The NVIDIA H100 and H200 look similar on paper, but the H200's larger memory and higher bandwidth make it the smarter buy for many AI workloads. This guide compares the specs, the real UK pricing and the workloads each GPU handles best, so you can decide with confidence.
H100 vs H200: The Short Answer
The H200 is the H100 with a bigger, faster memory system. Both cards use the same Hopper GH100 die, so raw compute is identical. The H200 replaces 80GB of HBM3 memory with 141GB of HBM3e, and pushes bandwidth from 3.35TB/s to 4.8TB/s. For memory-bound AI workloads, that is the difference between a model that fits in GPU memory and one that does not.
For UK enterprises, the practical question is not which card is faster. It is which card fits your models, your rack power budget and your budget. This guide compares both across the specs that actually matter, with real prices from the Eton Technology range.
H100 vs H200: Full Specification Comparison
The table below compares the common variants Eton Technology supplies. The H100 ships as an SXM module on HGX baseboards and as a PCIe card; the H200 ships as the SXM version and the NVL PCIe card.
| Specification | NVIDIA H100 (SXM) | NVIDIA H100 (PCIe) | NVIDIA H200 (SXM) | NVIDIA H200 (NVL PCIe) |
|---|---|---|---|---|
| GPU architecture | Hopper GH100 | Hopper GH100 | Hopper GH100 | Hopper GH100 |
| GPU memory | 80GB HBM3 | 80GB HBM2e | 141GB HBM3e | 141GB HBM3e |
| Memory bandwidth | 3.35 TB/s | 2 TB/s | 4.8 TB/s | 4.8 TB/s |
| FP8 Tensor Core | 3,958 TFLOPS* | 1,979 TFLOPS* | 3,958 TFLOPS* | 3,958 TFLOPS* |
| FP16 / BF16 | 1,979 TFLOPS* | 989 TFLOPS* | 1,979 TFLOPS* | 1,979 TFLOPS* |
| NVLink | 900 GB/s | None | 900 GB/s | 900 GB/s (bridge) |
| TDP | Up to 700W | Up to 400W | Up to 700W | Up to 600W |
| Form factor | SXM5 on HGX board | PCIe Gen5 dual-slot | SXM5 on HGX board | PCIe Gen5 dual-slot |
| Server fit | 4 or 8 GPU baseboards | Standard PCIe servers | 4 or 8 GPU baseboards | Standard PCIe servers, 2/4-way NVLink |
*With sparsity. Dense throughput is half these figures.
Memory and Bandwidth: Where the H200 Pulls Ahead
The H200's 141GB of HBM3e is 76% more capacity than the H100's 80GB, and the 4.8TB/s bandwidth is around 43% faster than the SXM variant. That matters most for large language models, because the KV cache grows with context length. A bigger memory pool means longer contexts, larger batch sizes and fewer times the model spills to CPU memory.
Inference benchmarks from NVIDIA put the H200 at up to 1.4x to 1.7x faster than the H100 on memory-bound LLM workloads. Training throughput gains are smaller, because training is often compute-bound rather than memory-bound.
The caveat: if your workloads are small models, image pipelines or classic HPC simulation with modest memory needs, the extra memory buys you little. The H100 still handles those workloads at the same speed, because the compute silicon is identical.
Compute: The Two Cards Share the Same Die
Both the H100 and H200 are built on the Hopper GH100 die with 132 SMs. FP8 Tensor Core throughput is 3,958 TFLOPS with sparsity on the SXM variants, and the CUDA core counts are identical. There is no generational leap in compute between them, which is why NVIDIA positions the H200 as a memory upgrade rather than a new architecture.
That has a practical consequence for buyers: if you already run H100s and your bottleneck is compute, upgrading to H200 will not help. If your bottleneck is memory capacity or bandwidth, the H200 is the fix. This is the single most useful way to frame the buying decision.
Form Factors: SXM, PCIe and the NVL Difference
The H100 is available as an SXM5 module (fitted to 4 or 8 GPU HGX baseboards) and as a PCIe card. The H200 follows the same pattern, with one important addition: the H200 NVL is a PCIe card that supports 2-way and 4-way NVLink bridges, so two or four cards can share a pooled memory pool of 282GB or 564GB.
For UK data centres, the NVL form factor is often the easier route in. It slots into standard PCIe Gen5 servers from HPE, Dell and Supermicro, avoids the power and cooling demands of a full HGX baseboard, and scales incrementally. The HGX route delivers more density per rack but needs the infrastructure to support it.
Eton Technology supplies H100 and H200 HGX baseboards and PCIe cards across its NVIDIA GPU range.
NVLink and Multi-GPU Scaling: What Changes
NVLink is where the H200 NVL changes the game for PCIe buyers. The H100 PCIe card has no NVLink support at all, which means multi-card H100 PCIe setups communicate over the PCIe bus only. That works for embarrassingly parallel workloads, but it is a serious bottleneck when a single model needs to span multiple GPUs.
The H200 NVL supports 2-way and 4-way NVLink bridges at 900GB/s per GPU. Two cards pool into a 282GB memory space and four cards pool into 564GB. That opens the door to running models on PCIe hardware that previously required an HGX baseboard: a four-card H200 NVL node can hold a large LLM in GPU memory with fast all-to-all communication between the cards.
The practical implication for UK buyers is that the H200 NVL is a genuine scale-out option in standard servers. If you expect to grow from one card to two or four, the H200 NVL keeps that path open. The H100 PCIe does not.
The H200 NVL technical guide covers bridge configurations, pooled memory and server platform compatibility in more detail.
Power and Cooling: What UK Data Centres Need to Know
Both cards run hot, and the power envelope is a real constraint in UK facilities where rack power is billed by the kilowatt. The H200 SXM is rated up to 700W and the H100 SXM up to 700W, so a fully loaded 8-GPU HGX baseboard draws in the region of 5.6kW to 6kW plus the host server. Dense HGX racks typically need liquid cooling or high-power air handling.
