | Liquid | 36 Grace Blackwell chips per rack |
| GB300 NVL72 (Blackwell Ultra) | 2025 | 192 GB HBM3e per GPU | Rack system | Liquid | ~50% more FP4 perf vs GB200 NVL72 |
*Per-rack figure for NVL72 system
ITAD note. Blackwell is effectively sold out in new channels through mid-2026; meaningful secondary-market volume is not expected before 2026–2027.
GB200 NVL72 is a rack-scale liquid-cooled system, not a discrete GPU card. Disposition requires specialized liquid-cooling logistics and whole-rack handling rather than card-level treatment. SXM6 is a new socket — it is not interchangeable with HGX H100 / SXM5 servers. B200A (PCIe) targets air-cooled deployments and will be more accessible to standard rack environments.
Hopper · H-series (2022–present)
Hopper is the dominant training and high-end inference architecture of the current cycle. H100 SXM5 cards plug into HGX H100 baseboards; the same baseboard accepts H200 SXM5 cards, which carry the same GH100 die but roughly double the memory capacity.
| Model | Year | Memory | Form factor | TDP | Notes |
|---|---|---|---|---|---|
| H100 SXM5 80GB | 2022 | 80 GB HBM3 | SXM5 | 700W | Flagship training; NVLink 4.0 (900 GB/s) |
| H100 PCIe 80GB | 2022 | 80 GB HBM2e | PCIe | 350W | Lower bandwidth; no NVSwitch |
| H100 NVL 94GB | 2023 | 94 GB HBM3 | PCIe | 400W | NVLink pair config; 188 GB combined |
| GH200 Grace Hopper | 2023 | 96 GB HBM3 + 480 GB LPDDR5 | SXM | 900W | Grace CPU + H100 superchip |
| H800 SXM 80GB | 2023 | 80 GB HBM3 | SXM5 | 700W | China export-compliantCHINA-ONLY |
| H800 PCIe 80GB | 2023 | 80 GB HBM2e | PCIe | 350W | China export-compliantCHINA-ONLY |
| H20 | 2024 | 96 GB HBM3 | PCIe | 200W | China export-compliant; reduced computeCHINA-ONLY |
| H200 SXM 141GB | 2024 | 141 GB HBM3e | SXM5 | 700W | Same GH100 die as H100; ~2× memory |
| H200 NVL 141GB | 2024 | 141 GB HBM3e | PCIe | 400W | NVLink pair; 282 GB combined |
ITAD note — form factor. H100 SXM5 vs PCIe is the single largest valuation split in current secondary-market activity. SXM5 requires an HGX H100 baseboard and commands a 30–50% premium over the PCIe variant. Misidentifying form factor at intake is a significant pricing-error risk. H200 SXM5 is board-level compatible with HGX H100 baseboards.
H800 and H20 are China export-control SKUs under BIS Entity List rules. Re-export to restricted parties or jurisdictions may be unlawful — flag for compliance review before disposition.
Firmware entitlement caveat
nvfwupd and NVIDIA firmware packages for H100 and H200 are gated behind NVIDIA Enterprise Support Portal entitlements that do not transfer with secondary-market hardware. Units acquired through ITAD channels may have no supported firmware upgrade path. This is a material disclosure for any disposition where the buyer expects ongoing firmware maintenance, and it is a topic that secondary market documentation rarely surfaces.
Ada Lovelace · L-series (2022–present)
Ada Lovelace is the air-cooled enterprise AI line — GDDR6 rather than HBM, PCIe rather than SXM. L40S has effectively replaced A40 as the standard enterprise AI card in air-cooled deployments. L4 is the spiritual successor to T4 — single-slot, low-watt, designed for broad cloud and enterprise inference deployment.
| Model | Year | Memory | Form factor | TDP | Notes |
|---|---|---|---|---|---|
| L4 | 2023 | 24 GB GDDR6 | PCIe | 72W | T4 successor; Google Cloud G2 instances |
| L40 | 2022 | 48 GB GDDR6 | PCIe | 300W | Rendering / inference |
| L40S | 2023 | 48 GB GDDR6 | PCIe | 350W | AI + graphics; FP8 Tensor Cores; A40 successor |
| L20 | 2023 | 48 GB GDDR6 | PCIe | 200W | China export-compliant L40SCHINA-ONLY |
| L2 | 2023 | 24 GB GDDR6 | PCIe | 60W | China export-compliant L4CHINA-ONLY |
ITAD note. L40S is the current air-cooled enterprise AI workhorse and is positioned as the A40 successor. Secondary-market volume is still emerging; expect meaningful supply as hyperscaler refresh cycles reach 2026–2027.
L4’s 72W single-slot envelope removes most deployment barriers, which is why it shows up across so many cloud and enterprise refresh streams. L20 and L2 are China-market-only — flag them for re-export compliance review at intake.
