AIFEATURE

Mixed-Precision Training and FP8: How 16-Bit and 8-Bit Formats Cut Compute and Memory Costs

E
EffectStory 編輯部Editorial Team
Published · Updated
Mixed-precision training combines multiple numerical formats in one computation. Moving from 32-bit to 16-bit formats such as BF16, and now to 8-bit FP8, tracks a hardware trend already visible in NVIDIA's Volta Tensor Cores, which deliver 8x the throughput of single-precision pipelines, while half-precision training cuts memory consumption by nearly 2x and a full-precision weight copy protects accuracy.

What Is Mixed-Precision Training?

Mixed precision is the combined use of different numerical precisions within a single computational method, according to NVIDIA's technical documentationCITE:E1. Rather than running every operation at one fixed bit-width, a training method can mix formats — for example carrying out arithmetic at lower precision while keeping certain values at higher precision — within the same computationCITE:E1.

How Do FP16 and BF16 Compare as 16-Bit Formats?

Bfloat16 (BF16) is a 16-bit floating-point format built from one sign bit, eight exponent bits, and seven mantissa bits, Google Cloud explained in an August 24, 2019 blog postCITE:E5. Google Cloud also stated that bfloat16 has a greater dynamic range — that is, more exponent bits — than FP16CITE:E6. Both formats total 16 bits, but the split between exponent and mantissa bits differs, which is why BF16 trades some numeric precision for a wider representable range compared with FP16CITE:E5CITE:E6.

What Is FP8 and How Are E4M3 and E5M2 Encoded?

FP8 is an 8-bit floating-point binary interchange format proposed in a paper published on arXiv on September 12, 2022, by Micikevicius et al., consisting of two encodingsCITE:E7. The first, E4M3, uses a 4-bit exponent and a 3-bit mantissa; the second, E5M2, uses a 5-bit exponent and a 2-bit mantissaCITE:E7. The two encodings mirror the FP16/BF16 trade-off one bit-width down: E4M3 keeps more mantissa bits for precision, while E5M2 keeps more exponent bits for rangeCITE:E7.

FormatSign BitsExponent BitsMantissa BitsTotal Bits
BF1618716
FP8 E4M3shared438
FP8 E5M2shared528

How Do Tensor Cores Accelerate Mixed-Precision Math?

NVIDIA's Volta generation of GPUs introduced Tensor Cores that provide 8x more throughput than single-precision math pipelines, according to NVIDIA's documentationCITE:E4. This hardware-level throughput multiplier is what makes reduced-precision formats like FP16, BF16, and FP8 attractive for training beyond their memory benefits aloneCITE:E4.

How Much Memory Does Lower Precision Save?

Training with half precision reduces the memory consumption of deep learning models by nearly 2x, Micikevicius et al. reported in a paper published on arXiv on October 10, 2017CITE:E3. That memory reduction comes from storing activations, gradients, and weights in a 16-bit format instead of 32-bit during the bulk of training computationCITE:E3.

How Is Accuracy Preserved in Mixed-Precision Training?

Micikevicius et al. recommend maintaining a single-precision (FP32) copy of the weights that accumulates gradients after each optimizer stepCITE:E2. This FP32 master copy acts as the accuracy safeguard underneath the lower-precision forward and backward computation, so the memory and throughput gains from reduced precision do not come at the cost of losing small gradient updatesCITE:E2.

What This Means

Across the evidence, the same pattern repeats at two different bit-widths: NVIDIA's Volta Tensor Cores already delivered an 8x throughput gain for reduced-precision mathCITE:E4, the 2017 paper measured a nearly 2x memory reduction from moving to half precisionCITE:E3, and that same paper's fix for the resulting precision loss — a persistent FP32 weight copy accumulating gradientsCITE:E2 — was published alongside the format comparison between FP16 and BF16 that Google Cloud later detailedCITE:E5CITE:E6. The 2022 FP8 proposal splits precision and range across two distinct encodings, E4M3 and E5M2CITE:E7, following the same exponent-versus-mantissa trade-off already visible in the 16-bit formatsCITE:E5CITE:E6.

