Platforms
H100 NVL vs L40S: What Changes in the Server Option?
H100 NVL and L40S can appear as separate accelerator options in the same server family, but the options are not interchangeable by GPU count. NVIDIA lists different memory types and capacities, and explicitly lists MIG and NVLink support for H100 NVL but not L40S. Compare those capabilities with the application requirement, then confirm the complete OEM configuration. Neither the board specifications nor a shared chassis name establish workload performance or site suitability.
Name H100 NVL before comparing it with L40S
An equipment request that says only “H100 or L40S” leaves a material product distinction unresolved. NVIDIA’s current H100 specification table separates the SXM and NVL variants. H100 SXM is not the PCIe-board alternative used in this comparison. The NVL column identifies a dual-slot, air-cooled PCIe form factor and 94GB of memory per GPU; the SXM column lists 80GB.
NVIDIA also describes H100 NVL with a bridge and 188GB across a pair. Keep the per-GPU and paired descriptions separate. A quotation for one card, two cards or a complete server should identify which unit each quantity describes. An advertised pair-level figure does not establish a single memory pool available to an application, nor prove that a particular server includes the bridge.
L40S is another specific product name. Do not silently substitute L40, an RTX workstation card or a newer server GPU when checking a returned option. This article compares documented product characteristics; it does not establish current inventory, orderability or continued support for an actual supplier quotation.
Compare the hardware flags that change the shortlist
Memory is one difference, but two yes-or-no capabilities can change the question before any speed comparison. Multi-Instance GPU, or MIG, is a named NVIDIA capability for dividing supported GPUs into GPU instances. NVLink is a GPU interconnect capability. A request for one does not establish a requirement for the other.
Use the exact product columns below to screen a requirement. They do not specify the instance profiles, virtualization stack or interconnect arrangement that a complete server will deliver. NVIDIA’s MIG supported-products table separately includes H100 PCIe variants; it is not a statement that MIG has been configured in a purchased system.
Scroll the table horizontally on smaller screens.
| Question | H100 NVL | L40S | Buyer consequence | Source |
|---|---|---|---|---|
| Advertised device memory | 94GB per GPU; HBM3 in NVIDIA’s NVL description | 48GB GDDR6 with ECC per GPU | Preserve memory type, unit and variant; neither number is an application-capacity result. | [N1][N2][N4] |
| MIG capability | Supported in NVIDIA’s product specifications | Not supported in NVIDIA’s specifications | An explicit MIG requirement cannot be met by changing the L40S quantity alone. | [N1][N2][N3] |
| NVLink capability | Listed for H100 NVL | Not supported in NVIDIA’s specifications | Identify the actual supported connection arrangement; a server name or pair of cards does not prove it. | [N1][N2] |
| Board interface and form factor | PCIe Gen5 card; dual-slot, passive in the product brief | PCIe Gen4 x16; dual-slot, passive in the L40S specification | These board descriptions do not replace the selected server’s configuration and installation documents. | [N1][N2][N4] |
Primary specifications checked October 16, 2026. Buyer consequences are original interpretation of the named fields. No peak FLOPS, memory-bandwidth ranking, power budget, model-fit calculation or cross-system performance claim is derived.
Use one server family to expose the difference
Dell’s PowerEdge R770 Installation and Service Manual lists up to two H100 NVL 94GB double-width GPUs and, separately, up to two L40S 48GB double-width GPUs. This is a useful comparison because the server-family name and the advertised maximum GPU count can match while the accelerator capabilities differ.
Read this as two entries in the R770 GPU specifications, not a configurable order or permission to exchange boards. The entry alone does not identify the installed risers, supported component combination, associated parts, complete chassis configuration or an NVLink bridge arrangement. It also does not say the two GPU models may be mixed. Ask the supplier to identify the exact supported server configuration represented by each returned option.
Keep the system document attached to the decision. A GPU product sheet describes the accelerator; an OEM server configuration identifies the system being supplied. Using an H100 SXM specification or another PowerEdge model’s GPU matrix to fill a missing R770 answer would change the comparison.
