Massed Compute

A 1x instance is $2.89 an hour as of 16 September 2026. Confirm live 1x and 8x rates on GPU pricing before you launch.

Stay on 1x when weights plus KV still fit 80 GB. Take 8x when you need 16-bit 70B, a cache that spills one card, or tensor parallel across NVLink.

If the model already fits 32 GB or 48 GB, start on that card. This page is the 80 GB Hopper SXM5 product.

Pick the H100 SKU

Massed Compute sells three H100 products. Each one has its own SKU and fabric. Name the SKU so you land on the right board.

This article is H100 SXM5. Hopper SXM, 80 GB HBM3, NVLink on the 8x box. SKUs gpu_1x_H100_SXM5 and gpu_8x_H100_SXM5. The live catalog has no 2x or 4x SXM5.

H100 PCIe 80 GB is a different board. Live 1x is gpu_1x_h100. 2x is gpu_2x_h100. 4x is gpu_4x_h100. Guides that say “launch a 2x H100” mean this family.

H100 NVL is a third family. Live 1x is gpu_1x_h100_nvl. Marketplace GPU type H100 NVL is the 94 GB module, not the 80 GB SXM5 board. 2x is gpu_2x_h100_nvl. 4x is gpu_4x_h100_nvl. Separate NVLink NVL SKUs exist in the catalog (gpu_2x_h100_nvl_nvlink). This pull showed no live capacity on those NVLink NVL rows.

A PCIe H100 spot SKU also exists (gpu_1x_h100_spot). Spot can stop. Use on-demand SXM5 for this product.

Same Hopper generation. Different board, different disk, different interconnect. A PCIe pair can split a job over PCIe. SXM5 NVLink is the 8x SKU.

Recipe posts that say “2x H100” mean gpu_2x_h100. That is 40 vCPU and 2,500 GB disk on PCIe.

View current GPU pricing

1x and 8x shapes

On demand has no contract and no minimum. You pay for the hours you use. Confirm live pricing before you launch.

gpu_1x_H100_SXM5 is one 80 GB module, 14 vCPU, 122 GiB RAM, and 350 GB disk. gpu_8x_H100_SXM5 is eight modules (640 GB combined), 126 vCPU, 976 GiB RAM, and 2,800 GB disk.

You are renting one eight-GPU NVLink box, not eight separate 1x VMs. Memory, RAM, and disk scale with the eight modules. vCPU does not land on a clean 8×. You get 126, not 112. A request that already fits 80 GB does not get eight times faster on eight cards.

Use 8x when the model, the KV cache, or the training shard plan does not fit one 80 GB GPU. Stay on 1x when a single module already holds the work.

One 1x SXM5 has 350 GB of disk and 14 vCPUs. The PCIe 1x H100 has 1,250 GB of disk and 20 vCPUs. Same Hopper name, different shape. If you keep several 70B checkpoints plus a container image, that 350 GB fills faster than the PCIe 1x box.

View current GPU pricing

What we checked on a live 1x

We launched a Massed Compute 1x H100 SXM5 on 10 September 2026. Ubuntu Server 24.04 with NVIDIA drivers. SKU gpu_1x_H100_SXM5. Hardware only. No model was served, so this page does not invent 1x tok/s.

ECC uncorrected was 0. Persistence mode was on. Idle draw was 68 W against a 700 W cap. Temperature was 24 C at 0% util. Device nodes /dev/nvidia0, /dev/nvidiactl, and /dev/nvidia-uvm were present.

nvidia-smi topo -m printed a single GPU0 cell. The NV# bonded set is not there because there is no peer GPU. That is expected on 1x. The NVLink box is gpu_8x_H100_SXM5.

80 GB, FP8, and the 8x fabric

If weights plus KV fit in 80 GB, the job runs. If they do not, you usually get an out of memory error rather than a slow-but-correct run. Some stacks can offload to system RAM. That path is much slower.

On gpu_8x_H100_SXM5 the eight modules share NVLink. Tensor parallel and pipeline parallel belong here because the shards have a fast path between cards. PCIe 2x and 4x can split a job too. They are a different interconnect.