Ampere · A-series (2020–2023)
Ampere introduced the current architecture-letter naming convention and the HGX baseboard model that carries through to Hopper and Blackwell. A100 SXM4 cards plug into HGX A100 baseboards (4-GPU or 8-GPU variants) and are not interchangeable with PCIe servers.
| Model | Year | Memory | Form factor | TDP | Notes |
|---|---|---|---|---|---|
| A10 | 2021 | 24 GB GDDR6 | PCIe | 150W | Inference / graphics; single-slot |
| A16 | 2021 | 4× 16 GB GDDR6 | PCIe | 250W | VDI / vGPU; quad-GPU single card |
| A30 | 2021 | 24 GB HBM2e | PCIe | 165W | Mid-range HPC/AI |
| A40 | 2020 | 48 GB GDDR6 | PCIe | 300W | Large-VRAM inference / rendering |
| A100 40GB SXM4 | 2020 | 40 GB HBM2 | SXM4 | 400W | Flagship training |
| A100 80GB SXM4 | 2020 | 80 GB HBM2e | SXM4 | 400W | Extended-memory training |
| A100 40GB PCIe | 2020 | 40 GB HBM2 | PCIe | 300W | — |
| A100 80GB PCIe | 2021 | 80 GB HBM2e | PCIe | 300W | — |
| A800 80GB SXM | 2022 | 80 GB HBM2e | SXM4 | 400W | China export-compliant A100CHINA-ONLY |
| A800 40GB PCIe | 2022 | 40 GB HBM2 | PCIe | 300W | China export-compliantCHINA-ONLY |
ITAD note. A100 SXM4 is not a drop-in PCIe card — it requires an HGX-compatible baseboard, and the 4-GPU and 8-GPU HGX A100 boards are different products. Verify the donor server before pricing.
A800 is the China export-control variant of A100 with reduced NVLink bandwidth (400 GB/s vs A100’s 600 GB/s). Treat as a distinct SKU for pricing and confirm end-destination compliance — re-export to restricted parties or jurisdictions may be unlawful under BIS rules.
A40 is the large-VRAM enterprise workhorse for the generation: PCIe form factor, no HGX requirement, 48 GB on a single card. A10 is the high-volume inference card that appears across Dell and HPE mid-cycle refreshes.
Turing · T4 (2018–2022)
Turing in the consumer market is the RTX 20-series. In the datacenter line, only the T4 shipped — but it became one of the most widely deployed datacenter GPUs ever produced, on the strength of its 70W single-slot envelope.
| Model | Year | Memory | Form factor | TDP | Notes |
|---|---|---|---|---|---|
| Tesla T4 | 2018 | 16 GB GDDR6 | PCIe | 70W | Dominant cloud inference GPU |
ITAD note. T4 is among the highest-volume datacenter GPUs in the secondary market — AWS G4dn and Google Cloud T4 instances drove enormous deployment. OEM-rebranded variants (Dell, HPE, Lenovo) are ubiquitous in enterprise refresh streams. Per-unit value is modest, but the SKU moves quickly.
Volta · V100 (2017–2020)
V100 is the first generation where NVIDIA explicitly named datacenter products by architecture rather than product brand. SXM2, SXM3, and PCIe variants shipped over a four-year span at 16 GB and 32 GB capacities.
| Model | Year | Memory | Form factor | TDP | Notes |
|---|---|---|---|---|---|
| Tesla V100 SXM2 16GB | 2017 | 16 GB HBM2 | SXM2 | 300W | DGX-1 GPU; NVLink 2.0 |
| Tesla V100 PCIe 16GB | 2017 | 16 GB HBM2 | PCIe | 250W | — |
| Tesla V100 SXM2 32GB | 2018 | 32 GB HBM2 | SXM2 | 300W | — |
| Tesla V100 PCIe 32GB | 2018 | 32 GB HBM2 | PCIe | 250W | — |
| Tesla V100S PCIe 32GB | 2019 | 32 GB HBM2 | PCIe | 250W | Faster memory variant |
| Tesla V100 SXM3 32GB | 2018 | 32 GB HBM2 | SXM3 | 350W | DGX-2 only — not interchangeable |
ITAD note. The SXM2 vs SXM3 distinction is a critical valuation split. SXM3 is DGX-2 exclusive and has effectively no aftermarket beyond that single platform, despite premium specifications. PCIe variants are the most liquid form factor across all V100 capacities. The 32 GB capacity commands a meaningful premium over 16 GB for memory-bound training workloads.
Pascal (2016–2018)
Pascal introduced HBM2 memory and the SXM2 form factor to the datacenter line. The P100 was the first NVLink-capable datacenter GPU; the P40 has had unusual longevity in the inference market.
| Model | Year | Memory | Form factor | TDP | Notes |
|---|---|---|---|---|---|
| Tesla P4 | 2016 | 8 GB GDDR5 | PCIe | 75W | Low-power inference |
| Tesla P6 | 2016 | 16 GB GDDR5 | MXM | 90W | Server blade |
| Tesla P40 | 2016 | 24 GB GDDR5 | PCIe | 250W | Training / large-batch inference |
| Tesla P100 SXM2 | 2016 | 16 GB HBM2 | SXM2 | 300W | First HBM datacenter GPU; NVLink |
| Tesla P100 PCIe | 2016 | 16 GB HBM2 | PCIe | 250W | PCIe variant; no NVLink |
ITAD note. P40 retains durable inference demand — 24 GB GDDR5 at low secondary cost is an attractive combination for AI hobbyists and cost-sensitive inference deployments. P100 SXM2 requires an NVLink-capable server (DGX-1, HGX-1) and should not be priced as a drop-in PCIe card; verify the donor server before quoting. P4 is high-volume but low-value per unit.
