AMD AI Chip Roadmap: MI450 and Helios Explained

In AI hardware, a new chip is usually shown on a stage long before anyone can plug it in. Months can pass between the keynote and the first server that runs it. Investors trade that gap, because an order only turns into sales once the chip and everything around it can be built.

The headlines make that gap hard to see. They use different words for a product at different stages, such as announced, launched, in production and shipping. They also mix up a chip's own name with the name of its family and of the rack it goes into, so it is hard to tell what a buyer can install today.

For a chip stock, the product roadmap works like a calendar of checks. Each date is a promise the company can keep on time, keep late or move, and investors can mark most of those dates well ahead of time.

Advanced Micro Devices (AMD) is at that stage with its MI450 chips and the Helios rack built around them, the hardware behind its largest AI deals. The two products carry much of AMD's wider AI business, and traders can follow their dates through the AMDUSDT Perpetual, BloFin's futures contract on AMD's share price.

What is AMD's AI chip roadmap?

AMD's AI chip roadmap is its schedule of Instinct GPUs and the Helios racks built from them. The MI400 Series launched on July 23, 2026, with MI500 due in 2027 and MI600 in 2028 (source: AMD, July 23, 2026). AMD put the start of Helios shipments on track for the third quarter of 2026 (source: AMD Q2 2026 earnings call).

A graphics processing unit (GPU) is a chip that does huge numbers of simple sums side by side, which is how AI models are trained and run. Instinct is the brand name AMD uses for the GPUs it sells to data centers. A rack is a tall cabinet of servers wired together, and Helios is AMD's plan for filling one with its newest chips.

Five status terms describe where a chip stands, and headlines often blur them. Announced means AMD has named the product in public. Launched means AMD has formally introduced it, often with a spec sheet, even if buyers wait many months for it. In production means the chip is being made, shipping means it is leaving the factory for customers, and in revenue means AMD has booked the sales in its results. Each generation sits on a different rung of that ladder:

The MI350 Series is the generation already in customers' data centers. In a release of November 11, 2025, AMD called it the fastest-ramping product in the company's history, already deployed at scale by cloud providers such as Oracle Cloud Infrastructure (source: AMD, November 11, 2025).

AMD says it will keep an annual cadence of new CPUs, GPUs, networking and racks through 2030 (source: AMD, July 23, 2026). A cadence is a steady rhythm, and on its August 4, 2026 call AMD said it plans to launch a new rack-scale AI platform every year (source: AMD Q2 2026 earnings call). Each of those launches hands investors a fresh date to check against.

The chip names get easier once you set today's chip beside the next one.

AMD MI455X specs compared with the MI355X

The MI455X is the GPU inside Helios, and the MI355X is the MI350 Series chip already running in cloud data centers. The MI455X carries 432 GB of HBM4 memory and is rated at 40.3 petaflops of peak FP4 compute (source: AMD MI455X specs). The MI355X carries 288 GB of the older HBM3E and 10.1 petaflops (source: AMD MI355X specs).

HBM, or high-bandwidth memory, is memory stacked right beside the GPU so that data reaches it fast. HBM4 is the newest generation of it, and HBM3E is the one before. Memory bandwidth is how much data can move in and out each second, counted here in terabytes per second (TB/s).

FP4 is a short, 4-bit number format that AI models can use when they answer requests, and a petaflop is a thousand trillion calculations a second. Peak means the ceiling the design allows on paper, so real programs run at or under it. The process node is the manufacturing generation at TSMC, the Taiwan company that makes the wafers for AMD's most advanced chips, and a smaller number means a newer process. A wafer is the thin silicon disc that chips are made on. The architecture is the GPU's design generation, and CDNA 5 is the one after CDNA 4.

Say you set the two chips side by side. Worked out from those figures, 432 GB against 288 GB is 1.5 times the memory, and 40.3 petaflops against 10.1 is about four times the peak FP4 compute. Bandwidth rises almost threefold, from 8 TB/s to 23.3 TB/s, and the fourfold FP4 jump matches AMD's own claim of four times the peak FP4 of its previous Instinct GPUs (source: AMD MI400 page).

AMD also gives a measured figure, and it needs one more term: a token is a small piece of text that an AI model reads or writes. AMD says the MI455X delivers 34 times the token throughput of the MI355X. That comes from its own lab tests in July 2026 on the Deepseek V4 Flash model (source: AMD, July 23, 2026).

The MI400 Series is the family, and the MI455X is the member AMD builds into Helios (source: AMD MI400 page). AMD's November 2025 release and its August 2026 call both called the Helios GPU part of the "MI450 Series" (source: AMD, November 11, 2025). So an MI450 in a headline and an MI455X on a spec sheet point to the same rack. Memory is the widest gap in AMD's own comparison with Nvidia, and HBM4 supply is also one of the bottlenecks on the ramp.