Hopper also runs FP8 in the Transformer Engine. That is why 1x SXM5 is a strong fit for high-throughput and long-context serving when the model still fits one card. You shrink the weights, keep more of the 80 GB for KV, and stay on one GPU.

We have a public vLLM run of GLM 5.3 Flash on gpu_8x_H100_SXM5 in the gpu-benchmark repo. Official vllm/vllm-openai:glm53-flash image. Tensor parallel 8. KV in bfloat16. Output token throughput at concurrency 32 was 698 tok/s. Same capture was 1041 tok/s at concurrency 16 and 1824 tok/s at concurrency 64. Median time to first token at concurrency 32 was 321.2 ms. Those figures are for that model and engine. They are not a general SXM5 speed rating, and they are not a 1x number.

A public Llama 3.1 70B TGI page exists too. That run used 2x H100 PCIe, not SXM5. Use it as PCIe context. Those charts are a different SKU and engine.

1x or 8x

Stay on gpu_1x_H100_SXM5 when weights plus a real context window still fit 80 GB at FP8 or 4-bit. One serving process. One GPU. An OpenAI-shaped API. LoRA or QLoRA on a base that already fits. The other seven cards stay off the bill.

Move to gpu_8x_H100_SXM5 when BF16 or FP16 weights are larger than 80 GB (a 70B at 16-bit is this case), when FP8 weights fit but KV at your context and batch spills the rest of the card, or when you are sharding one model with tensor parallel or pipeline parallel. Full-parameter train and continued train with DeepSpeed, FSDP, or Axolotl across eight GPUs belong here too.

Extra cards buy aggregate memory and a fabric. They do not make a single request eight times faster if that request already ran on one GPU. The RTX PRO 4500 guide shows the same pattern on a smaller card. Decode stayed flat while memory per card fell. Treat 8x SXM5 the same way.

Disk is part of the decision. 1x SXM5 is 350 GB. 8x is 2,800 GB. A 70B-class checkpoint is tens of gigabytes. Two copies plus a container image plus logs will press the 1x disk. If you need several weights on disk, 8x storage is the headroom, or you pull one checkpoint at a time on 1x.

vCPU and system RAM scale too. 1x is 14 vCPU and 122 GiB. 8x is 126 vCPU and 976 GiB. Tokenizers, dataloader workers, and CPU-side preprocess sit in that RAM. The 8x hourly covers that extra CPU, RAM, disk, and NVLink.

70B-class sizing (sizing, not a benchmark)

Llama 3.3 70B is a named 70B-class instruct model. The math below is sizing only. We did not run that checkpoint on SXM5.

Parameter count times bytes per weight. 70 billion times 2 bytes is about 140 GB. Times 1 byte is about 70 GB. Times 0.5 bytes is about 35 GB. Activations and the KV cache sit on top of that.

KV is often what spills after the weights already fit. It grows with context length and with batch size. A short chat on FP8 70B is a 1x job. A long window, a wide batch, or a second model in the same process is how you walk off 80 GB.

For a Llama-class 70B with grouped-query attention (80 layers, 8 KV heads, 128-dim), 16-bit KV is about 0.31 MB per token. That is sizing math, not a bench.

Raise batch size and those KV numbers multiply. Raise precision of the cache and they grow again. Cut KV to FP8 and they shrink. The table is a planning sketch. Measure your engine’s real reservation with nvidia-smi on the SKU you launch.

If you need 16-bit 70B, 1x is the wrong shape. That is an 8x SXM5 job, or a different product with more memory per GPU. RTX PRO 6000 Blackwell is 96 GB. Those are different cards. Their tok/s are not interchangeable.

This page does not claim a Llama 3.3 70B tok/s number on SXM5. We did not serve that model on the 1x smoke.

Use cases

70B-class inference on 1x

Serve a 70B-class instruct model with vLLM, SGLang, or TensorRT-LLM on gpu_1x_H100_SXM5.

This is the default when weights and a real context window still fit 80 GB at FP8 or 4-bit. One GPU keeps the serving stack simple. You do not pay for seven idle cards.

Those recipe posts were measured on other SKUs.

