Post

Zyphra发布首个MI355X推理评测：长文本直逼B200

币界网
币界网
币界网消息，Zyphra发布了基于AMD旗舰芯片MI355X的端到端大模型推理实测，测试在真实单节点环境下运行DeepSeek V3.2、Kimi K2.6与GLM-5.1，并对标NVIDIA B200。实测显示，AMD的288GB显存远超B200的180GB，使其在处理超长文本时能更高效地利用缓存，降低硬件成本。然而，AMD在芯片间互联上处于劣势，B200通过NVLink交换机实现900GB/s带宽，而MI355X采用点对点直连，通信效率较低。为弥补这一短板，Zyphra开发了张量序列并行(TSP)与树状注意力(Tree Attention)算法。最终结果显示，尽管在单请求速度上NVIDIA B200仍领先，但在长文本场景下，Zyphra推理栈的吞吐量逐渐逼近B200，证明AMD在长文本生产环境中具备竞争力。
NVDA-0.91%
AMD-3.40%
18:44 · 05/19/26·4.25K views·
Visit source
0
0

Disclaimer: OKX Orbit content is provided for informational purposes only. Learn more

Replies

No comments yet. Be the first to reply!

Related Flash News

Source avatarChainCatcher•2h ago

U.S. stock index futures edged lower, and semiconductor stocks continued to fall

Source avatar币界网•2h ago

Bitget Wallet adds 130 new Xstocks tokenized stocks

Source avatarChainCatcher•4h ago

Institutions: AI data center demand may drive the revaluation of Bitcoin miners, and more than $90 billion in cooperation has been announced

Source avatar币界网•5h ago

Dell has teamed up with Palantir to launch a native AI operating system: the underlying storage replaces Ceph, directly targeting defense and core banks

Source avatar币界网•10h ago

Analog Devices plans to acquire Empower for $1.5 billion, challenging Nvidia's supplier MPS

Source avatarTechFlow•16h ago

WEEX Labs: AI and the financial sector resonate stronger, and U.S. stock risk appetite continues to pick up

Source avatarChainCatcher•19h ago

The three major stock indexes closed mixed, with Li Auto falling nearly 10%

Source avatarChainCatcher•23h ago

Ondo: Tokenized stock TVL exceeds $1.5 billion, with a concentration of about 25% of top assets

Source avatarChainCatcher•23h ago

Former OpenAI researcher Aschenbrenner is betting heavily on Bitcoin miners while shorting Nvidia and AMD

Source avatarChainCatcher•23h ago

The three major U.S. stock indexes turned down, with Micron Technology falling nearly 3%