MI430X, MI440X and MI350P

The MI455X gets the headlines, but the MI400 Series has other members built for other buyers. The MI430X is aimed at high-performance computing and sovereign AI, meaning AI computing a country builds and runs for itself. It delivers up to 288 teraflops of FP64, the high-precision 64-bit math used in scientific computing (source: AMD, July 23, 2026). A teraflop is a thousand times smaller than a petaflop.

The MI430X shows why launched and available are different rungs. AMD's second-quarter results release lists it among the MI400 chips AMD launched (source: AMD Q2 2026 results). AMD's own product page says the MI430X is expected to be available in 2027 (source: AMD MI400 page). The chip will power supercomputers including Discovery at Oak Ridge National Laboratory and Alice Recoque, France's first exascale supercomputer (source: AMD, January 5, 2026). On August 31, 2026, AMD added LUMI-AI in Kajaani, Finland, an MI430X system scheduled for deployment in the second half of 2027 (source: AMD, August 31, 2026).

AMD announced the MI440X earlier, on January 5, 2026, for companies that run AI in their own data centers. It comes in a compact eight-GPU form that fits into existing infrastructure, and AMD aims it at training, fine-tuning and inference, the work of putting a finished model to use on real questions (source: AMD, January 5, 2026).

AMD also launched the MI350P at the July 23, 2026 event, a GPU meant to bring AI acceleration to the infrastructure that customers already run (source: AMD, July 23, 2026).

Taken together, one keynote can launch several chips whose customer dates fall in different years. When a headline says an AMD chip launched, check the availability date AMD gave for that exact model before you treat it as shipping.

What is AMD Helios?

AMD Helios is AMD's reference design for a complete AI server rack, holding 72 Instinct MI455X GPUs and 18 EPYC "Venice" processors (source: AMD, July 23, 2026). AMD calls it a blueprint for partners to build their own branded systems, and it rates each rack at up to 2.9 exaflops of FP4 compute (source: AMD Helios page).

An exaflop is a thousand petaflops. A reference design means AMD draws up the whole rack and its parts list, then hands the plan to other companies to build. In AMD's words, Helios is "a reference design, not a product for sale" (source: AMD Helios page). Those builders are OEMs, which sell servers under their own brand, and ODMs, which design and build hardware for other companies to sell or run.

AMD's quarterly filing puts it in legal terms. AMD develops Helios and licenses it for customers and their manufacturing partners to build, and while it supplies certain products used in each system, it does not make or sell the completed racks (source: AMD 10-Q, Q2 2026).

Most of the rack's headline figures are the GPU's figures multiplied by 72. Take the MI455X's 432 GB of HBM4 and multiply it by 72, and the worked-out total is 31,104 GB, the 31 TB AMD lists for the rack. The same sum on 23.3 TB/s gives about 1,678 TB/s, which is the 1.7 petabytes per second AMD lists, since a petabyte is a thousand terabytes (source: AMD MI400 page).

Networking inside the rack is called scale-up, because it lets the 72 GPUs work as one large system. Networking between racks is called scale-out, because it lets a data center add more racks to the same workload. AMD rates Helios at 260 TB/s of scale-up bandwidth and 43 TB/s of scale-out bandwidth (source: AMD Helios page).

The cabinet follows Open Rack Wide, a rack standard Meta introduced through the Open Compute Project (OCP). It is double-wide, where earlier OCP racks were single-wide, and adds liquid cooling and wider trays for large accelerator boards (source: AMD MI400 page). AMD first showed Helios in public as a static display at the OCP Global Summit in San Jose on October 14, 2025 (source: AMD, October 14, 2025).

ZT Systems, Sanmina and who builds Helios racks

AMD bought rack design skills before it had a rack to offer. In March 2025 it completed its purchase of ZT Systems, which supplied AI and computing infrastructure to the largest cloud companies, for a total purchase price of $4.4 billion (source: AMD 10-Q, Q2 2026).

AMD kept the part of ZT that designs systems, with certain staff and intellectual property from those design operations (source: AMD 10-K, 2025). It sold the factory side, the ZT Manufacturing Business, to Sanmina in October 2025 for $2.4 billion in cash plus Sanmina shares, and it signed a manufacturing services agreement with Sanmina with a first term of five years (source: AMD 10-Q, Q2 2026).

That split shapes who builds the racks. Sanmina and Wiwynn are infrastructure partners for Helios, and the server brands Bull, HPE, Lenovo and Supermicro will offer Helios systems too (source: AMD, July 23, 2026). AMD's May 2026 Taiwan release named Wistron and Inventec as well, alongside Sanmina and Wiwynn, among the ODMs helping to build Helios-based systems (source: AMD, May 21, 2026).

So when a server maker announces a Helios system, the GPUs, CPUs and networking chips inside come from AMD. The builder puts the rack together, adds parts from other suppliers and sells the finished system, so the pace of each rollout depends on those companies as well as on AMD. The name on a Helios launch is therefore often a builder's, and a new OEM system shows that another company has taken up the design.

