The AMD Instinct MI400 series is AMD's 2026 generation of datacenter GPUs. AMD launched it on July 23, 2026, with the MI455X (432 GB of HBM4, up to 23.3 TB/s) built for the 72-GPU Helios rack, and the MI430X for HPC and sovereign AI. "MI450" is the earlier name for the same generation; AMD's July materials name the MI455X and MI430X.
Status, plainly: these parts are launched but not generally available. AMD's product page says the Helios reference design "is being shared with partners now, with volume deployments expected in 2H 2026," and the MI430X "expected to be available in 2027." Aquanode does not rent MI400-series parts. The box near the end of this post shows what you can run today.
This guide covers:
- What the MI400 series, MI450, MI455X, MI430X and MI440X names mean
- The MI455X and Helios specs, with AMD's comparison claims
- Dated availability status and who has committed to buy
- How it compares with the MI355X you can use today
TL;DR
- Names: MI400 is the series. AMD's July 23, 2026 launch lists the MI455X (frontier AI, Helios) and the MI430X (HPC and sovereign AI). AMD's earlier materials, including its Q2 2026 results release, still use "MI450 Series" for the family.
- Per GPU: the MI455X has 432 GB of HBM4, up to 23.3 TB/s of bandwidth, and up to 40 PFLOPS of peak FP4 and 20 PFLOPS of peak FP8 (AMD). That is 1.5x the MI355X's 288 GB of memory.
- Per rack: Helios has 72 MI455X GPUs, 18 EPYC "Venice" CPUs, 31 TB of HBM4, and up to 2.9 exaFLOPS of FP4 (AMD).
- Status: launched July 23, 2026; AMD expects volume deployments in the second half of 2026. OpenAI says it expects Helios online starting in the fourth quarter of 2026, with deployments accelerating through 2027.
- Verdict: if you are planning 2027 capacity, treat Helios as a rack-scale NVIDIA Vera Rubin NVL72 alternative to evaluate. If you need GPUs this quarter, the MI355X and MI300X are the AMD parts that exist in the field.
The names: MI400, MI450, MI455X, MI430X, MI440X
AMD's naming has shifted over the past year, and search results mix the labels, so here is the map.
| Name | What it is | Source and date |
|---|---|---|
| MI400 Series | The generation after the MI350 Series | Previewed June 12, 2025; launched July 23, 2026 |
| MI450 Series | Earlier family label. AMD's Q2 2026 results release still describes a deal "to deploy up to 2 gigawatts of MI450 Series GPUs in AMD Helios racks" | AMD results release, August 4, 2026 |
| MI455X | The Helios rack GPU, for frontier AI and AI factories | AMD launch release, July 23, 2026 |
| MI430X | Sovereign AI and HPC part, up to 288 TFLOPS of FP64 | AMD launch release, July 23, 2026 |
| MI440X | An eight-GPU, on-premises enterprise AI system announced at CES | AMD CES release, January 5, 2026 |
Two cautions. First, the July 23 launch release defines the family as MI455X and MI430X only; the MI440X is not listed in that definition, and we found no AMD availability date for it. Second, the same Anthropic partnership is described as "MI450 Series" in AMD's Q2 results release and as "MI455X" in its July 23 release. Read "MI450" as the generation, and "MI455X" as the specific GPU AMD ships in Helios.
MI455X and Helios specs
AMD's product page lists the following. Everything in this table is AMD's peak theoretical figure, and AMD's endnotes say results are subject to change when products are released in market.
| Spec | MI455X (per GPU) | MI355X (per GPU, for reference) |
|---|---|---|
| Architecture | 5th Gen CDNA | 4th Gen CDNA |
| Memory | 432 GB HBM4 (12 stacks) | 288 GB HBM3E |
| Memory bandwidth | up to 23.3 TB/s | 8 TB/s |
| Peak FP4 | up to 40 PFLOPS | 10.07 PFLOPS (dense MXFP4) |
| Peak FP8 | up to 20 PFLOPS | 5.03 PFLOPS (dense) |
| Peak FP16 / BF16 | 5 PFLOPS (AMD comparison table) | 2.52 PFLOPS (dense) |
| Work group processors | up to 256 | 256 compute units |
| Scale-up bandwidth per GPU | up to 3.6 TB/s | 7 x 153.6 GB/s Infinity Fabric |
The jump is large on paper: 1.5x the memory, about 2.9x the bandwidth, and roughly 4x the peak FP4 and FP8 throughput. The MI455X figures are AMD design targets from a launch; the MI355X figures come from a shipping-product datasheet. Hold the MI455X numbers loosely until independent measurements exist. For a deeper look at the memory change, see our HBM3E vs HBM4 post, and for the per-GPU predecessor see the MI355X guide.