• vLLM was documented on gpu_1x_l40 (L40, 48 GB), Ubuntu Server 24.04 with NVIDIA drivers, ungated Qwen 2.5 0.5B on port 8000
• SGLang was validated on an L40 (gpu_1x_l40) on 20 August 2026
• TensorRT-LLM was validated on PCIe gpu_1x_h100

The install steps still apply. Launch gpu_1x_H100_SXM5 when you want this product. Keep the L40 and PCIe SKUs for the cards those posts actually measured.

High-throughput, long context, and the 8x box

FP8, a wide context, and a busy batch all grow KV. Stay on 1x while the cache fits. Move to gpu_8x_H100_SXM5 when the model or the KV spills one GPU.

Tensor parallel across NVLink is why the 8x SKU exists. You are buying 640 GB of combined GPU memory and a fabric that can split one large model.

The GLM 5.3 Flash 8x page is the public example of that shape. Tensor parallel 8 on gpu_8x_H100_SXM5. Cite those tok/s only for that model.

Pick 8x when you are sharding one model or one train job across eight Hopper GPUs. Keep 1x when a single 80 GB module already holds the work.

Fine-tune and continued train

LoRA and QLoRA on 1x when the base model fits. That is the simple path for adapters and short continued-train runs.

When the model does not fit one 80 GB card, use DeepSpeed, PyTorch FSDP, or Axolotl on gpu_8x_H100_SXM5. NVLink is the point of that SKU.

DeepSpeed and Axolotl were measured on PCIe gpu_2x_h100. Retarget the same tools at SXM5. Use gpu_2x_h100 when you want that PCIe pair. Use gpu_8x_H100_SXM5 when you want this NVLink box.

The DeepSpeed post is explicit about 70B full fine-tune. That job was not run on the PCIe pair. The SKU it names for that work is 8x H100 SXM / NVLink. That is gpu_8x_H100_SXM5 on this marketplace.

FSDP is the PyTorch-native sharded-train path on the same 8x box. Same reason as DeepSpeed. You want eight Hopper GPUs that can talk over NVLink.

Tool paths and SKU-safe launches

Five common stacks. The public walkthroughs are real. Each one was measured on another SKU. Launch gpu_1x_H100_SXM5 or gpu_8x_H100_SXM5 when this is the product you want.

A full vLLM bootstrap is a longer install than a hardware check. This article’s 1x smoke did not serve a model.

When H100 SXM5 is not the fit

Pick the card that matches the model. A mid-size model that fits 32 GB or 48 GB is a better match for those cards than Hopper SXM. Confirm live rates on GPU pricing before you launch.

The RTX PRO 4500 Blackwell guide covers that 32 GB card in depth. Use it for 8B to 32B work that fits. Native FP4 is the reason to pick it over an older 32 GB or 48 GB card at a similar hourly.

L40S (gpu_1x_l40s) is 48 GB. That is a strong Ada inference card when 32 GB is tight and 80 GB is more than you need.

RTX PRO 6000 Blackwell 1x (gpu_1x_pro_6000_blackwell) is 96 GB. A 70B at FP8 that is tight on 80 GB can be a 1x 96 GB job instead of an 8x Hopper box. It does not give you Hopper NVLink across eight GPUs.

Blackwell SXM is the step above Hopper when you need that next generation on the same marketplace. This pull lists gpu_8x_b200_SXM6 and gpu_8x_b300_SXM6. This pull showed no live capacity for those SKUs, so they are catalog rows, not a launch CTA.

How to launch an H100 SXM5 instance

On Deploy, GPU type is a dropdown. Quantity is a row of tiles. For this family the type label is H100 SXM5 (80GB). The tiles that belong to it are ×1 and ×8. There is no ×2 or ×4 tile for this type.

1. Sign in to the Massed Compute marketplace and open Deploy
2. Under GPU type, pick H100 SXM5 (80GB)
3. Under Quantity, pick ×1 (gpu_1x_H100_SXM5) or ×8 (gpu_8x_H100_SXM5)
4. Choose a Category, then an Image or setup recipe that matches your work. Ubuntu Server 24.04 with NVIDIA drivers is the usual start.
5. Add an SSH Key and a Startup Command if you want them. Both are optional.
6. Launch

The instance comes up in a few minutes with a public IP. The 1x smoke on 10 September 2026 reached Running with the GPU visible to nvidia-smi.

