The accelerators — and what they will not do

Two things people are surprised by

They cannot encode video. At all.

NVIDIA’s own encode and decode support matrix records zero NVENC engines on the A100 and zero on the H100. They decode — five engines on the A100, seven on the H100 — but there is no encoder on the board, so every frame these cards produce has to be encoded by the processor instead. If your pipeline has video coming out of it, this is the wrong shelf, and the L40S with its three AV1 encoders is the right one. That is one page across at /l40s.

These cards also have no display outputs. They are headless compute parts. Nothing that expects a framebuffer — a virtual desktop, a remote workstation, a game server — belongs on them.

The A100 is not obsolete, and the 40 GB card is not a mistake

The A100 is the previous generation, and it reads its weights at 1,555 GB/s with 40 GB and 1,935 GB/s with 80 GB. For a model that fits, that read rate is what decides how many tokens a second you get, and it remains far ahead of every non-stacked card in the catalogue. The 40 GB card holds a thirty-billion-parameter model at eight-bit precision comfortably, which is most of what people actually serve, at a fraction of what the H100 costs each month.

Buy the H100 when the model genuinely needs Hopper — the transformer engine, FP8, the newer tensor units — or when the run has to finish sooner rather than merely finish. Buy an A100 when the question is whether it fits and how fast it reads, which is the more common question.

The specifications, as published

Figures from NVIDIA’s own datasheets and the Tesla specification tables, cited in the source of this page. Where a figure could not be sourced against the exact part we fit, this page prints nothing rather than a guess.

The H100 shows no core count and no bandwidth above, and that is deliberate. Our part is recorded in the store under a core count belonging to the SXM module, which is a card that bolts to a baseboard rather than into a slot; the 80 GB HBM2e memory it also records belongs to the PCIe card, which is what actually goes into these machines. Rather than pick one and print it confidently, the page prints the memory both sources agree on and omits the rest. Ask us and we will tell you exactly which board is going in yours.

The three of them side by side, and what sits above

From our own shelves. All of these are orderable today, in the same chassis, so the question is which is worth the difference.

The last row is there because it is the honest comparison and most listings hide it: the RTX PRO 6000 Blackwell holds 96 GB, more than any accelerator on this page. Read its memory and its money against the H100’s in the table above rather than taking our word for either. What it gives up is read rate and NVLink; what it gains is four video encoders and ray tracing. If your question is will it fit, read /rtx60 before you commit to this shelf.

Which machine they go in, and what link they get there

The A100 is a PCIe Gen 4 card and the H100 is Gen 5, while our platforms are Gen 3 and Gen 4. The H100 therefore never gets the full bus it was designed for on any machine we run, and this page says so rather than leaving you to find out.

Whether that costs you anything depends entirely on the shape of the work. An inference endpoint whose weights are resident barely touches the slot and will not notice. A training loop streaming a fresh batch from host memory on every step will notice on every step. If you are not sure which you are, tell us what the job is before you order — it decides the machine, not the card.

What actually fits in 80 GB

The 40 GB A100 holds the same table one rung down: a thirty-to-thirty-four-billion-parameter model at eight-bit precision fits, the same model at sixteen-bit does not. A seventy-billion model at sixteen-bit does not fit any single card we sell — that is a two-card build, and the reason the NVLink line above matters.

What they cost, with the machines they go in

The card is one line on the invoice and the machine is another. Both halves below come from one reading of the order catalogue, so the price and the specification beside it always belong to each other.

Bandwidth is unmetered in both directions with no egress charge, which on this shelf is usually the largest hidden cost elsewhere — a training run pulling a dataset in and pushing checkpoints out moves a great many bytes, and per-gigabyte egress is charged on every one of them. One IPv4 address is included, there is no contract and one month is the default term. For every other card we fit, priced the same way, read the whole catalogue, priced. For machines racked and ready today, /instant.