AIMPACT News, May 16 (UTC+8), Google disclosed the architecture details of the eighth-generation TPU (TPU 8t) rack-level connection to the Virgo network. The network uses high-radix switches and a flat two-layer non-blocking topology, increasing data center network bandwidth by four times compared to the previous generation, with a single structure capable of connecting over 134k TPU 8t chips, providing 47 Pb/s of non-blocking bidirectional bandwidth and nearly linear scaling performance of over 1.7K ExaFlops. The TPU 8t itself adopts a 3D torus topology, with a single super pod scalable to 9,600 chips, and supports expansion to over one million chips via JAX and Pathways. Key technologies include SparseCore accelerators, VPU/MXU overlap and balanced scaling, native FP4 support, and integrated Arm-based Axion CPUs to eliminate host bottlenecks. This design addresses the evolution of AI models from dense large language models to large-scale mixture-of-experts models and inference-intensive architectures. (Source: InFoQ)

GOOGLX0.98%

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.

10 Likes

Reward
10
10
2
Share

Comment

Add a comment

Semi-MeltedIceCream

· 3h ago

High-base-number switches are not cheap, right? How long does it take for cloud vendors to recoup their costs with this pricing?

View OriginalReply0

HypeVaccinated

· 7h ago

The supercomputing center was just completed when it was beaten by Google Cloud; who can keep up with this iteration speed?

View OriginalReply0

LeverageWhisperer

· 8h ago

SparseCore and native support for FP4 have some advantages, but the inference cost is also being reduced.

View OriginalReply0

GateUser-6da8ed4c

· 8h ago

Arm Axion CPU integrated in, so there's no need to connect an external host anymore, the design is quite clever.

View OriginalReply0

Stop-LossLineForTheEveningGlow

· 9h ago

JAX/Pathways directly scaled to millions, Google is forcing other frameworks to keep up.

View OriginalReply0

GateUser-7919e6b9

· 9h ago

134k-chip single-structure—how do we split this fault domain? Curious how operations and maintenance handle it.

View OriginalReply0

PaperSculptureOctopus

· 9h ago

Wait, is 8t the eighth generation? I haven't even tried the hot TPU v5 yet.

View OriginalReply0

GateUser-9d67589f

· 9h ago

3D torus topology + two-layer non-blocking, the network has indeed been heavily optimized.

View OriginalReply0

SpiralSeaSalt

· 9h ago

A million-chip cluster... Is this going to train Skynet?

View OriginalReply0

Post-RainCandlestick

· 9h ago

Google has really outdone itself with the TPU this time; 47 Pb/s—what a concept. My home broadband would faint from crying in the bathroom.

View OriginalReply0

Trending Topics
View More
#
StockTradingChallengeUpTo17000U
16M Popularity
#
TrumpBacksCFTCAuthorityOverPredictionMarkets
825.76K Popularity
#
GatePredictionMarketAddsSmartMoneyTracking
12.48M Popularity
#
MicronMarketCapBreaks1Trillion
40.85K Popularity
#
TradeCFDWinGold
3.08M Popularity

Pinned

Sitemap

Google releases the eighth-generation TPU 8t rack-scale network architecture details

Trending Topics

StockTradingChallengeUpTo17000U

TrumpBacksCFTCAuthorityOverPredictionMarkets

GatePredictionMarketAddsSmartMoneyTracking

MicronMarketCapBreaks1Trillion

TradeCFDWinGold

Pinned