Maxwell (2014–2016)
Maxwell shifted toward inference, virtual desktop infrastructure, and power-efficient deployment. The M60 was widely deployed in Citrix and VMware vGPU environments.
| Model | Year | Memory | Notes |
|---|---|---|---|
| Tesla M4 | 2015 | 4 GB GDDR5 | Low-power inference |
| Tesla M10 | 2015 | 4× 8 GB GDDR5 | VDI / virtual desktop |
| Tesla M40 | 2015 | 12 or 24 GB GDDR5 | Training |
| Tesla M60 | 2015 | 2× 8 GB GDDR5 | VDI / vGPU (Citrix, VMware) |
Kepler (2012–2016)
Kepler was the dominant compute architecture of the early HPC and early cloud era. The K80 in particular powered AWS P2 instances and a large fraction of first-wave cloud GPU deployments.
| Model | Year | Memory | Notes |
|---|---|---|---|
| Tesla K10 | 2012 | 2× 4 GB GDDR5 | Dual-GPU inference card |
| Tesla K20 | 2012 | 5 GB GDDR5 | Scientific HPC |
| Tesla K20X | 2012 | 6 GB GDDR5 | Titan supercomputer GPU |
| Tesla K40 | 2013 | 12 GB GDDR5 | HPC workhorse |
| Tesla K80 | 2014 | 2× 12 GB GDDR5 | Dual-GPU; dominant early cloud GPU |
ITAD note. K80 inventory remains visible in HPC refresh streams — the install base was large and many units are still being decommissioned. K40 and K80 are end-of-driver-support, so retained value is workload-specific (CUDA-compatible legacy applications and budget HPC users) rather than broad market.
Tesla microarchitecture & Fermi (2007–2013)
The earliest NVIDIA datacenter compute parts. CUDA was introduced on the original Tesla microarchitecture; ECC memory, double-precision performance, and the SXM/server-module concept arrived with Fermi. These cards are largely retired from production deployment but still surface occasionally in legacy refresh streams.
| Model | Year | Memory | Notes |
|---|---|---|---|
| Tesla C870 | 2007 | 1.5 GB GDDR3 | First CUDA compute card |
| Tesla S1070 | 2008 | 4× 4 GB GDDR3 | 1U server appliance |
| Tesla C1060 | 2008 | 4 GB GDDR3 | Workstation compute |
| Tesla S2050 | 2010 | 4× 3 GB GDDR5 | Fermi server appliance |
| Tesla C2050 | 2010 | 3 GB GDDR5 ECC | Fermi workstation |
| Tesla C2070 | 2010 | 6 GB GDDR5 ECC | Fermi workstation |
| Tesla M2050 | 2010 | 3 GB GDDR5 ECC | Fermi server module |
| Tesla M2070 | 2010 | 6 GB GDDR5 ECC | Fermi server module |
| Tesla M2090 | 2011 | 6 GB GDDR5 ECC | Higher-end Fermi server |
Form factor compatibility
SXM sockets are not cross-compatible across generations. Form-factor identification at intake determines which servers a card is usable in, and is one of the primary sources of valuation error in secondary-market disposition.
| Form factor | Architecture | Server requirement |
|---|---|---|
| PCIe x16 | All generations | Any standard PCIe server |
| SXM2 | Pascal, Volta | DGX-1 / HGX-1 baseboard |
| SXM3 | Volta (DGX-2 only) | DGX-2 exclusive — not interchangeable |
| SXM4 | Ampere (A100) | HGX A100 baseboard (4-GPU or 8-GPU) |
| SXM5 | Hopper (H100, H200) | HGX H100 or H200 baseboard |
| SXM6 | Blackwell (B100, B200) | HGX B200 baseboard — not SXM5 compatible |
| MXM | Pascal (P6) | Blade-server specific |
Export-control SKU map
BIS export control rules have produced restricted variants of multiple GPU generations. These are distinct SKUs with different specifications and potential re-export restrictions; do not treat as equivalent to the base SKU for pricing.
| Export SKU | Based on | Primary restriction | Market |
|---|---|---|---|
| A800 | A100 | NVLink BW: 400 GB/s (vs 600 GB/s) | China only |
| H800 | H100 | Reduced NVLink bandwidth | China only |
| H20 | Hopper | Significantly reduced compute | China only |
| L20 | L40S | Reduced interconnect performance | China only |
| L2 | L4 | Reduced interconnect performance | China only |
Compliance action at intake
Flag any of the above SKUs during intake. Verify end-destination compliance before disposition. Re-export to restricted parties or countries may be unlawful under BIS Entity List rules.
Need a SKU-level valuation or compatibility check?
GPU Resource provides current and forward fair-market value, OEM part-number cross-references, and disposition advisory across the SKUs catalogued on this page. If you have a specific intake to value or a configuration to verify, get in touch.