References: [D1]
Compare the complete physical system, including what stays the same
The physical comparison is more specific than “H100 needs a different rack.” In the named PowerEdge R770 example, Dell identifies a 2U server and separately lists up to two double-width H100 NVL GPUs or up to two double-width L40S GPUs, with 350 W against each GPU entry. Those common OEM fields do not disappear because the memory and feature flags differ. They also do not identify a complete, supported order. These physical fields were rechecked on October 19, 2026.
At board level, the retained comparison describes two passive, dual-slot PCIe products with different PCIe generations. NVIDIA's L40S table gives a 4.4-inch height and 10.5-inch length. The H100 NVL product brief describes a full-height, full-length 10.5-inch dual-slot board. A similar nominal board envelope does not establish identical retention, auxiliary cabling, riser population or thermal qualification in every server. Passive describes the card's cooling arrangement; it does not mean the installed server can dispense with its prescribed cooling.
There is also a physical interconnect distinction. The March 2024 H100 NVL brief describes a bridged pair using three NVLink bridges and identifies additional space around the card for that assembly. L40S has no NVLink support in NVIDIA's table. A bridged H100 NVL assembly therefore has parts and an envelope that an L40S comparison cannot simply inherit. This is a description of NVIDIA's named assembly, not an instruction to add bridges to an R770. Dell's two-GPU listing alone does not establish that arrangement. No bridge placement, clearance value, CPU topology or installation step is adopted here.
Power figures also have a source boundary. The H100 NVL board brief includes a 400 W maximum in a specified board-power context, while the named Dell GPU entry is 350 W. Preserve the OEM-specific entry when describing that server option; do not overwrite it with the board's headline maximum or conclude that an installer should change a power setting. Neither figure is the complete server's input demand, rack capacity or a branch-circuit allowance.
Scroll the table horizontally on smaller screens.
| Boundary | H100 NVL option | L40S option | Physical planning consequence | Source |
|---|---|---|---|---|
| Complete server family | R770 is a 2U server; Dell lists up to two double-width H100 NVL GPUs. | The same R770 family lists up to two double-width L40S GPUs. | The GPU name alone does not require a different chassis height. Select the actual complete configuration before assessing rack space, rails and access. | [D1][D2] |
| Board package and host interface | Full-height, full-length 10.5-inch dual-slot PCIe Gen5 board in the H100 NVL brief. | 4.4-inch by 10.5-inch dual-slot PCIe Gen4 x16 board in NVIDIA's L40S table. | Similar nominal size does not approve a swap or determine the actual server's retention and connection kit. | [N2][N5] |
| GPU-to-GPU physical assembly | NVIDIA describes a pair with three NVLink bridges and additional assembly space. | NVIDIA lists no NVLink support. | The H100 NVL bridge arrangement is a distinct supplied-assembly question. It is not established by a two-card OEM option entry. | [N2][N5] |
| GPU power field in the named OEM listing | 350 W against the R770 H100 NVL entry. | 350 W against the R770 L40S entry. | These listed GPU fields can match while the products differ. They are not measured or maximum whole-server demand. | [D1] |
| Board thermal description | Passive in the H100 NVL product brief. | Passive in NVIDIA's L40S specification. | The actual complete-server cooling configuration and its operating requirements still govern; passive is not a room-air or fan allowance. | [N2][N5] |
A source-qualified comparison of selected published fields, not a compatibility matrix, rack design, installation method, power calculation or approval of an actual configuration. The old capability comparison and its source qualifications remain intact; N5 supplies the separately checked physical brief scope.
Work a six-GPU comparison at the correct physical boundary
Fictional EX-PD137 compares two proposed populations for six GPUs. A names three R770 servers with two H100 NVL GPUs per server; B names three R770 servers with two L40S GPUs per server. The quantities are stipulated for this example. They are not actual supplier configurations, quotations or supported BOMs, and no mixing of GPU models within a server is proposed.