📊 Evidence

FAQ

What Is Mixed-Precision Training?

Mixed precision is the combined use of different numerical precisions within a single computational method, according to NVIDIA's technical documentationCITE:E1…

How Do FP16 and BF16 Compare as 16-Bit Formats?

Bfloat16 (BF16) is a 16-bit floating-point format built from one sign bit, eight exponent bits, and seven mantissa bits, Google Cloud explained in an August 24,…

What Is FP8 and How Are E4M3 and E5M2 Encoded?

FP8 is an 8-bit floating-point binary interchange format proposed in a paper published on arXiv on September 12, 2022, by Micikevicius et al.

How Do Tensor Cores Accelerate Mixed-Precision Math?

NVIDIA's Volta generation of GPUs introduced Tensor Cores that provide 8x more throughput than single-precision math pipelines, according to NVIDIA's documentat…

📎 Sources

  1. docs.nvidia.com
  2. arxiv.org
  3. cloud.google.com
  4. arxiv.org

Related data

Author's TakeEffectStory 編輯部

The jump from 16-bit to 8-bit formats reads less like a leap and more like the same trade-off repeating one bit-width down: BF16 already sacrificed mantissa bits for exponent range against FP16, and FP8's E4M3-versus-E5M2 split is that same exponent-versus-mantissa choice made explicit as two separate encodings. What stands out is that the throughput case for going lower precision was already made at the hardware level — Volta's Tensor Cores hit 8x over single-precision pipelines before FP8 was even proposed — so the format work is catching up to capability that existed years earlier. The open question this history raises is whether FP8 training will need the same accuracy backstop the 2017 paper relied on for FP16: a persistent FP32 master weight copy accumulating gradients after every optimizer step. Whether that safeguard, or some FP8-specific equivalent, gets adopted in practice is the detail worth watching next.

E
EffectStory 編輯部Editorial Team

Related

BRIEF

CNA Launches Taiwan's First News MCP Tool, AskCNA, Priced at NT$200 a Month

Central News Agency (中央社) launched CNA MCP on August 31, 2026, Taiwan's first news tool built on Anthropic's Model Context Protocol (released November 2024), letting AI agents such as Claude, ChatGPT, and Grok retrieve and cite its archives in real time. The tool integrates nearly 5 million newswire stories, 3.5 million photos, and open data from about 150 government agencies, priced at NT$200 a month with an early-bird bonus-quota plan, and received funding from Google Taiwan's nDX Digital Innovation Grant Program.

EffectStory 編輯部 ·
BRIEF

Sony Music and Warner Chappell Sue Anthropic Over Alleged 'Brazen Campaign' of Copyright Theft

Sony Music Publishing and Warner Chappell, joined by other music publishers, sued Anthropic and co-founders Dario Amodei and Benjamin Mann in the U.S. District Court for the Northern District of California, alleging illegal torrenting, scraping, and downloading of copyrighted lyrics and sheet music. The publishers seek up to $150,000 per work and $25,000 per instance of stripped copyright data, a total that could reach several billion dollars. The filing follows Anthropic's earlier $1.5 billion settlement in the Bartz case.

EffectStory 編輯部 ·
BRIEF

Why Anthropic Turned to Nscale and Lambda for $45B and $35B GPU Compute Deals

Anthropic has assembled compute capacity across at least four NVIDIA-linked providers: a $35 billion contract with Lambda tied to a Hut 8-built Texas data center, a $45 billion, six-year deal with Nscale for a West Virginia campus running NVIDIA Vera Rubin systems, a $10 billion contract with startup Volta in Norway, and a reported (unconfirmed) tenancy at Riot Platforms' Rockdale, Texas site. NVIDIA sits inside nearly every arrangement — as investor, lessor, or chip supplier.

EffectStory 編輯部 ·