2026加速卡Accelerator
NVIDIA Rubin:288GB HBM4 / 22TB/sNVIDIA Rubin: 288GB HBM4 / 22TB/s
单卡 288GB HBM4(12-Hi stack)、22TB/s、336B 晶体管、224 SM / 896 Tensor Cores;NVFP4 推理最高 50 PFLOPS;NVLink 6 3.6TB/s。Per GPU: 288GB HBM4 (12-Hi stacks), 22TB/s, 336B transistors, 224 SM / 896 Tensor Cores; up to 50 PFLOPS NVFP4 inference; NVLink 6 at 3.6TB/s.
- HBMHBM
- 288GB HBM4
- 带宽BW
- 22TB/s
- NVFP4NVFP4
- 50 PFLOPS(inference)
- NVLink 6NVLink 6
- 3.6TB/s
- 晶体管transistors
- 336B
为什么值得看Why it matters它把单卡带宽从 8TB/s 拉到 22TB/s,是长上下文与高并发推理的硬上限变化;与 B300 同为 288GB,但 HBM 代际与互联不同。It lifts per-GPU bandwidth from 8TB/s to 22TB/s, a hard ceiling change for long-context and high-concurrency inference. It shares 288GB with B300 but differs in HBM generation and interconnect.
边界 / 它没主张什么Boundary / what it does not claim50 PF 是 NVFP4 推理性能,不是所有精度;不能与 AMD MXFP4 直接横比。它是单卡数字,不是 Vera Rubin NVL72 整机。50 PF is NVFP4 inference, not every precision; it is not directly comparable with AMD MXFP4. It is a per-GPU figure, not the Vera Rubin NVL72 rack.
来源Sources:NVIDIA Rubin 架构博客NVIDIA Rubin architecture blog
2026加速卡Accelerator
NVIDIA B300:288GB HBM3e / 8TB/sNVIDIA B300: 288GB HBM3e / 8TB/s
Blackwell Ultra 单卡:288GB HBM3e、8TB/s、15 PFLOPS dense NVFP4、160 SM / 640 Tensor Cores、最高 1400W。Blackwell Ultra per GPU: 288GB HBM3e, 8TB/s, 15 PFLOPS dense NVFP4, 160 SM / 640 Tensor Cores, up to 1,400W.
- HBMHBM
- 288GB HBM3e
- 带宽BW
- 8TB/s
- NVFP4 denseNVFP4 dense
- 15 PFLOPS
- 功耗TGP
- 最高 1,400W
为什么值得看Why it matters它是 2026 年仍在交付的「上一代但成熟」选项;与 Rubin 同为 288GB,差别在 HBM 代际、带宽与 NVFP4 算力,适合按供货与成本做现实选择。It is the “previous generation but mature” option still shipping in 2026; it shares 288GB with Rubin but differs in HBM generation, bandwidth and NVFP4 compute — a practical choice driven by supply and cost.
边界 / 它没主张什么Boundary / what it does not claim15 PF 是 dense NVFP4;不能与 Rubin 的 50 PF 推理数字横比。1400W 是最高 TGP,实际功耗随负载变化。15 PF is dense NVFP4 and cannot be compared directly with Rubin’s 50 PF inference figure. 1,400W is max TGP; real power varies with workload.
来源Sources:NVIDIA Blackwell Ultra 博客NVIDIA Blackwell Ultra blog
2026-07-23加速卡Accelerator
AMD MI455X:432GB HBM4 / 23.3TB/sAMD MI455X: 432GB HBM4 / 23.3TB/s
单卡 432GB HBM4(12 stack)、23.3TB/s、320B 晶体管;OCP MXFP4 40.3 PFLOPS、FP8 20.1 PFLOPS;UALoE 3.6TB/s;DLC。Per GPU: 432GB HBM4 (12 stacks), 23.3TB/s, 320B transistors; OCP MXFP4 40.3 PFLOPS, FP8 20.1 PFLOPS; UALoE 3.6TB/s; direct liquid cooling.
- HBMHBM
- 432GB HBM4
- 带宽BW
- 23.3TB/s
- MXFP4MXFP4
- 40.3 PFLOPS
- 互联scale-up
- UALoE 3.6TB/s
- 冷却cooling
- DLC
为什么值得看Why it matters432GB 是当前公开单卡最大容量档;对把超大模型或长 KV 缓存留在单卡上的场景,容量比峰值算力更先决定「能不能跑」。432GB is the largest publicly documented per-GPU capacity here; for keeping a very large model or long KV cache on one card, capacity decides feasibility before peak compute.
边界 / 它没主张什么Boundary / what it does not claim40.3 PF 是 OCP MXFP4 峰值,不是实测;不能与 NVIDIA NVFP4 直接横比。DLC 意味着风冷机房不能直接上。40.3 PF is OCP MXFP4 peak, not measured; it is not directly comparable with NVIDIA NVFP4. DLC means an air-cooled facility cannot simply take it.
