GLM-4.5 hardware requirements
Needs at least 4 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 32 x H100 SXM5 80GB at $58.56/hr or 48 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $24.00/hr.
Top options for GLM-4.5
The strongest part each vendor offers for GLM-4.5 at BF16 with 32,768 tokens per user, rentable parts first. Each card is one replica, the fewest accelerators that hold the model, and the line it draws on the graph below. Select a card to plot it.
Tokens per second as users grow
Each line is one replica, the fewest accelerators that hold the model, serving more users at once. It climbs while memory bandwidth is the limit, flattens once compute is, and stops where memory runs out at 32,768 tokens per user. Hover a line to bring it forward.
8 x GB200 NVL72 GPU 186GB: peaks at 1.3k tokens per second when memory fills at 57 users.4 x Instinct MI355X OAM: peaks at 554 tokens per second when memory fills at 31 users.8 x Ascend 950DT: peaks at 413 tokens per second when memory fills at 31 users.8 x Gaudi 2 96GB: peaks at 138 tokens per second when memory fills at 1 users.4 x TPU v7 192GB: peaks at 236 tokens per second when memory fills at 1 users.
Memory fills at: 8 x GB200 NVL72 GPU 186GB 57 users, 4 x Instinct MI355X OAM 31 users, 8 x Ascend 950DT 31 users, 8 x Gaudi 2 96GB 1 users, 4 x TPU v7 192GB 1 users. Roofline estimates at BF16.
Bare minimum
- Tokens/s
- 325
- Per hour
- $23.96
- Per million tokens
- $20.47
- Memory used
- 65%
- Tokens/s
- 137
- Per hour
- $4.00
- Per million tokens
- $8.11
- Memory used
- 98%
- Tokens/s
- 395
- Per hour
- $8.00
- Per million tokens
- $5.63
- Memory used
- 76%
Configurations
Every datacenter GPU in the catalog, sized for these settings. Change precision, context and concurrency and the whole page follows. Each point is one replica: the smallest tensor-parallel group of that GPU that holds the model.
8 x Instinct MI355X OAM: 1.1k tokens/s per replica at $47.92/hr.8 x GB300 NVL72 GPU 288GB: 1.1k tokens/s per replica at $99.36/hr.8 x GB200 NVL72 GPU 186GB: 1.1k tokens/s per replica at $84.00/hr.8 x Instinct MI350X OAM: 1.0k tokens/s per replica at $49.28/hr.8 x B300 SXM 262GB: 1.0k tokens/s per replica at $39.92/hr.8 x B200 SXM 180GB: 1.0k tokens/s per replica at $32.72/hr.8 x Instinct MI325X OAM: 726 tokens/s per replica at $30.40/hr.8 x Instinct MI300X 192GB: 662 tokens/s per replica at $19.12/hr.16 x H200 SXM 141GB: 1.0k tokens/s per replica at $39.68/hr.16 x H100 SXM5 80GB: 771 tokens/s per replica at $29.28/hr.16 x H200 NVL 141GB: 959 tokens/s per replica at $57.82/hr.16 x H100 NVL 94GB: 829 tokens/s per replica at $51.04/hr.16 x H100 PCIe 80GB: 483 tokens/s per replica at $40.00/hr.16 x RTX PRO 6000 Blackwell Workstation Edition: 408 tokens/s per replica at $28.80/hr.16 x RTX PRO 6000 Blackwell Server Edition: 372 tokens/s per replica at $11.04/hr.16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition: 395 tokens/s per replica at $8.00/hr.16 x Gaudi 2 96GB: 491 tokens/s per replica at $17.60/hr.16 x A100 PCIe 80GB: 377 tokens/s per replica at $21.60/hr.16 x A100 SXM4 80GB: 390 tokens/s per replica at $22.40/hr.
Priced configurations only; 27 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.
