mirror of
https://github.com/hanzo-apps/hanzo.ai.git
synced 2026-10-29 07:20:43 +00:00
zeekay
8b354c25ad
Savings (components/sales/Savings, lib/savings): the tokens a team uses a month, by model or by spend, at the providers' list rates, on Hanzo Cloud moved to an Enso SKU plus the members on Team, and on the reader's own GPUs on all month. Every rate is read when the page opens: the price list (GET /v1/pricing, its on-demand GPU hour too), the plan list (Team's member month) and the model list (what Hanzo bills a provider's model). The volumes, people and GPUs are the reader's, and the page says it opens with an example. Each figure says what it assumes; the GPU figure leaves the agreement out and says so, with where it meets list and that sizing is done with the customer. /calculator carries the same calculator. Your own GPUs: Zen 6's prefill as Hanzo measured it on one DGX Spark (lib/data/zen6-spark.json, scripts/zen6.py, from the enterprise-story branch) and what a Zen 6 Max rack is made of, from NVIDIA's pages (Rack, with Zen 6 Max's page). That branch's Cloud against Enterprise board is left off /benchmarks: it prints a 96.0% the page's gate refuses. Proof shows six of the data room's case studies. Partners are the programmes that exist: affiliates, White-label and the agency's plans as the plan list sells them, and the clouds Hanzo deploys on, captioned as vendors, not partners. No company is called a partner: none is on record.
30 lines
1.8 KiB
TypeScript
30 lines
1.8 KiB
TypeScript
/**
|
|
* /enterprise's own GPUs: what Hanzo measured on one DGX Spark, and what a rack
|
|
* for Zen 6 Max is made of.
|
|
*
|
|
* The prefill sweep is read from lib/data/zen6-spark.json (`python3
|
|
* scripts/zen6.py`). A server's memory, network and storage are NVIDIA's own
|
|
* figures, each with the page it was read from; the planning around them is
|
|
* Hanzo's, and final capacity comes from qualifying the customer's workload.
|
|
*/
|
|
import zen from '@/lib/data/zen6-spark.json' with { type: 'json' }
|
|
|
|
/** The day each vendor page below was read and its figure checked. */
|
|
export const READ = '2026-10-12'
|
|
|
|
/** Where every vendor figure comes from. */
|
|
export const SOURCES = {
|
|
spark: { label: 'NVIDIA DGX Spark: GB10, 273 GB/s memory bandwidth', url: 'https://www.nvidia.com/en-us/products/workstations/dgx-spark/' },
|
|
b300: { label: 'NVIDIA DGX B300 user guide: 8 x 288 GB GPU memory', url: 'https://docs.nvidia.com/dgx/dgxb300-user-guide/introduction-to-dgxb300.html' },
|
|
b200: { label: 'NVIDIA DGX B200 user guide: 1,440 GB total GPU memory', url: 'https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html' },
|
|
} as const
|
|
|
|
/** The measured prefill on the Spark, by prompt size (cold, quiet box), and the box and model it was measured on. */
|
|
export const SPARK = { box: zen.box, model: zen.model, measured: zen.sweep.measured, prefill: zen.sweep.prefill }
|
|
|
|
/** Zen 6 Max: what a rack for it is made of. Memory per server is NVIDIA's; the rest is planning. */
|
|
export const SERVERS = [
|
|
{ name: 'DGX B300', gpus: '8 x B300', memory: '8 x 288 GB, 2,304 GB', network: '8 x 800 Gb/s InfiniBand or Ethernet', nvme: '8 x 3.84 TB', source: 'b300' },
|
|
{ name: 'DGX B200', gpus: '8 x B200', memory: '1,440 GB', network: '8 x ConnectX-7, up to 400 Gb/s', nvme: '8 x 3.84 TB', source: 'b200' },
|
|
] as const
|