The Helios rack
AMD's July 23, 2026 release describes Helios as combining "72 high-performance AMD Instinct MI455X GPUs and 18 powerful 6th Gen AMD EPYC Venice CPUs, connected by AMD Pensando front-end, scale-up and scale-out networking." The key rack figures from AMD:
| Helios rack | AMD figure |
|---|---|
| GPUs | 72 MI455X in 18 four-GPU compute trays |
| Peak FP4 | up to 2.9 exaFLOPS |
| Peak FP8 | up to 1.4 exaFLOPS |
| HBM4 memory | 31 TB |
| Memory bandwidth | up to 1.7 PB/s |
| Scale-up fabric | UALink over Ethernet (UALoE), 72 GPUs in one domain |
| Scale-out | Pensando Vulcano 800 AI NICs, aligned with the Ultra Ethernet Consortium |
| Rack standard | Open Rack Wide (ORW), submitted to the Open Compute Project, designed with Meta |
Per-GPU scale-up of up to 3.6 TB/s across 72 GPUs works out to roughly 260 TB/s for the rack, which matches the 260 TB/s that IT Pro's launch coverage cites (our arithmetic, not an additional AMD claim). The scale-up fabric is the notable design choice: Helios uses the open UALink protocol carried over Ethernet rather than a proprietary switched fabric. Our UALink vs NVLink post compares the two approaches, and our NVLink explainer covers NVIDIA's side.
AMD's claims against NVIDIA Vera Rubin
AMD compares Helios with NVIDIA's Vera Rubin NVL72 rack, which is not shipping either. Treat every number below as AMD's claim, computed against NVIDIA's published preliminary specifications.
- Per GPU: AMD's MI455X table lists 40 PFLOPS of FP4 against 35 for Vera Rubin, 20 against 18 for FP8/FP6, and 5 against 4 for FP16/BF16. It lists 432 GB against 288 GB of memory, and 23.3 TB/s against 22.0 TB/s of bandwidth. AMD's summary is 15% more MXFP4 peak performance.
- Per rack: AMD says Helios offers 15% more peak FP4 performance, 50% more HBM capacity, 6% more HBM bandwidth and 50% more scale-out bandwidth than the leading competitive solution.
- Cost per token: AMD says Helios delivers "up to 30% more tokens per dollar." AMD's endnote says this is modeling, not measurement: its Performance Labs modeled token throughput per GPU on Kimi K2 Thinking (32K input, 8K output) at three interactivity points against NVIDIA's "published modeled specifications" for a Vera Rubin NVL72 rack (July 2026).
The NVIDIA side of that comparison is covered in our Rubin guide and the Vera Rubin NVL72 guide. We have not seen independent MLPerf results for either rack, and we will not estimate any.
AMD also said in June 2025 that the MI400 series was expected to deliver "up to 10x more performance running inference on Mixture of Experts models" than the previous generation. That was an engineering projection at the time and AMD has not republished it as a measured result, so treat it as a target. See mixture of experts for why MoE inference leans on memory capacity.
Availability and who has committed
Dates matter here, so each one has its source.
- June 12, 2025: AMD previews Helios and the MI400 series at Advancing AI 2025 (AMD press release).
- January 5, 2026: AMD shows Helios at CES, introduces the MI440X, and previews the MI500 series for 2027 (AMD press release).
- February 23, 2026: AMD's data center GM said AMD was "highly confident of ramping Helios in high volume in the second half of the year," and denied reports of a thermal-related delay. In the same article, the analyst firm SemiAnalysis, not AMD, was reported to expect engineering samples and low-volume production in the second half of 2026 and mass production by Q2 2027 (Next Platform). That is a third-party view.
- July 23, 2026: AMD launches the MI400 series and Helios at Advancing AI 2026. AMD's release says leading AI labs and cloud providers including OpenAI, Anthropic, Meta, Microsoft, Oracle, HUMAIN, Tensorwave, Vultr and Cirrascale are choosing Helios, with systems from Bull, HPE, Lenovo and Supermicro. AMD's August results release goes further and says Helios "is being deployed by" a similar list. The release says OpenAI "expects to bring Helios online beginning in the fourth quarter of 2026, with deployments accelerating throughout 2027." It also describes Anthropic's plan to deploy up to 2 gigawatts of MI455X GPUs in Helios racks.
- August 4, 2026: AMD's Q2 results release says Helios "begins to ramp."
- MI430X: AMD's product page says it is "expected to be available in 2027."
What we did not find: an AMD statement that Helios racks have shipped to customers in volume, or any public price. "Launched" and "being deployed" are AMD's words for a ramp that is starting. If you are planning capacity, ask your cloud or OEM contact for a delivery date in writing, not a launch date.
Infrastructure needs
The MI455X is a rack-scale part. You do not slot it into an existing eight-GPU server.
- Rack: Helios is a double-wide Open Rack Wide (ORW) rack. It needs datacenter space built or retrofitted for that standard.