Both populations contain three planned server enclosures. Applying the named 2U chassis description gives a 6U sum of chassis heights for each alternative. That is not a rack layout or a claim that either population fits an available rack. Rails, usable mounting positions, cable management, service access, weight and the actual delivered configuration are not supplied by this arithmetic.
Multiplying the two selected 350 W GPU entries gives a 700 W GPU-field subtotal per proposed server and 2,100 W across each six-GPU population. The equality is useful: it shows why these particular OEM GPU entries cannot support a claim that one population necessarily needs a larger facility power allocation. The subtotal excludes the rest of each server and is neither measured consumption nor a whole-server maximum. It cannot size a rack PDU, circuit or cooling system.
Now change the requested assembly. A's application brief explicitly requires a bridged pair in each server, while B does not require NVLink. The H100 NVL brief explains the three-bridge assembly at the GPU-pair level, but the cited R770 option list does not confirm that it is supplied or supported in A. Do not convert three planned pairs into a bridge order or an installation instruction. The missing OEM configuration answer is material even though the chassis and selected GPU-power arithmetic match. For B, the negative NVLink field is a known product characteristic, not an unresolved promise that more L40S cards could supply that feature.
The resulting decision is concrete: preserve the common chassis and GPU-field quantities, distinguish the interconnect assembly, and obtain each complete server's applicable physical configuration and system-level demand before releasing site-dependent work. No amount of label matching settles those remaining decisions. The earlier application-feature screens still apply, and actual operating performance, installation permission and acceptance remain separate.
Scroll the table horizontally on smaller screens.
| Comparison item | A: stipulated H100 NVL population | B: stipulated L40S population | What can be concluded | Source |
|---|---|---|---|---|
| Server and GPU units | Three proposed enclosures; two GPUs each; six GPUs total. | Three proposed enclosures; two GPUs each; six GPUs total. | Equal stated populations, not equivalent application capability or approved system configurations. | Planning analysis |
| Chassis-height arithmetic | 3 × 2U = 6U of chassis heights. | 3 × 2U = 6U of chassis heights. | The stipulated comparison has the same chassis-height sum. No actual rack fit, placement or service-space allowance is derived. | [D2] |
| Selected OEM GPU-field arithmetic | 2 × 350 W = 700 W per proposed server; 6 × 350 W = 2,100 W across the population. | The same 700 W and 2,100 W GPU-field subtotals. | Only a subtotal of the selected listing fields; actual whole-system and facility requirements are absent. | [D1] |
| Interconnect requirement | A requests bridged pairs; its complete OEM arrangement is not returned. | B does not request NVLink; L40S has no NVLink support. | A has an unresolved supplied-assembly dependency. Neither matching count nor a general GPU product brief resolves it. | [N2][N5] |
Original fictional planning arithmetic only. No populated rack, delivery availability, OEM validation, measured load, thermal result, supported bridge quantity or price is asserted.
Work through two fictional returned options
Fictional example EX-OPT-16: a buyer receives two proposed R770 configurations. Option A names two H100 NVL GPUs. Option B names two L40S GPUs. These are illustrative descriptions, not actual orders, supported bills of materials, prices or customer results. The supplier has not yet established the complete configuration of either option.
Consider three alternative briefs from the application team, rather than assuming a single combined operating mode. One asks for MIG-backed instances; another asks for an explicit NVLink arrangement instead of MIG. A comparison based only on “two GPUs per server” would hide the distinction. The following screens do not establish that MIG and NVLink can be used together in a particular configuration.
