跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators未公布单价No public price

Google TPU 8tx(机密推理)Google TPU 8tx (confidential inference)

未公布单价No public price未公布单价No public price · 官方页未给公开单价;这类产品通常按整机、云实例或询价成交,不拿经销商价冒充官方价。No public unit price on the official page; these usually sell as systems, cloud instances or on request, so dealer quotes are not presented as official.

回答「需要在云上处理敏感数据、又不想把数据搬到另一个集群,该看什么加速器」。8tx 复用 8t 硅片,配第五代 AMD EPYC 主机,把机密计算做进加速器硬件层:数据与模型权重在使用中也被隔离保护;同一系统还能做私有数据微调、蒸馏和 RL。
不主张:2× 延迟改善是相对 TPU 8t 的厂商口径,不是独立实测;「near-zero performance penalty」也是 Google 的说法;机密计算不等于合规结论,法律与合同问题仍需单独确认;没有公开价格与第三方吞吐数据。
Answers which accelerator to look at when sensitive data must be processed in the cloud without moving it to another cluster. 8tx reuses 8t silicon with 5th-gen AMD EPYC hosts and puts confidential computing into the accelerator hardware layer: data and model weights stay isolated even in use; the same system can also do private-data fine-tuning, distillation and RL.
It does not claim: the 2x latency improvement is a vendor figure versus TPU 8t, not an independent measurement; "near-zero performance penalty" is also Google’s wording; confidential compute is not a compliance conclusion — legal and contractual questions remain separate; no public pricing or third-party throughput data.

关键参数Key specs

定位 / role
机密推理 + 小规模训练、微调、蒸馏、RL
主机 / host
第五代 AMD EPYC
延迟 / latency
较 TPU 8t 最多低 2×(Google 口径)
安全 / security
硬件级机密计算(使用中保护数据与权重)
网络 / network
Boardfly 低延迟网络
同代 / family
TPU 8t、TPU 8i、TPU 8tx、TPU 8i Max
状态 / status
2026-11-09 发布

官方原文怎么说What the source says

TPU 8tx provides hardware-level security for sensitive data, while TPU 8i Max handles massive context windows.
1 https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/
Built on the same silicon as TPU 8t and paired with 5th Generation AMD EPYC CPU hosts, TPU 8tx builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use.
2 https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/
delivering up to 2x lower latency compared to TPU 8t with a near-zero performance penalty.
3 https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/
TPU 8tx enables teams to fine-tune models using their private data, distill large models into smaller ones, and train their agents through reinforcement learning
4 https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/

资料与许可Source and licence

资料时间Source checked
2026-11-20
许可 / 类型Licence / kind
vendor-claim
许可说明Licence evidence
https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/
类目Category
加速卡Accelerators
状态Status
已发布 / 未出货Announced
价格口径Price basis
未公布单价No public price · not-published

同类目还有More in this category

下一步Next step

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute