Skip to content
Updated minutes ago
Meta

Llama 3 8B

Meta · dense · 8B parameters · 8,192 context

Quality
63.0

Parameters

8B

Context Window

8K tokens

Architecture

Dense

Best GPU

A30

Cheapest API

$0.20/M

Quality Score

63/100

Intelligence Brief

Llama 3 8B is a 8B parameter DENSE model from Meta, featuring Grouped Query Attention (GQA) with 32 layers and 4,096 hidden dimensions. With a 8,192 token context window, it supports structured output, code, math, multilingual. On standardized benchmarks, it achieves MMLU 68.4, HumanEval 37, GSM8K 73. The most cost-effective API deployment is via together at $0.20/M output tokens. For self-hosted inference, A30 delivers optimal throughput at $332/month.

Architecture Details

TypeDENSE
Total Parameters8B
Active Parameters8B
Layers32
Hidden Dimension4,096
Attention Heads32
KV Heads8
Head Dimension128
Vocab Size128,256

Memory Requirements

BF16 Weights

16.0 GB

FP8 Weights

8.0 GB

INT4 Weights

4.0 GB

KV-Cache per Token131072 bytes
Activation Estimate1.00 GB

GPU Compatibility Matrix

Llama 3 8B is compatible with 90% of GPU configurations across 41 GPUs at 3 precision levels.

BF16 (Full)
FP8 (Half)
INT4 (Quarter)
Blackwell(7 GPUs)
B200 NVL (pair)360GB
B300288GB
B100 SXM192GB
GB200 NVL72 (per GPU)192GB
Hopper(7 GPUs)
H100 NVL 94GB (per GPU pair)188GB
H200 SXM141GB
H2096GB
GH20096GB
Ada Lovelace(11 GPUs)
L40S48GB
L4048GB
RTX 6000 Ada48GB
L2048GB
Ampere(16 GPUs)
A100 80GB SXM80GB
A100 80GB PCIe80GB
A1664GB
RTX A600048GB
Legend:No fitVery tightTightModerateGoodExcellent

GPU Recommendations

A30optimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

314.9 tok/s

Latency (ITL)

3.2ms

Est. TTFT

1ms

Cost/Month

$332

Cost/M Tokens

$0.40

Use this config →
RTX 4090optimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

340.2 tok/s

Latency (ITL)

2.9ms

Est. TTFT

1ms

Cost/Month

$370

Cost/M Tokens

$0.41

Use this config →
RTX 3090optimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

315.9 tok/s

Latency (ITL)

3.2ms

Est. TTFT

1ms

Cost/Month

$180

Cost/M Tokens

$0.22

Use this config →

Deployment Options

API

API Deployment

together

$0.20/M

output tokens

Self-Hosted

Single GPU

A30

$332/mo

Min VRAM: 8 GB

Scale

Multi-GPU

RTX 3060 x2

190.4 tok/s

TP· $114/mo

API Pricing Comparison

ProviderInput $/MOutput $/MBadges
together$0.20$0.20
Cheapest

Cost Analysis

ProviderInput $/MOutput $/M~Monthly Cost
togetherBest Value$0.20$0.20$2

Cost per 1,000 Requests

Short (500 tok)

$0.14

via together

Medium (2K tok)

$0.56

via together

Long (8K tok)

$2.00

via together

Performance Estimates

Throughput by GPU

A30
314.9 tok/s
RTX 4090
340.2 tok/s
RTX 3090
315.9 tok/s

VRAM Breakdown (A30, BF16)

Weights
Act
Weights 16.0 GBKV-Cache 2.1 GBActivations 8.0 GBOverhead 1.3 GB

Precision Impact

bf16

16.0 GB

weights/GPU

~314.9 tok/s

fp8

8.0 GB

weights/GPU

int4

4.0 GB

weights/GPU

Quality Benchmarks

Average
66th percentile across all models
MMLU
68.4
Below Average (28th pctile)
HumanEval
37.0
Bottom 25% (23th pctile)
GSM8K
73.0
Below Average (34th pctile)
MT-Bench
76.0
Bottom 25% (0th pctile)

Capabilities

Features

Tool Use Vision Code Math Reasoning Multilingual Structured Output

Supported Frameworks

vllmsglangtgitensorrt-llmollama

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy Llama 3 8B

Similar Models

Aya 23 8B

8B params · dense

Quality: 50

from $0.60/M

Lower quality, More expensiveCompare →
Larger context, Higher qualityCompare →

Llama Guard 3 8B

8B params · dense

Quality: 50

from $0.20/M

Larger context, Lower qualityCompare →

Ministral 8B

8B params · dense

Quality: 50

from $0.10/M

Larger context, Lower quality, CheaperCompare →

Minitron 8B

8B params · dense

Quality: 62

from $0.10/M

Frequently Asked Questions

How much VRAM does Llama 3 8B need for inference?

Llama 3 8B requires approximately 16.0 GB of VRAM at BF16 precision, 8.0 GB at FP8, or 4.0 GB at INT4 quantization. Additional VRAM is needed for KV-cache (131072 bytes per token) and activations (~1.00 GB).

What is the best GPU for Llama 3 8B?

The top recommended GPU for Llama 3 8B is the A30 using BF16 precision. It achieves approximately 314.9 tokens/sec at an estimated cost of $332/month ($0.40/M tokens). Score: 100/100.

How much does Llama 3 8B inference cost?

Llama 3 8B API inference starts from $0.20/M input tokens and $0.20/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.