There is a faster way to do all of this. The Massed Compute MCP server connects an AI assistant like Claude, Cursor, or ChatGPT straight to your account. You describe the machine you want in plain language, and it checks live inventory, launches the instance, and hands back your SSH details. It can also track what you are spending and shut things down when you are finished. See what the MCP server can do and how to set it up

The instances_launch sku names for this product are

gpu_1x_H100_SXM5
gpu_8x_H100_SXM5

Copy one of those names into MCP or the API. Use any to let the platform pick available capacity.

Launch an H100 SXM5 instance

Guides that name H100, with the right SKU

These walkthroughs are live and useful. Each one was measured on another SKU. Launch gpu_1x_H100_SXM5 or gpu_8x_H100_SXM5 when you want this family. A “2x H100” checkout path on those posts is PCIe gpu_2x_h100.

• Deploy vLLM with an OpenAI-compatible API (documented on L40 gpu_1x_l40)
• Deploy SGLang with an OpenAI-compatible API (validated on L40 gpu_1x_l40)
• Deploy TensorRT-LLM on NVIDIA GPUs (validated on PCIe gpu_1x_h100)
• Fine-tune with Axolotl on multi-GPU VMs (validated on PCIe gpu_2x_h100)
• Train with DeepSpeed on multi-GPU VMs (validated on PCIe gpu_2x_h100)

vLLM is the common 8x tensor-parallel engine on this card as well. The GLM 5.3 Flash gpu-benchmark page is the public 8x SXM5 example.

Common questions about H100 SXM5

What does 1x H100 SXM5 cost per hour?

$2.89 on demand with no contract, current as of 16 September 2026. Confirm live 1x and 8x rates on GPU pricing before you launch.

What GPU type do I pick on Deploy?

H100 SXM5 (80GB). Then pick quantity ×1 or ×8. That maps to gpu_1x_H100_SXM5 and gpu_8x_H100_SXM5.

Is there a 2x or 4x H100 SXM5?

No. The live catalog has 1x and 8x only. 2x and 4x H100 in the catalog are PCIe (gpu_2x_h100, gpu_4x_h100).

Is 8x eight times the 1x price?

No. The 8x SKU includes more RAM, more disk, and NVLink. You rent one eight-GPU box. Check live 1x and 8x rates on GPU pricing.

When should I pick 8x instead of 1x?

When the model or the KV cache does not fit 80 GB, or when you are training with DeepSpeed, FSDP, or Axolotl across eight GPUs. If the job already fits 1x, stay on 1x. 8x is capacity. One request that already fit 80 GB does not get eight times faster.

How is SXM5 different from PCIe H100?

SXM5 is the Hopper SXM module with 80 GB HBM3. The 8x SKU adds NVLink. PCIe H100 is a different SKU, with 1x, 2x, and 4x shapes. 1x SXM5 also has less disk (350 GB) and fewer vCPUs (14) than PCIe 1x (1,250 GB and 20 vCPUs). A PCIe pair can split a job. The SXM5 NVLink domain is gpu_8x_H100_SXM5.

Does 1x SXM5 have NVLink?

Not as a multi-GPU domain. On a live 1x host on 10 September 2026, nvidia-smi showed NVLink links inactive and topology showed a single GPU with no NV# peer. NVLink is why you launch gpu_8x_H100_SXM5.

How big is the 1x disk?

Catalog storage is 350 GB. The 1x host we checked showed 344G formatted on /. PCIe 1x H100 is 1,250 GB. Plan checkpoints against 350 GB on 1x SXM5.

What image should I use?

Ubuntu Server 24.04 with NVIDIA drivers. The 1x smoke on that image showed driver 580.126.16 and CUDA 13.0 in nvidia-smi. That is the same image family the public vLLM, SGLang, TensorRT-LLM, Axolotl, and DeepSpeed posts start from.

Can I launch with MCP?

Yes. Install the Massed Compute MCP and launch gpu_1x_H100_SXM5 or gpu_8x_H100_SXM5. Or open Deploy, pick H100 SXM5 (80GB) under GPU type, and choose ×1 or ×8.

Launch H100 SXM5

Check live GPU pricing, then launch gpu_1x_H100_SXM5 or gpu_8x_H100_SXM5.