Scroll the table horizontally on smaller screens.
| Buyer input | Option A: two H100 NVL GPUs | Option B: two L40S GPUs | Source |
|---|---|---|---|
| Brief 1 explicitly requires MIG-backed instances. | The advertised hardware capability does not rule it out. Actual instance profile, software stack and complete-system support remain unconfirmed. | The specified GPU lacks MIG. A proposal describing vGPU support has not answered the stated MIG requirement. | [N1][N2][N3] |
| Alternative brief 2 requires an NVLink connection rather than MIG. | The GPU lists NVLink support; the server’s actual supported connection arrangement and supplied parts are still unanswered. | The GPU does not list NVLink support. More cards do not establish that missing capability. | [N1][N2] |
| Alternative brief 3 requires neither feature and instead specifies an application latency target. | The earlier capability screen no longer chooses the option. Comparable workload evidence is needed. | The earlier capability screen no longer rejects the option. Comparable workload evidence is needed. | Planning analysis |
Original fictional decision aid, not a recommendation to purchase, configure or operate either system. Source anchors support only the repeated product capability facts. The final row is planning analysis; no measured result, license entitlement or application fit is asserted.
Separate the graphics and sharing questions
NVIDIA positions L40S for graphics and media as well as AI workloads. If the request includes rendering, a virtual workstation or video processing, identify the actual application and delivery mode. An AI-only specification line does not show whether those requirements were evaluated. This product positioning does not prove that a particular application is supported or performs adequately on a returned server.
L40S also illustrates why “no MIG” should not become “no shared GPU use of any kind.” NVIDIA lists vGPU software support separately from its negative MIG field. That is a different software-support question, not another name for MIG. The software owner needs to confirm the relevant product, release, supported server and entitlement for the intended delivery mode. This comparison does not approve a hypervisor, VM profile, tenant-isolation design or licensing arrangement.
The useful outcome is a clear requirement: a named MIG capability, a defined graphics or media application, a specified sharing model, or none of these. Then request evidence for that requirement. Adding every available feature to the brief without a user need can obscure the comparison just as much as omitting a necessary one.
Keep the performance decision separate
A hardware capability can make an option ineligible for an explicit requirement. It does not establish that the remaining option is fastest, best value or adequate for the workload. This article intentionally does not rank H100 NVL and L40S by peak arithmetic rates or translate GPU memory into a supported model size.
When both options meet the feature brief, compare a defined application result: the model or application version, workload, accuracy or quality condition, concurrency, software stack and complete system need to be identifiable. A latency target and a throughput target also describe different outcomes. Preserve the actual evidence and its conditions rather than transferring a headline from another benchmark. The linked GPU-speed guide covers this comparison in detail.
If the result is missing, the honest status is an unresolved selection question. A matching GPU count, a larger memory label or the word “enterprise” cannot supply the missing application evidence.
Pass a selected server configuration to the physical project
Once the customer and supplier have selected and confirmed the complete server configuration, use its actual OEM documents for the physical project. A PCIe board specification is not a rack-mount system drawing. The equipment supplied, mounting package, site conditions and authorized work determine the installation scope. An apparent match within one server family does not establish that a different option reuses every installation input.
Leviathan can discuss separately scoped rack-and-stack installation, physical equipment mounting, structured cabling, routing, dressing, labeling and agreed physical verification. Handling requirements remain site and equipment specific. Any formal testing requires confirmation of equipment, qualified personnel, method, acceptance criteria and report format for that engagement. Identify the appointed equipment/material supplier and delivery responsibilities. MIG or vGPU configuration, application validation and full system commissioning are not standard services. Liquid-side work, including coordination or subcontracting, and electrical installation are excluded.
Bring the confirmed server configuration and the physical work you need to the scope discussion. That connects the accelerator decision to a real installation package without presenting a product comparison as installation or workload approval.
Common planning questions
Is every H100 an H100 NVL?
No. NVIDIA separates H100 SXM and H100 NVL in its current product table, with different memory and form-factor entries. This comparison uses the H100 NVL option; an unspecified H100 quotation needs an exact variant before its fields can be compared.
Does the same R770 server name make H100 NVL and L40S interchangeable?
No. Dell lists them as separate GPU options, not a blanket exchange or mixing approval. Their advertised MIG, NVLink and memory characteristics differ. The exact supported OEM configuration and the application requirement must both be confirmed.
Does L40S having no MIG mean it cannot support any shared GPU use?
