ASICAWS

AWS Trainium2

Curated Aggregate·2024-11-016 workloads · 6 records

MyAI RatingNot scoredInsufficient comparison evidence

VRAM

96 GB

TDP

500 W

MSRP

$21k

Perf/W

0.18 tok/s/W

Cost/1K tok

$0.0024/k

Tested

2024-11-01

Quick answer

How many tokens per second does AWS Trainium2 produce on Llama 3 70B Q4?

AWS Trainium2 has a source-attributed result of 58.0 tok/s on Llama 3 70B Q4 (batch 1, 8192-token context, Q4_K_M; runtime not documented). This is a reference report, not an independently verified lab result. A 70B Q4_K_M artifact needs roughly 40GB for weights alone; smaller memory configurations require explicit offload or model splitting and do not establish full-GPU residency.

Verdict

AWS Trainium2 with 96GB VRAM at 500W TDP, scored across 6 workloads with 6 benchmark records.

Reference workload

Llama 3 70B Q4

58 tok/s

Quantization

Q4_K_M

8K context · batch 1

LLM Inference Performance

Benchmarks (6 workloads)

Workload	Score	Quant	Context	σ	Status	Tested
Llama 3 70B Q8 llm	92.0tok/s	INT8	8K	,	Curated Aggregate	2024-11-01
Llama 3 8B FP16 llm	320.0tok/s	FP16	8K	,	Curated Aggregate	2025-12-29
Llama 3 8B Q4 llm	215.0tok/s	Q4_K_M	8K	,	Curated Aggregate	2026-03-15
Llama 3 70B Q4 llm	58.0tok/s	Q4_K_M	8K	,	Curated Aggregate	2026-03-28
Mistral 7B llm	280.0tok/s	INT8	4K	,	Curated Aggregate	2024-12-04
Qwen 2.5 14B llm	158.0tok/s	INT8	8K	,	Curated Aggregate	2025-01-20

MyAI Score: not scored

There is insufficient comparable evidence to score AWS Trainium2. A score needs results for multiple devices with matching workload, runtime version, quantization, context and batch size.

Workload Fit

Planning estimates for 96GB: weights plus at least 2GB or 10% overhead. Actual KV cache depends on model, context and cache format; confirm the artifact before buying.

FP16

7-8B

Excellent

13-14B

Excellent

32B

Excellent

70B

Excellent

Good

Won't fit

Reported batch-one examples

Llama 3 70B Q458 tok/s

Llama 3 8B Q4215 tok/s

Llama 3 8B FP16320 tok/s

Source

Two-chip pod.

View source How we benchmark →

Public Trust Layer

Trust score

6/10

MyAI rating

Not scored

Runs

Not documented

Freshness

Aging

Source-linked row with explicit verification status.

Record date: 2026-03-28; 165 days old.

Open primary source

Enterprise buying note

Where to buy

Reseller compare

Retailer we'd check first

Amazon search plus reseller quotes

For datacenter and accelerator parts, Amazon is useful for spotting live listings, accessories, or used pulls, but serious procurement usually happens through integrators, brokers, or cloud partners.

Enterprise procurement, not retail

This silicon is typically acquired through an authorized OEM partner, system integrator, or hyperscaler reseller. For on-demand access, compare hourly rates at RunPod, Vast.ai, Lambda Labs, or your existing cloud provider before committing to capital expenditure.

+Use $21k only as a rough anchor. Enterprise street pricing moves with supply, warranty, and included accessories.
+Confirm cooling, power delivery, and return terms before you purchase. These parts often ship without consumer-friendly safeguards.
+If this is for production, compare against authorized reseller quotes before you commit.

There is insufficient comparable evidence for a purchase recommendation. Check model fit, software support and a current seller quote.

Affiliate note: we do not have a device-level ASIN yet, so this opens tagged Amazon search results for the exact product name.

Same Tier

Groq LPU (per chip)

0.23GB

Cerebras WSE-3 (CS-3)