Will AI accelerators improve in FLOPs/watt by 100x of an NVidia H100 by 2033?

100Ṁ3383

2033

86%

chance

ALL

Compared to an H100, will tensor TFLOPs/Watt improve by 100x by 2033? AI accelerators in scope for this question must be deployed significantly - with at least 100k units or $100M (in 2024 dollars) in production, and have published perf/watt numbers.

This market will count peak FLOPs/watt at k bits of precision, adjusted by a factor of 2^(1 - 32/k). That is, 16 bit precision counts 1/4 as much as 32 bit, which counts 1/4 as much as 64 bit precision.

Technology

Technical AI Timelines

AI Impacts

AI Safety

Get

1,000

to start trading!

People are also trading

Will an AI model use more than 1e28 FLOPS in training before 2026?

8% chance

What will be the maximum achievable flop utilization on the next generation of Nvidia server chips?

Will there be an announcement of a model with a training compute of over 1e30 FLOPs by the end of 2025?

5% chance

When will a US government AI run overtake private AI compute by FLOP?

Will any chip maker other than NVIDIA, Intel, and AMD create accelerators for Deep Learning and be profitable by 2030?

90% chance

Will China develop domestic GPU with comparable performance to NVIDIA H100 by 2026?

86% chance

Will a machine learning training run exceed 10^26 FLOP in China before 2026?

52% chance

Will a machine learning training run exceed 10^27 FLOP in China before 2030?

79% chance

Will a machine learning training run exceed 10^25 FLOP in China before 2027?

82% chance

At least one of the most powerful neural nets at end of 2030 will be trained using 10^27 FLOPs

Sort by:

Two questions:

What's the TFLOPs/Watt of an H100 today?
What floating point precision is this about? Double-precision?

GPUs usually don't do higher than single precision (32 bit), It's considered unnecessary for graphics or neural networs. But good question, modern ai systems do often use even lower precision.

Well H100 supports double precision 🤷‍♂️

In practice modern AI is usually much less than 32 bit; 16 bit is the large end. Honestly I'd recommend counting any of 8, 16, or 32 bit flop/s/w performance for this question.

This market will use an adjusted flop/s/w by adjusting k-bit precision by a factor of 2^(1 - 32/k). That is, 16 bit precision counts 1/4 as much as 32 bit, which counts 1/4 as much as 64 bit precision.

Whichever precision gets the highest score is counted as the FLOP/s/W for the chip.