Compute ratio for SOTA model training workloads by hardware architecture by EOY 2028 | Manifold

Compute ratio for SOTA model training workloads by hardware architecture by EOY 2028

1

Ṁ100Ṁ30

2028

1H

6H

1D

1W

1M

ALL

20%

GPU/CPU

20%

FPGA (Reprogrammable ASIC)

59%

Custom ASIC

The ratio of compute (per FLOP) used for the flagship language models of the 10 leading companies/organizations at market close.

Epistemic standards: company/lab announcements, consensus reporting

(Hopefully) forward-compatible definitions:

GPU: a general-purpose CPU specialized to handle embarassingly parallel workload tasks

FPGA: an ASIC that implements logic gates through a reprogrammable medium (currently, LUTs and routing fabric), meaning the logic can be cost-effectively modified after finishing manufacturing, at a price at least one OOM less than manufacturing a new one.

ASIC (excluding FPGAs): a non-reprogrammable IC that directly implements logic that would normally be handled at the programming-language layer in a CPU directly in the digital logic.

Market context

Get

1,000

to start trading!

Sort by:

What if there aren't ten leading companies or orgs by then?

People are also trading

Most training run compute greater than 2e27 FLOP by EOY 2026?

What will be true of the SOTA AI on the FrontierMath benchmark, before 2027?

SOTA AI at EOY 2026 a reasoning model?

GPT-4 performance and compute efficiency from a simple architecture before 2026

What will be true of the SOTA AI on the FrontierMath benchmark, before 2028?

Will any of CPUs/GPU created by hackerfab generate at least 10^18 operations of compute by EOY2028?

BIG-bench accuracy 75% #4: Will SOTA for a single model on BIG-bench pass 75% by the start of 2027?

BIG-bench accuracy 75% #5: Will SOTA for a single model on BIG-bench pass 75% by the start of 2028?

[Carlini questions] Cost of single SOTA AI system training run by 2030

Will the transformer architecture be replaced in SOTA LLMs by 2028?

Related questions

Most training run compute greater than 2e27 FLOP by EOY 2026?

What will be true of the SOTA AI on the FrontierMath benchmark, before 2027?

SOTA AI at EOY 2026 a reasoning model?

GPT-4 performance and compute efficiency from a simple architecture before 2026

What will be true of the SOTA AI on the FrontierMath benchmark, before 2028?

Will any of CPUs/GPU created by hackerfab generate at least 10^18 operations of compute by EOY2028?

BIG-bench accuracy 75% #4: Will SOTA for a single model on BIG-bench pass 75% by the start of 2027?

BIG-bench accuracy 75% #5: Will SOTA for a single model on BIG-bench pass 75% by the start of 2028?

[Carlini questions] Cost of single SOTA AI system training run by 2030

Will the transformer architecture be replaced in SOTA LLMs by 2028?

© Manifold Markets, Inc.•Terms•Privacy