No. NVIDIA separately lists L40S vGPU software support and no MIG support. Do not treat those fields as synonyms. The intended sharing model, supported software and server configuration, entitlement and actual delivery still need confirmation by the software owner.
What physical work can Leviathan scope, and does it include MIG or vGPU configuration?
Leviathan offers physical installation and cabling for the agreed OEM configuration and site, with the equipment supplier and delivery responsibilities identified. MIG or vGPU configuration, application validation and full system commissioning are not standard services. Specific hardware availability, permitted work and acceptance requirements must be confirmed for the engagement.
Does the same two-GPU R770 option imply different rack space for H100 NVL and L40S?
Not from those fields alone. Dell describes the R770 as a 2U server and lists up to two double-width GPUs for each named option. The actual chassis configuration, rails, mounting and service access still govern the physical layout. A GPU model label is not a rack-space calculation.
Why can the H100 NVL product brief say 400 W while the R770 GPU entry says 350 W?
They describe different source boundaries: a board maximum in a specified NVIDIA product-brief context and a named Dell system option entry. Preserve both at their stated scopes. This comparison does not prescribe a power setting, explain an unverified OEM implementation or turn either number into complete-server or facility demand.
Primary sources
- [N1] NVIDIA H100 product specifications: SXM and NVL columnsChecked 2026-10-16. Only named H100 NVL/SXM memory, form factor, MIG and NVLink fields and the separate NVL paired-memory description are adopted. Marketing speed, price, license inclusion, support and availability claims are not adopted.
- [N2] NVIDIA L40S full specificationsChecked 2026-10-16. Only exact L40S memory, PCIe/form-factor, MIG, NVLink and separate vGPU-support fields and broad graphics/media positioning are used. No performance, display-output, power, cooling-design, licensing or availability conclusion.
- [N3] NVIDIA MIG User Guide: supported GPUsChecked 2026-10-16. Corroborates named H100 PCIe hardware support only. No instance creation, driver selection, VM configuration, profile allocation or isolation guarantee.
- [N4] NVIDIA H100 NVL Product Brief PB-11773-001_v01, March 2024Checked 2026-10-16. Current H100 page links this March 2024 edition. Only named board identity, HBM3/94GB, MIG and NVLink capability are corroborated; no historical driver, bridge installation/topology, thermal/power or operating instructions adopted.
- [D1] Dell PowerEdge R770 Installation and Service Manual: GPU specificationsChecked 2026-10-16. Only separate up-to-two H100 NVL 94GB and L40S 48GB double-width entries for R770 are adopted. No arbitrary combination, actual order, mixing, bridge support, stock, site fit or installation method is approved.
- [L1] Leviathan rack-and-stack service scopeChecked 2026-10-16. Current physical installation scope and exclusions. Equipment/material sourcing has its own approved purchasing and delivery scope; accelerator/software configuration and system commissioning are not standard services.
- [L2] Leviathan physical verification and formal-test boundariesChecked 2026-10-16. Formal equipment, qualified personnel, method, acceptance criteria and report format require actual engagement confirmation. No default formal-testing resources or procedure implied.
- [D2] Dell PowerEdge R770 Installation and Service Manual: system configurations and featuresChecked 2026-10-19. Only the named complete server's 2U identity is adopted. It does not establish a particular GPU/riser/bridge/rail/cooling configuration, actual site fit, availability or work authority. Other configuration, operating and warranty instructions are not transferred.
- [N5] NVIDIA H100 NVL Product Brief PB-11773-001_v01: physical comparison fieldsChecked 2026-10-19. Separate new physical-field use of the March 2024 brief, alongside the unchanged narrower N4 citation: full-height/full-length 10.5-inch dual-slot/passive identity, a three-bridge pair with additional space, and its source-specific 400 W maximum only. No numeric clearance, placement/topology, auxiliary power connection, setting, operating, installation or actual OEM-system approval adopted. D1's two named 350 W GPU entries and N2's nominal form-factor/passive fields were also rechecked October 19; their original source objects and dates remain exact.