Google has been designing its own AI accelerators since 2015, longer than any other hyperscaler, and by 2026 that program has produced two generations enterprises are actively deploying: Trillium (TPU v6e), generally available since late 2024, and Ironwood (internally TPU7x), Google’s seventh-generation chip announced in April 2025 and explicitly built for inference rather than training. The two chips mark a real architectural leap in a short span — roughly 2.5x the raw compute and up to 6x the memory capacity, per Google’s own published specifications. Here’s what actually changed under the hood, straight from Google Cloud’s own documentation.
At a Glance: Official Per-Chip Specifications
| Spec (per chip) | TPU v5p | Trillium (TPU v6e) | Ironwood (TPU7x) |
|---|---|---|---|
| Peak compute (BF16) | 459 TFLOPs | 918 TFLOPs | 2,307 TFLOPs |
| Peak compute (FP8) | 459 TFLOPs | 918 TOPs (Int8) | 4,614 TFLOPs |
| HBM capacity | 95 GiB | 32 GiB (31.25 GiB usable) | 192 GiB |
| HBM bandwidth | 2,765 GBps | 1,638 GBps | 7,380 GBps |
| ICI bandwidth | 1,200 GBps | 800 GBps | 1,200 GBps |
| Max chips per pod | 8,960 | 256 | 9,216 |
Figures sourced directly from Google Cloud’s own TPU documentation. Note that Trillium’s per-chip HBM capacity (32 GiB) is lower than the older v5p’s (95 GiB) — Trillium was optimized differently, trading some per-chip memory for higher compute density and a design meant to pack more chips efficiently per pod, while Ironwood restores and substantially exceeds v5p-class memory capacity.
Trillium (TPU v6e): The Workhorse Generation
Per Google’s own documentation, Trillium delivers 918 TFLOPs of peak bf16 compute and 1,836 TOPs at Int8 per chip, backed by 32 GB of HBM (31.25 GiB usable after system and runtime reservations) at 1,638 GBps of bandwidth. Chips connect over a 2D-torus Inter-Chip Interconnect (ICI) fabric rated at 800 GBps bidirectional per chip across four ICI ports. A full Trillium Pod tops out at 256 chips (arranged 16×16, with 8 chips per host), delivering a pod-level 234.9 PFLOPs of bf16 compute, 102.4 TB/s of all-reduce bandwidth, and 3.2 TB/s of bisection bandwidth. Slice configurations range from a single chip up to the full 256-chip pod, with an 8-chip single-VM slice specifically optimized for inference workloads. Google’s documentation describes v6e as sharing design similarities with the earlier v5e generation, scaled up substantially in compute and networking.
Ironwood (TPU7x): Built for the “Age of Inference”
Ironwood, which Google’s own blog post calls “the first Google TPU for the age of inference,” pushes every major spec further. Per chip, it delivers 2,307 TFLOPs of bf16 compute and 4,614 TFLOPs at FP8 — roughly 2.5x Trillium’s bf16 throughput. Memory jumps to 192 GiB of HBM (6x Trillium’s capacity) at 7,380 GBps of bandwidth (about 4.5x Trillium’s), and ICI bandwidth rises to 1,200 GBps bidirectional across six links, 1.5x Trillium’s figure and matching the older v5p generation. A full Ironwood pod scales to 9,216 chips, organized as 144 cubes of 4x4x4, which Google rates at 42.5 exaflops of pod-level peak compute — a figure Google says is more than 24 times the roughly 1.7 exaflops-per-pod rating of El Capitan, which it describes as the world’s largest supercomputer. Google also claims a 2x improvement in performance-per-watt over Trillium, and says Ironwood is nearly 30x more power-efficient than Google’s original 2018 Cloud TPU, attributing part of that gain to liquid cooling that it says sustains up to twice the performance of standard air cooling under heavy load.
The “inference” framing is deliberate and specific: Google’s blog describes a shift from AI that responds to queries toward AI that proactively retrieves information and generates insights and multi-step reasoning on its own — the kind of workload Google calls “thinking models,” including large mixture-of-experts systems. Ironwood’s enhanced SparseCore, a dedicated unit for handling large embeddings efficiently, is aimed specifically at these larger, sparser model architectures, as well as at ranking and recommendation workloads in finance and science. Google states Ironwood chips can be combined via its Pathways software into pools of hundreds of thousands of chips working together, though the company’s blog post stops short of giving a maximum chip count for that combined configuration beyond the 9,216-chip single pod figure.
How the Two Generations Fit Together
Google doesn’t frame Ironwood as a strict replacement for Trillium; the two generations are positioned for different jobs. Trillium remains a capable, widely deployed general-purpose TPU generation with a smaller per-chip memory footprint tuned for cost-efficient scale-out. Ironwood is explicitly inference-first, with the memory capacity and bandwidth that large, parameter-heavy production models need to serve efficiently at scale, while also being capable of training. Both are available through Google Cloud as part of what Google calls its AI Hypercomputer architecture, which bundles the TPUs with Google’s specific networking, storage, and software stack rather than selling the chips as a standalone product the way a CPU or GPU vendor typically would.
Notably, Google’s own public materials for both chips avoid naming GPU competitors or giving direct GPU comparisons; the company’s competitive claims are made relative to its own earlier TPU generations and, for Ironwood, against the El Capitan supercomputer. That’s worth flagging for readers trying to benchmark TPUs directly against Nvidia or AWS silicon: Google has not published a like-for-like comparison, so any such comparison circulating elsewhere is a third-party estimate, not a Google-sourced figure.
Bottom Line
The jump from Trillium to Ironwood is one of the larger generational leaps in Google’s TPU history by its own published numbers — roughly 2.5x the compute, 6x the memory capacity, and a dedicated design orientation toward serving today’s largest reasoning and mixture-of-experts models efficiently rather than purely training smaller ones faster. For teams already on Google Cloud, the practical choice between the two generations should hinge on whether the workload is inference-heavy and memory-hungry (favoring Ironwood) or a more standard training/serving mix where Trillium’s wider availability and lower per-chip memory footprint remain a reasonable fit. For anyone evaluating TPUs against GPU or other custom-silicon alternatives, the key caveat is that Google has not published a direct, apples-to-apples comparison against any competing chip — all of the figures above are Google’s own, measured against its own prior generations.
FAQ
Is Ironwood the same as “TPU v7”?
Google’s own technical documentation refers to it as TPU7x, while its marketing materials call it Ironwood and describe it as the seventh generation of Cloud TPU; both names refer to the same chip.
Why does Trillium have less HBM per chip than the older v5p?
Per Google’s own published specs, Trillium (32 GiB) has less per-chip HBM than v5p (95 GiB); Google’s documentation frames v6e as sharing design lineage with the smaller v5e generation rather than v5p, reflecting a different scale-out design trade-off rather than a straightforward upgrade on every single spec.
How big is a full Ironwood pod?
Up to 9,216 chips, arranged as 144 cubes of 4x4x4, rated by Google at 42.5 exaflops of peak pod-level compute.
Does Google compare Ironwood directly to Nvidia GPUs?
No. Google’s official announcement and documentation compare Ironwood only to its own prior TPU generations and, for scale, to the El Capitan supercomputer — it does not publish a direct GPU comparison.
Can Trillium and Ironwood chips be combined into larger clusters?
Google states that Pathways, its distributed-training software, can combine hundreds of thousands of Ironwood chips; the single-pod maximums (256 chips for Trillium, 9,216 for Ironwood) describe the largest tightly interconnected unit within one pod, not an absolute ceiling on total deployable chips.