The PCIe variants are easier to manage. The H200 NVL is rated up to 600W and the H100 PCIe up to 400W. A 4-GPU H200 NVL node draws roughly 2.4kW for the GPUs alone, which sits comfortably within a standard 8-12kW UK rack budget. Air cooling is sufficient for most NVL deployments, which is why the form factor is popular for facilities that cannot retrofit liquid cooling.
When comparing costs, remember to include power and cooling in the total cost of ownership. A cheaper card that forces a rack upgrade or liquid-cooling retrofit is rarely the cheaper option once you add the facility costs.
Talk to the Eton Technology team about power, cooling and rack planning before you commit to a configuration.
Which Workloads Suit Which Card?
Use this as a starting point. The right answer for your environment depends on model size, serving concurrency and whether you are training or running inference.
| Workload | H100 | H200 | Why |
|---|---|---|---|
| LLM inference (large models) | Runs, but memory-constrained | Best fit | 141GB fits larger models and longer contexts |
| LLM fine-tuning / LoRA | Works well | Better for larger models | Bigger memory pool handles bigger batches |
| Small model inference | Good | No advantage | Compute-bound, memory not the limit |
| HPC simulation (GROMACS, NAMD) | Good | Better for memory-heavy sims | Bandwidth helps memory-bound codes |
| RAG / embeddings | Good | Good | Fits comfortably in either |
| Video / image processing | Good | No advantage | Rarely memory-bound |
| Training at scale | Good | Similar | Training is compute-bound |
Real UK Pricing: What These Cards Cost
Prices move with supply and demand, and the H200 has been surprisingly competitive in the UK market. The figures below are the list prices on etontechnology.com at the time of writing.
| Product | Price (excl. VAT) |
|---|---|
| NVIDIA H100 80GB PCIe (new retail) | £29,425.00 |
| NVIDIA H200 NVL 141GB PCIe (new) | £27,760.00 |
| NVIDIA H200 141GB 8-way HGX baseboard | £209,675.00 |
| NVIDIA H100 HGX SXM5 8-way (liquid cooled) | Price on request |
Prices checked August 2026. Contact Eton Technology for current stock, lead times and volume pricing.
The headline takeaway: the H200 NVL PCIe card is priced below the H100 PCIe card, even though it carries 76% more memory and 43% more bandwidth. That is an unusual market position, and it is why many UK buyers now spec H200 for new builds.
See the current list price on the H200 NVL product page, or compare with the H100 80GB PCIe listing.
H100 or H200: A Decision Framework
Choose the H200 if:
- You run or plan to run large language models where context length and batch size are hitting the 80GB limit.
- Your workloads are memory-bound: LLM inference, RAG over large corpora, memory-heavy HPC simulations.
- You are building new GPU capacity and want the best memory per pound on a PCIe card.
- You want the option of 2-way or 4-way NVLink pooling on PCIe cards without committing to an HGX baseboard.
Choose the H100 if:
- Your models fit comfortably in 80GB and your bottleneck is compute, not memory.
- You are expanding an existing H100 fleet and want consistency across nodes.
- Your budget favours the lower entry price of the PCIe variant, or you found H100 stock at a strong price.
- You need the mature ecosystem and validated driver stacks that come with a card already proven in production.
Our Recommendation for UK Enterprises
For most new AI infrastructure projects, the H200 is the better buy. The pricing position in the UK makes it hard to justify the H100 for new builds: more memory, more bandwidth, comparable price. The H100 remains the sensible choice for teams expanding existing fleets or running compute-bound workloads on a tight budget.
If you are unsure, the deciding question is simple: does your workload currently hit the 80GB memory ceiling? If yes, the H200 pays for itself in fewer nodes and simpler serving. If no, save the money and stick with the H100.
Browse both ranges in the NVIDIA GPU collection, or read the H200 buyers guide and the H200 NVL technical guide for deeper detail.
Frequently Asked Questions
Is the H200 faster than the H100?
For memory-bound workloads, yes, up to around 1.4x to 1.7x faster in NVIDIA's LLM inference benchmarks. For compute-bound workloads, the two cards are effectively identical because they share the same Hopper die.
What is the main difference between H100 and H200?
Memory. The H200 has 141GB of HBM3e at 4.8TB/s, while the H100 has 80GB of HBM3 at 3.35TB/s. Compute capability is the same on both.
Is the H200 worth the extra cost?
In the UK market right now, the H200 NVL PCIe card is actually priced below the H100 PCIe card, so there is no extra cost for more memory. For new builds, the H200 is usually the better value.
Can the H200 replace the H100 in existing servers?
Not always. The H100 SXM and H200 SXM use the same HGX baseboard form factor, but the H200 NVL PCIe card needs a PCIe Gen5 server with sufficient power delivery. Check compatibility with your server platform before buying.
Which GPU is best for running LLMs in production?
For large models with long contexts, the H200. The 141GB memory pool fits larger models and bigger batches, which means higher serving throughput and lower cost per token.
Does Eton Technology supply both H100 and H200?
Yes. Eton Technology supplies new H100 and H200 GPUs in PCIe and HGX form factors, plus compatible HPE, Dell and Supermicro servers, with UK next-day shipping available.
Get in Touch
Eton Technology Solutions is a UK-based supplier of NVIDIA enterprise GPU infrastructure, stocking H100 and H200 GPUs, HGX baseboards and compatible server platforms from HPE, Dell and Supermicro.
Browse the NVIDIA GPU collection or contact the team for current pricing, stock and lead times.
Phone: (+44) 333 999 7768 | Email: info@etontechnology.com