来源Sources:AMD MI455X 规格页AMD MI455X spec page
2026-11-12加速卡Accelerator
AMD MI450X:288GB HBM4 / 23.3TB/sAMD MI450X: 288GB HBM4 / 23.3TB/s
与 MI455X 同为 CDNA5、同为 23.3TB/s,但容量为 288GB(8 个 HBM4 stack);MXFP4 40.3 PFLOPS;2026-11-12 上市。Same CDNA5 and 23.3TB/s as MI455X, but 288GB (eight HBM4 stacks); MXFP4 40.3 PFLOPS; launched 2026-11-12.
- HBMHBM
- 288GB HBM4(8 stacks)
- 带宽BW
- 23.3TB/s
- MXFP4MXFP4
- 40.3 PFLOPS
- 上市launch
- 2026-11-12
为什么值得看Why it matters它把 MI455X 的带宽/算力保留在 288GB 档,给不需要 432GB 的集群一个更现实的容量点;型号差在容量,不在架构代际。It keeps MI455X bandwidth/compute at the 288GB tier, a more realistic capacity point for clusters that do not need 432GB; the difference is capacity, not architecture generation.
边界 / 它没主张什么Boundary / what it does not claim不能因为型号更大就假定 MI455X 更强;两者单卡显存不同,机架配置也不同。MXFP4 与 NVFP4 不可直接横比。Do not assume MI455X is stronger just because the number is higher; per-GPU memory and rack configurations differ. MXFP4 and NVFP4 are not directly comparable.
来源Sources:AMD MI450X 规格页AMD MI450X spec page
2026-04-22加速卡 / ASICAccelerator / ASIC
Google TPU7x(Ironwood):192GiB / 7.38TB/sGoogle TPU7x (Ironwood): 192GiB / 7.38TB/s
第七代 TPU,单芯片 192GiB HBM3E、约 7.38TB/s、4614 TFLOPS FP8 / 2307 TFLOPS BF16;9,216 chips per pod;ICI 1.2TB/s;JAX / PyTorch 支持,TensorFlow 不支持。Seventh-generation TPU: 192GiB HBM3E per chip, about 7.38TB/s, 4614 TFLOPS FP8 / 2307 TFLOPS BF16; 9,216 chips per pod; ICI 1.2TB/s; JAX / PyTorch supported, TensorFlow not.
- HBMchip
- 192GiB HBM3E
- 带宽BW
- 约 7.38TB/s
- FP8FP8
- 4614 TFLOPS
- PodPod
- 9,216 chips
- 框架frameworks
- JAX / PyTorch
为什么值得看Why it matters它是非 GPU 路线里最完整的一档:单芯片容量、Pod 规模与互联都公开;对大规模训练、且能接受 JAX/PyTorch 技术栈的团队,是 GPU 之外的真实选项。The most complete non-GPU option: per-chip capacity, pod scale and interconnect are public. For large-scale training with a JAX/PyTorch stack, it is a real alternative to GPUs.
边界 / 它没主张什么Boundary / what it does not claimPod 规模不是单卡吞吐;4614 TF 是 FP8 峰值。TensorFlow 不支持是官方明确写的,迁移成本要单算。不能把 TPU 与 GPU 按同一 CUDA 生态逻辑比较。Pod scale is not per-card throughput; 4614 TF is FP8 peak. The docs explicitly say TensorFlow is unsupported, so migration cost is separate. TPUs and GPUs cannot be compared through the same CUDA-ecosystem logic.
来源Sources:Google Cloud TPU7x 文档Google Cloud TPU7x docs
2026整机 / 机架System / rack
AMD Helios:72× MI455X,31TB HBM4AMD Helios: 72× MI455X, 31TB HBM4
机架级方案:72 张 MI455X 或 MI450X,31TB HBM4,1.4 EF FP8 / 2.9 EF FP4;不是单卡规格。Rack-scale system: 72 MI455X or MI450X GPUs, 31TB HBM4, 1.4 EF FP8 / 2.9 EF FP4; not a per-GPU spec.
- GPUGPU
- 72× MI455X / MI450X
- HBM 合计total HBM
- 31TB HBM4
- FP8FP8
- 1.4 EFLOPS
- FP4FP4
- 2.9 EFLOPS
为什么值得看Why it matters它回答「一个机架能装多少 HBM、给多少算力」,是集群采购与机房规划的单位;与单卡条目分开,避免把整机数字当成单卡。It answers how much HBM and compute one rack holds, the unit for cluster procurement and facility planning; kept separate from per-GPU entries so rack numbers are not mistaken for card numbers.
边界 / 它没主张什么Boundary / what it does not claim31TB 是 72 张卡的合计;1.4 EF FP8 与 2.9 EF FP4 也是整机架。功耗、冷却与机房条件另算,不是「买到就能跑满」。31TB is the sum of 72 cards; 1.4 EF FP8 and 2.9 EF FP4 are rack-level too. Power, cooling and facility conditions are separate; buying the rack is not a guarantee of peak runtime.
来源Sources:AMD Instinct MI400 系列页AMD Instinct MI400 series page