Groq told the LPU speed story. Cerebras tells the wafer-scale story: giant chips historically used for training clusters, now also offered as Cerebras Inference (cloud API) for supported LLMs. Same engineering job as Together/Groq—rent fast tokens—different silicon and catalog. OpenRouter may list Cerebras as an upstream; you can also call Cerebras direct with an OpenAI-compatible client (Vol. 18 OpenAI SDK).
Module close: Perplexity (search product + Sonar API), which is not an inference chip. Hardware context: Vol. 18.3 GPU vs specialist ASICs—do not assume CUDA mental models transfer 1:1.
Learning Objectives
By the end of this lesson, students should be able to:
- Describe Cerebras as wafer-scale silicon + inference (and historically training) cloud, not a chatbot brand.
- Call Cerebras Inference with an OpenAI-compatible or official client.
- Compare Cerebras vs Groq vs Together vs GPU self-host qualitatively.
- Expect a curated model list (ports), not every Hub checkpoint.
- Measure TTFT/tok-s on your prompts; refuse fake wafer benchmarks.
- Know when GPU clusters (Vol. 18.4) still win for custom training.
Cerebras builds wafer-scale engines (WSE)—an entire silicon wafer as one accelerator—and sells systems plus cloud services. Cerebras Inference exposes supported LLMs over an API (often OpenAI-compatible). Older narrative is training supercomputers; newer student-facing product is fast inference SKUs. Cerebras is not Perplexity, not Groq, not Hugging Face, and not xAI.
Wafer-Scale vs GPU vs LPU
You do not need microarchitecture trivia to buy inference. You need: which models are offered, what the latency/quality trade looks like on your eval set, data residency, and price sheet. WSE marketing emphasizes on-chip memory and very high decode; Groq emphasizes LPU scheduling; Together emphasizes GPU fleets and catalog. All three still serve transformer weights. Custom training of a 70B+ model may still be a GPU-cluster or Cerebras training conversation—separate SKU from Inference API.
Pricing / Strengths / Weaknesses (Qualitative)
| Dimension | Cerebras Inference | GroqCloud | Together / GPU cloud | Self-host GPUs |
|---|---|---|---|---|
| Pricing posture | Per-token inference on WSE-hosted models; enterprise systems priced separately—confirm live cloud table | Per-token LPU | Per-token or instance-hour | Capex/opex GPUs |
| Strengths | Speed-specialist inference; OpenAI-shaped API; distinctive silicon for very large memory stories; training heritage for orgs that need it | Also speed; often easy signup | Broader model/fine-tune catalog | Full control |
| Weaknesses | Curated model list; fewer multimodal/video SKUs; availability/regions vary; not a search engine; speed ≠ closed-model quality | Also curated | Usually not the tok/s headline (measure) | Ops (Vol. 18.4) |
Inference API
- Chat completions
- Pin model id
- Same FastAPI wrap
Via OpenRouter
- Prototype hop
- Still log upstream
- Direct for SLA later
Training systems
- Not this lecture’s default buy
- Different procurement
- Vol. 18.4 cluster mindset
Pick Cerebras Inference when
- Needed model is on their list
- Interactive latency matters
- You already like OpenAI clients
- A/B vs Groq is in the test plan
Look elsewhere when
- You need arbitrary HF checkpoints tomorrow
- Diffusion/video (Replicate)
- Search-with-citations (Perplexity)
- Air-gapped GPU-only policy
Python: OpenAI-Compatible Cerebras Client
Official SDK and base URLs evolve. Pattern: env key, config model id, OpenAI client. Confirm current base_url in Cerebras cloud docs before a spike.
Related Lectures
| Lecture | Role |
|---|---|
| Groq / Together / OpenRouter | Inference peers / gateway |
| OpenAI SDK / GPU / Tier 5 clusters | Client + hardware contrast |
| Latency | Why you A/B silicon |
| Llama family | Typical hosted weights |
| Perplexity | Next: search product, not a chip |
“Cerebras is only a training supercomputer; there is no API.” Inference cloud exists—confirm current product. Second: wafer-scale means every HF model runs. Third: Cerebras is Groq with a different logo. Fourth: tok/s from a keynote is your p95. Fifth: Cerebras Inference is Perplexity Sonar. Sixth: buying Inference includes unlimited on-prem WSE hardware.
Knowledge Check
- Short Answer: What silicon approach is Cerebras known for? Answer: Wafer-scale engines (WSE) — an accelerator the size of a wafer.
- True/False: Cerebras Inference is the same product as Perplexity.ai search. Answer: False.
- Multiple Choice: Client pattern is typically: (a) OpenAI-compatible API, (b) only Suno jobs, (c) only ComfyUI. Answer: (a).
- Short Answer: Name one qualitative peer for decode speed. Answer: Groq (or Together/GPU — Groq is the closest specialist).
- True/False: Every Hub checkpoint is guaranteed on Cerebras Inference. Answer: False — curated/ported list.
- Multiple Choice: Custom 70B training is usually: (a) a separate training/cluster decision, (b) a single chat.completions flag, (c) Udio. Answer: (a).
- Short Answer: Which Vol. 18 lecture covers GPUs as the default accelerator? Answer: GPU (18.3).
- True/False: Invent wafer tok/s benchmarks for the exam. Answer: False.
- Multiple Choice: OpenRouter may: (a) list Cerebras as an upstream, (b) replace silicon, (c) train WSE in the browser. Answer: (a).
- Short Answer: Next lecture in 22.5? Answer: Perplexity.
Key Takeaways
- Cerebras = wafer-scale silicon; Inference API is the student-facing buy.
- A/B vs Groq/Together on your prompts; curated catalogs on both specialists.
- OpenAI-shaped clients; pin model ids; Vol. 18 wrap.
- Training systems ≠ Inference SKU.
- Next: Perplexity—search-grounded answers, not chips.
Lab: Architecture memo: voice-agent LLM hop—Groq vs Cerebras vs Together. Students must name eval metrics (TTFT, task success, cost sheet lookup) and one reason they might still pick a frontier model. Optional live call if keys exist. Grade: no fake tok/s tables, WSE vs LPU vs GPU distinguished correctly.
Recap: Cerebras is wafer-scale inference (and historically training), not a search brand. Last 22.5 vendor: Perplexity.