| Rent | |||
|---|---|---|---|
| 4 | 1.5k tok/s | Specs → | |
| 8 | 2.5k tok/s | Specs → | |
TPU 8t announced | 8 | ~930 tok/s | Specs → |
TPU 8i announced | 8 | ~1.1k tok/s | Specs → |
| 8 | 1.1k tok/s | Rent → | |
| 8 | 1.1k tok/s | Rent → | |
| 8 | 1.1k tok/s | Rent → | |
| 8 | 1.0k tok/s | Rent → | |
| 8 | 974 tok/s | Specs → | |
Instinct MI430X OAM announced | 4 | ~1.1k tok/s | Specs → |
| 8 | 1.0k tok/s | Rent → | |
| 8 | 1.0k tok/s | Rent → | |
| 8 | 1.5k tok/s | Specs → | |
| 8 | 969 tok/s | Specs → | |
| 8 | ~1.5k tok/s | Specs → | |
| 8 | 726 tok/s | Rent → | |
| 8 | 662 tok/s | Rent → | |
| 8 | 894 tok/s | Specs → | |
H200 SXM 141GB 2 nodes | 16 | 1.0k tok/s | Rent → |
H100 SXM5 80GB 2 nodes | 16 | 771 tok/s | Rent → |
Data Center GPU Max 1550 128GB 2 nodes | 16 | 729 tok/s | Specs → |
H200 NVL 141GB 2 nodes | 16 | ~959 tok/s | Rent → |
H100 NVL 94GB 2 nodes | 16 | ~829 tok/s | Rent → |
| 16 | 708 tok/s | Specs → | |
H100 PCIe 80GB 2 nodes | 16 | 483 tok/s | Rent → |
H800 PCIe 80GB 2 nodes | 16 | 483 tok/s | Specs → |
| 16 | 734 tok/s | Specs → | |
| 16 | 408 tok/s | Rent → | |
| 16 | 373 tok/s | Specs → | |
| 16 | ~373 tok/s | Specs → | |
| 16 | 372 tok/s | Rent → | |
Gaudi 3 128GB 2 nodes | 16 | 921 tok/s | Specs → |
TPU v5p 95GB 2 nodes | 16 | 544 tok/s | Specs → |
| 16 | 395 tok/s | Rent → | |
| 16 | ~146 tok/s | Specs → | |
Gaudi 2 96GB 2 nodes | 16 | 491 tok/s | Rent → |
| 16 | 324 tok/s | Specs → | |
MI250X 128GB 2 nodes | 16 | 566 tok/s | Specs → |
Instinct MI250 128GB 2 nodes | 16 | 552 tok/s | Specs → |
A800 PCIe 80GB 2 nodes | 16 | 377 tok/s | Specs → |
A800 SXM4 2 nodes | 16 | 390 tok/s | Specs → |
A100 PCIe 80GB 2 nodes | 16 | 377 tok/s | Rent → |
A100 SXM4 80GB 2 nodes | 16 | 390 tok/s | Rent → |
H20 141GB HBM3e 2 nodes | 16 | 365 tok/s | Specs → |
H20 96GB 2 nodes | 16 | 365 tok/s | Specs → |
| 16 | 96.8 tok/s | Specs → |
By vendor
At BF16 the NVIDIA pick is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 395 tokens/s for $8.00/hr; the AMD pick is 8 x Instinct MI300X 192GB at about 662 tokens/s for $19.12/hr. Per rental dollar NVIDIA delivers 1.42x the tokens of AMD here.
NVIDIA: 49.3 tokens/s per dollar, 82.2 tokens/s per kW (16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition).AMD: 34.6 tokens/s per dollar, 110 tokens/s per kW (8 x Instinct MI300X 192GB).Huawei: no live price, no power figure (16 x Ascend 950DT).Intel: 27.9 tokens/s per dollar, 51.2 tokens/s per kW (16 x Gaudi 2 96GB).Google: no live price, no power figure (8 x TPU v7 192GB).Other: no live price, no power figure (16 x MTT S5000).
Fleet what-if
Compare whole fleets at these settings: a thousand of one part against a hundred of another. Each fleet splits into replicas of its tensor-parallel width; GPUs left over sit idle.
| Fleet | Tokens/s |
|---|---|
1,008 x H200 SXM 141GB | 64k tok/s |
104 x GB200 NVL72 GPU 186GB | 14k tok/s |
1,008 x H200 SXM 141GB: 64k tokens/s, $2,500/hr, 706 kW.104 x GB200 NVL72 GPU 186GB: 14k tokens/s, $1,092/hr, no power figure.
H200 SXM 141GB: 64k tokens/s at 1008 GPUs.GB200 NVL72 GPU 186GB: 133k tokens/s at 1008 GPUs.
Size GLM-4.5 for a tokens per second target
The calculator starts from your traffic instead of a fleet: set a target rate and read the replica count for every GPU.
Frequently asked questions
How much GPU memory does GLM-4.5 need?
The weights take 717 GB at the native BF16 precision (717 GB at BF16, 358 GB at FP8, 202 GB at INT4). Each concurrent stream adds 368 KB of KV cache per token: 1.5 GB at 4,096 tokens and 12.3 GB at 32,768 tokens.
What is the bare minimum to run GLM-4.5?
4 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00 per hour.
How many H100s do you need to run GLM-4.5?
16 H100s at BF16 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 16 H100s, and 1,000 tokens per second takes 32 H100s across 2 replicas.
How much does GLM-4.5 cost per million tokens on an H100?
About $10.55 per million output tokens at BF16, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.83 per GPU-hour. The cheapest part per token is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $5.63.
What precision does GLM-4.5 ship in?
The published checkpoint is BF16. FP8 and INT4 figures describe post-training quantisations that halve and quarter the footprint at a small accuracy cost.
All figures are roofline estimates from the model's config.json and each GPU's published memory, bandwidth and tensor-core peak, with fixed efficiency factors. Prices are the cheapest live on-demand listing per GPU when the site has one. They are a planning floor; a tuned serving stack can do better. Active parameters per the Z.ai model card.