Three Inference Chips in One Week: Google TPU 8tx and 8i Max, AMD MI450X, Intel Crescent Island
Between 9 and 12 November 2026, Google, AMD and Intel each put new inference silicon on the table. The three announcements look unrelated at first:…
Between 9 and 12 November 2026, Google, AMD and Intel each put new inference silicon on the table. The three announcements look unrelated at first: a confidential-computing TPU and a dual-chip TPU from Google, a lower-memory variant of AMD’s flagship rack GPU, and an Intel inference GPU that partner boards can fit with up to 480 GB of low-power DRAM. Read side by side, they tell the same story. In inference, memory capacity and the key-value (KV) cache now shape chip design as much as raw compute does. For the broader market context, see our overview of the AI accelerator market.
Why inference forces a memory rethink
Training is a throughput problem: keep thousands of chips busy on one giant job for weeks. Serving is different. Every request carries its own context, and reasoning models and AI agents carry a lot of it: long documents, codebases, multi-turn conversations.
To avoid recomputing the whole context each time a new token is generated, inference engines store intermediate attention results, the “keys” and “values”, for every token already processed. This is the KV cache. It grows with context length and with the number of concurrent users, and it has to sit somewhere fast. When it no longer fits, the system either evicts part of it and recomputes later, which costs time, or moves it to a slower, larger memory tier, which costs latency.
Vendors now say so openly. Google describes its newest TPU as being “built for a different problem: the memory wall” [1]. AMD pitches its new GPU as a way for customers “to match memory capacity with their AI workloads” [3]. Intel, according to Seoul Economic Daily, explicitly framed its standalone-chip strategy around how much DRAM different services need for their KV cache [10]. Three vendors, one constraint.
Google: two new TPU systems, built around existing chips
On 9 November 2026, Amin Vahdat, Google’s SVP and Chief Technologist for AI and Infrastructure, introduced two additions to the eighth-generation TPU family: TPU 8tx and TPU 8i Max [1]. They join TPU 8t (training) and TPU 8i (reasoning and serving), which Google unveiled at Cloud Next in April. Google says this is “the first time we have introduced four TPU systems in a single year” [1].
The key word is system. Google distinguishes a new chip, a multi-year effort, from a new system built around an existing chip by changing packaging, memory, networking and security. Both products are of the second kind [1].
TPU 8i Max: two reasoning chips, one package
The baseline is TPU 8i. In its April architecture deep dive, Google listed TPU 8i with 288 GB of HBM capacity, 384 MB of on-chip SRAM, 8,601 GB/s of HBM bandwidth and 10.1 peak FP4 petaflops per chip [2]. Google also wrote that the 3x increase in on-chip SRAM over the previous generation lets TPU 8i “host a larger KV Cache entirely on silicon,” reducing core idle time during long-context decoding [2]. In other words, the KV cache was already a design target for the base chip.
TPU 8i Max places two TPU 8i chips on a single package. According to Google, it delivers “double the compute, double the SRAM and double the HBM bandwidth of TPU 8i,” and fits twice the silicon into the existing rack footprint [1]. Google positions it for “massive context windows and the highest compute density per rack,” citing scenarios in which dozens of agents work across an entire codebase or hours of video [1].
Two cautions apply. First, Google’s announcement does not state an HBM capacity figure for TPU 8i Max. TechCrunch described the product as “doubling the memory capacity” [8], and two TPU 8i chips would logically carry twice the HBM of one, but that capacity figure is not in Google’s post, so we do not print it here. Second, TechTarget reported 768 MB of on-chip SRAM and 17.2 TB/s of HBM bandwidth for the dual-chip design [9]. Those numbers match a simple doubling of Google’s published TPU 8i figures, but they come from press coverage, not from the announcement itself.
Google paired the hardware with software. The same post introduced TPU Sync, an open-source “model-state layer” that treats chip HBM, host DRAM, shared storage and the data center network as one memory hierarchy. Google says it automatically tiers large KV caches off the chip and streams context between disaggregated prefill and decode nodes, cutting KV transfer latency and time to first token by more than 90% [1]. That figure is Google’s own claim.

TPU 8tx: confidential inference with an AMD host
TPU 8tx targets a different buyer: regulated enterprises and governments that want to run frontier or proprietary models on sensitive data. Google says it is “built on the same silicon as TPU 8t and paired with 5th Generation AMD EPYC CPU hosts,” and that it builds security into the hardware layer so data and model weights stay “protected and isolated even while in active use” [1]. It also borrows the low-latency Boardfly network from TPU 8i, which Google says delivers “up to 2x lower latency compared to TPU 8t with a near-zero performance penalty” [1].
The AMD detail is notable. In April, Google described Arm-based Axion CPU hosts across the eighth-generation TPU systems [2]. For the confidential variant, the official post names AMD EPYC hosts instead [1]. TechCrunch went further and described TPU 8tx as “designed with AMD” [8]. Google’s own post does not use co-design language, so the safe reading is the official one: TPU 8tx uses TPU 8t silicon with AMD EPYC host CPUs.