- Power and cooling: AMD's pages we reviewed do not publish per-GPU or per-rack power figures. Third-party coverage cites figures above 2 kW per GPU and a rack near 245 kW, but we could not confirm those from AMD and do not rely on them. Expect liquid cooling at this scale and ask the OEM for the figure.
- Networking: Scale-up is UALink over Ethernet inside the rack. Scale-out is Ultra Ethernet-aligned, with Pensando Vulcano 800 NICs.
- Software: ROCm, with AMD stating that the MI400 series runs on the same open stack as today's Instinct parts. If you plan ahead, validate your inference stack on an MI355X first; kernel and framework work you do now should carry forward, though AMD has not published a migration guide we can cite.
MI355X, MI300X or wait?
Run on an MI355X or MI300X now when: you need capacity this quarter, your model fits in 288 GB or 192 GB per GPU, and you want to be on ROCm early. The MI300X vs H100 vs H200 post covers the software-maturity trade-offs.
Plan for Helios when: you are sizing 2027 frontier-scale serving or training, you want an open-standards rack, and you can wait for volume. The 432 GB of memory per GPU and 31 TB per rack are aimed at very large mixture-of-experts models and long-context serving.
Look at NVIDIA when: you need a shipping rack-scale system today. The GB200 NVL72 is covered in its own guide, and the B300 is the current Blackwell Ultra part.
What to run today
No MI400-series hardware is rentable on Aquanode. Aquanode manages and optimizes GPUs for training and inference workloads, and the box below shows the closest GPUs you can rent on demand right now, with live prices. A row reading "None right now" means no offer at the moment.
What's next
AMD's roadmap, as stated in its July 23, 2026 release, has the MI500 series in 2027, with the Helios 500 rack to pair it with EPYC "Verano" CPUs, and the Helios 600 rack and MI600 series after that. The January 2026 CES release adds that the MI500 is planned on CDNA 6, a 2nm process and HBM4E memory, and that AMD claims up to a 1,000x AI performance increase over the MI300X (an AMD claim, dated January 5, 2026). For the whole picture across vendors, see the datacenter GPU guide.
FAQ
What is the AMD MI450?
MI450 is the name AMD used for the MI400-series family before launch, and AMD's Q2 2026 results release still uses "MI450 Series." The GPU AMD launched for the Helios rack on July 23, 2026 is the MI455X.
Is the MI450 or MI455X available?
Launched, not generally available. AMD says the Helios reference design is being shared with partners and volume deployments are expected in the second half of 2026. OpenAI says it expects Helios online starting in the fourth quarter of 2026. Aquanode does not rent it.
How much memory does the MI455X have?
432 GB of HBM4 at up to 23.3 TB/s per GPU (AMD). A 72-GPU Helios rack has 31 TB of HBM4.
What is AMD Helios?
A 72-GPU rack-scale system from AMD built on the MI455X, 18 EPYC Venice CPUs and Pensando networking, in the double-wide Open Rack Wide format. Scale-up uses UALink over Ethernet.
Is the MI455X faster than NVIDIA Vera Rubin?
AMD says so on peak FP4 (15% more) and memory (432 GB against 288 GB), and claims up to 30% more tokens per dollar for the rack. Those figures are AMD's modeling against NVIDIA's preliminary published specifications. No independent benchmark of either rack existed at the time of writing.
What is the difference between the MI455X and MI430X?
The MI455X targets frontier AI and the Helios rack. The MI430X targets HPC and sovereign AI, with up to 288 TFLOPS of hardware FP64, and AMD expects it to be available in 2027.
Sources
- AAI 2026: AMD Launches AMD Instinct MI400 Series GPUs, July 23, 2026
- AAI 2026: AMD Launches AMD Helios Rackscale Solution, July 23, 2026 (rack figures, 15% / 50% / 6% / 30% claims)
- AMD Instinct MI400 Series GPUs product page (MI455X specs, comparison table, availability FAQ, endnotes)
- AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era, July 23, 2026 (customers, OpenAI timing, Anthropic, MI500 and later roadmap)
- AMD second quarter 2026 financial results release, August 4, 2026 ("MI450 Series" wording, Helios "begins to ramp")
- AMD and its Partners Share their Vision for "AI Everywhere, for Everyone" at CES 2026, January 5, 2026 (MI440X, MI500 plans)
- AMD Advancing AI 2025 press release, June 12, 2025 (Helios preview, 10x MoE projection)
- IT Pro, AMD Advancing AI 2026 coverage (260 TB/s scale-up figure)
- The Next Platform, February 23, 2026: AMD says Helios racks and MI400 series GPUs on track for 2H 2026 (AMD executive quote, SemiAnalysis report)
- AMD Instinct MI355X GPU datasheet (MI355X reference figures)