Google says teams can also fine-tune, distill and run reinforcement learning on the same system that serves the model [1]. Google has not published availability dates or pricing for either new system, according to TechTarget [9].

AMD: same compute, a memory-capacity choice
AMD’s update came on 12 November 2026, through a newsroom post covering three products in the Instinct MI400 Series [3].
MI450X: a lower-memory option for Helios
The new product is the Instinct MI450X. AMD says it offers 288 GB of HBM4 per GPU “while delivering the same core compute, memory bandwidth and scale-up and scale-out connectivity capabilities as MI455X” [3]. Customers deploying AMD’s Helios rack-scale system can choose either GPU on the same rack architecture and ROCm software stack [3].
For comparison, AMD’s MI455X product page lists 432 GB of HBM4 across 12 stacks, 23.3 TB/s of peak memory bandwidth and 40.3 peak MXFP4 petaflops, in a direct-liquid-cooled module built for Helios [4]. The MI450X therefore keeps that compute and bandwidth profile but drops a third of the memory capacity.
AMD’s own framing is a clean statement of the KV-cache trade-off. For “workloads that fit within its memory capacity,” MI450X is pitched as a way to hit throughput and response-time targets while optimizing operating expenses. MI455X “provides additional memory headroom for larger models, longer contexts or higher concurrency” [3]. Longer contexts and higher concurrency are exactly the two variables that inflate the KV cache.
The MI455X performance claim, and how to read it
The same release carries AMD’s headline inference number. AMD says MI455X on Helios delivered 21,333 tokens per second per GPU on DeepSeek-R1, “resulting in up to 30% higher throughput than the Nvidia Vera Rubin latest published MLPerf offline result” [3]. These are AMD’s claims, based on AMD’s own testing.
The footnote matters. According to AMD, the result comes from testing by AMD Performance Labs in October 2026 on “an AMD-tested MLPerf Inference Offline configuration.” It is compared with the published MLPerf Inference v6.1 Closed Division result for Nvidia’s Vera Rubin NVL72 (entry 6.1-0106), retrieved from MLCommons on 11 November 2026 [3]. In plain terms, AMD’s number is an internal run set against a competitor’s published submission, not two results from the same official MLPerf round. Treat it as a vendor comparison.
AMD also claims Helios offers “up to 20% higher HBM bandwidth” and higher scale-up bandwidth than Vera Rubin, again based on AMD’s calculations from published specifications [3]. On its MI455X product page, AMD lists Vera Rubin at 288 GB and 22.0 TB/s for comparison [4]. Those competitor figures are as presented by AMD.
MI430X: memory for science
AMD also detailed the Instinct MI430X for high-performance computing and AI for science: 432 GB of HBM4, up to 23.3 TB/s of memory bandwidth and up to 288 TFLOPS of peak native FP64 [5]. It is not an inference chip, but AMD’s pitch rests on the same logic: more memory per GPU means fewer GPUs and fewer nodes for the same problem [5].
Intel: one chip, flexible DRAM
Intel’s contribution came from Seoul. At Intel AI Summit Seoul 2026 on 12 November, Anil Nanduri, who leads AI products and go-to-market in Intel’s Data Center Group, outlined plans for Crescent Island, the inference GPU now marketed as the Intel GPU D-series. The event was reported by Seoul Economic Daily on 13 November [10].
What Intel has officially published
Intel first announced Crescent Island in October 2025, describing an inference-optimized data center GPU on the Xe3P architecture with 160 GB of LPDDR5X, designed for air-cooled enterprise servers [7]. In October 2026, at the OCP Global Summit, Intel’s newsroom gave fuller specifications. The GPU D-series combines up to 480 GB of LPDDR5X, 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute within a 350-watt envelope, in a standard air-cooled PCIe form factor [6]. A footnote clarifies the memory range: Intel’s own branded PCIe card carries 160 GB of LPDDR5X, and the design lets partners build their own branded cards with options up to 480 GB [6]. Intel also says the GPU supports models of up to 300 billion parameters on a single card and more than one trillion parameters in an eight-GPU system, and that customer sampling was underway [6].
What was new in Seoul
Two points from the Seoul summit appear only in press coverage, so we attribute them. According to Seoul Economic Daily, Intel will sell Crescent Island as a standalone chip, so server makers can design their own cards and size DRAM to their workloads, from 160 GB on Intel’s design up to 480 GB on customer boards [10]. The paper also quotes Nanduri as saying Intel ran inference workloads in its labs within weeks of receiving physical samples, and that “we are currently shipping evaluation samples to customers” [10].
Seoul Economic Daily also reported the reasoning behind the strategy, and it is a KV-cache argument. When an AI service handles long documents or sustained conversations, it stores processed information in DRAM for reuse. If memory runs short and part of that cache is discarded, the work must be recomputed, which slows responses. That is why, the paper writes, DRAM needs vary by service even on the same GPU [10].
The memory choice is a bet. LPDDR5X offers far less bandwidth than HBM: Intel’s 1.5 TB/s [6] compares with 23.3 TB/s for AMD’s MI455X [4]. Intel is wagering that for many enterprise inference jobs, capacity and cost matter more than peak bandwidth. Seoul Economic Daily added that wider LPDDR5X use in AI servers could open supply opportunities for Samsung and SK hynix [10]; for the HBM side of that memory market, see our analysis of the HBM4 supply race.
Three answers to one question
Put the announcements side by side and the design choices line up along a single axis: how much memory to attach to each unit of compute, and at what cost.
- Google scales memory with compute. TPU 8i Max doubles compute, SRAM and HBM bandwidth in lockstep by packaging two chips together [1]. It then adds a software layer, TPU Sync, to spill KV caches across HBM, host DRAM and storage [1].
- AMD keeps compute fixed and varies memory. MI450X and MI455X share core compute, bandwidth and connectivity but differ in HBM4 capacity, 288 GB versus 432 GB [3][4]. Customers pick the capacity their models and concurrency require.
- Intel trades bandwidth for capacity. It uses LPDDR5X instead of HBM, reaches up to 480 GB per card on partner designs [6], and, according to Seoul Economic Daily, lets board makers choose the amount [10].
Google alone pulled a security thread this week: TPU 8tx makes confidential computing a dedicated inference product [1].
None of these approaches is universally better. What the week confirms is that the inference market is fragmenting by memory profile, as the training market did earlier by interconnect and rack scale (see our guide to multi-trillion-parameter training silicon).
What to watch next
- Availability and pricing. Google has not given dates or prices for TPU 8tx and 8i Max [9]. AMD has not priced MI450X [3]. Intel has given no general availability date for the GPU D-series in the sources reviewed here [6][10].
- Like-for-like benchmarks. AMD’s 30% claim against Vera Rubin rests on an AMD-run configuration [3]. An official MLPerf submission for MI455X would allow a direct comparison.
- Official TPU 8i Max memory figures. Google has not yet published HBM capacity per package [1].
- Intel’s standalone-chip model in practice. Watch for partner boards at 480 GB and for Intel primary material confirming the standalone-chip sales model reported from Seoul [6][10].
The common lesson is simple. In inference, the chip that wins is increasingly the one whose memory fits the workload, not just the one with the most compute.
Related reading
- The AI Accelerator Market: GPUs, ASICs and the Second-Source Race
- The HBM4 Supply Race: Samsung, SK hynix, and Micron Rewrite AI Memory Economics
- Multi-Trillion-Parameter Training: What the Next Silicon Generation Must Deliver
References / Sources
Primary / official
1. Google, Amin Vahdat, “Our newest TPUs: 8tx and 8i Max, built for the agentic era” (Nov. 9, 2026): https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/
2. Google Cloud Blog, Diwakar Gupta and Sabastian Mugazambi, “Inside the eighth-generation TPU: An architecture deep dive” (Apr. 22, 2026): https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive
3. AMD Newsroom, “AMD Advances the Instinct MI400 Series for AI and Scientific Computing” (Nov. 12, 2026), including endnotes MI400-031a, MI400-004a and MI400-013: https://newsroom.amd.com/news/amd-advances-instinct-mi400-series-ai-scientific-computing/
4. AMD, “AMD Instinct MI455X GPUs” product specifications page (accessed Nov. 16, 2026): https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html
5. AMD Blog, “Exascale Was the Beginning: AMD Instinct MI430X GPUs” (Nov. 12, 2026): https://www.amd.com/en/blogs/2026/exascale-was-the-beginning-amd-instinct–mi430x-gpus.html
6. Intel Newsroom, “Intel Accelerates Open Infrastructure for the Agentic AI Era” (OCP Global Summit, Oct. 2026): https://www.intel.com/content/www/us/en/newsroom/news/artificial-intelligence/intel-accelerates-open-infrastructure-agentic-ai-era.html
7. Intel Newsroom, “Intel to Expand AI Accelerator Portfolio with New GPU” (Oct. 14, 2025): https://www.intel.com/content/www/us/en/newsroom/news/artificial-intelligence/intel-to-expand-ai-accelerator-portfolio-with-new-gpu.html
Secondary / press (clearly attributed)
8. TechCrunch, Ram Iyer, “Google broadens its AI hardware portfolio with two new specialized chips” (Nov. 10, 2026): https://techcrunch.com/2026/11/10/google-broadens-its-ai-hardware-portfolio-with-two-new-specialized-chips/
9. TechTarget, Joab Jackson, “Google specializes TPUs for agentic, confidential computing” (Nov. 9, 2026): https://www.techtarget.com/news/google-specializes-tpus-for-agentic-confidential-computing/874084/
10. Seoul Economic Daily, Lee Seok-jin, “Intel to Sell AI Inference GPU as Standalone Chip, Tripling DRAM Capacity” (Nov. 13, 2026): https://en.sedaily.com/business/2026/11/13/intel-to-sell-ai-inference-gpu-as-standalone-chip-tripling
Images: original illustrations generated for AndBeyond.Tech; no logos or real people depicted.