AMD Helios vs Nvidia Vera Rubin, as AMD compares them

AMD calculated in June 2026 that the MI455X has 15% more peak FP4 performance and 50% more memory than Nvidia's Vera Rubin GPU, using Nvidia's published specifications, which for FP4 were preliminary (source: AMD MI400 page). Per rack, AMD estimates Helios gives up to 30% more tokens per dollar than a Vera Rubin NVL72 rack (source: AMD, July 23, 2026).

Vera Rubin is the Nvidia GPU and rack that AMD measures itself against, and every figure in the comparison comes from AMD. Each claim rests on a stated basis and a stated month:

Tokens per dollar is the number of tokens a system produces for each dollar its GPUs cost to use. Throughput is the amount of work a system finishes in a set time.

AMD has published no price for an MI455X or a Helios rack. So the dollar in its tokens-per-dollar claim comes from AMD's own projection of hourly GPU prices, which it bases on market conditions. AMD made the same kind of claim a generation earlier, when it said at the June 2025 launch that the MI355X gave up to 40% more tokens per dollar than competing solutions (source: AMD, June 12, 2025).

AMD's footnoted figures come from its own test labs. They show how AMD positions Helios on paper, which is one input into how AMD and Nvidia compare as stocks. The next test of these claims comes when customers run the racks.

UALink, Ultra Ethernet and the Open Rack Wide standard

AMD also sells Helios on how it is built. It designed the rack on open standards, which are shared sets of rules that any company can build products to. They include Open Rack Wide for the cabinet, Ultra Ethernet and UALink (Ultra Accelerator Link), a standard for scale-up links that connect GPUs directly to one another (source: AMD Helios page). Ultra Ethernet is the networking standard of the Ultra Ethernet Consortium, or UEC.

AMD's case for them is about choice: it says Helios reduces fragmentation by replacing "proprietary, one-off designs" with open standards such as OCP, UALink and UEC (source: AMD MI400 page). AMD's Pensando Vulcano AI network cards carry the rack's networking using UALink (source: AMD Helios page).

The first racks may use a halfway step, since AMD's spec page labels the MI455X's scale-up link "UALoE" and lists 3.6 TB/s of bandwidth for it (source: AMD MI455X specs). In February 2026, Tom's Hardware reported that at least the first Helios machines would run UALink over Ethernet, and said that means lower performance (source: Tom's Hardware).

The same report quoted the chief executive of Astera Labs, who said the company would be ready for the first customer platform ramps in 2027 (source: Tom's Hardware). A platform ramp is the move from first systems to volume, so on that view some of the UALink hardware reaches volume in 2027, after the quarter AMD gave for the first Helios shipments.

For an investor, the point is timing, because each standard has to be ready in working parts, from many suppliers, before the racks that depend on it can scale. A headline about UALink or Ultra Ethernet parts can therefore tell you as much about Helios timing as a headline about AMD's own chips.

When MI450 and Helios ship: AMD's own dates

In November 2025, AMD set Helios for the third quarter of 2026 (source: AMD, November 11, 2025). On August 4, 2026, AMD chief executive Lisa Su said Helios was in production. Its initial shipments remained on track to start later that quarter, then ramp in the fourth quarter and into 2027 (source: AMD Q2 2026 earnings call).

AMD has restated its timing several times since then, and each statement came in public, with a date. In AMD's wording, "2H 2026" means the second half of 2026, and a gigawatt means one billion watts of electric power.

A ramp is the climb from first units to full volume, and Su described Helios's ramp in three steps. She said the third quarter "is the very beginning of the ramp", the fourth quarter "is a step up", and the first quarter of 2027 "a further step up" (source: AMD Q2 2026 earnings call).

AMD kept that shape in September. On September 8, 2026, chief financial officer Jean Hu said the third quarter would be "the very beginning of our production shipment", with revenue expected in it and a very significant step up in the fourth (source: Citi conference transcript, September 8, 2026). The third-quarter 2026 report can show whether initial shipments began, but by AMD's own account the larger volumes come in the fourth quarter and through 2027.

Customers have set their own dates around that ramp. OpenAI, which has a supply agreement with AMD, expects to bring Helios online beginning in the fourth quarter of 2026, with deployments speeding up through 2027 (source: AMD, July 23, 2026). Anthropic's agreement is sized in gigawatts, and the first gigawatt of AMD GPUs under it begins deploying in the first half of 2027 (source: AMD and Anthropic).

AMD's balance sheet shows the build-up. Inventory is the stock of chips and materials a company holds before it sells them, and work in process is the part still being made. AMD's inventory rose from $7,920 million on December 27, 2025 to $8,468 million on June 27, 2026, and work in process made up $5,265 million of the later total (source: AMD 10-Q, Q2 2026).

Say you work out the change. $8,468 million minus $7,920 million is a rise of $548 million in six months, and about 62% of the June total was still in production. AMD said on the call that inventory rose "to support strong data center demand" (source: AMD Q2 2026 earnings call). The filing does not say how much of that inventory is MI450 or Helios parts.

Delay reports and AMD's answers

Outside reports have questioned the schedule. A Tom's Hardware report on February 17, 2026 carried a note from SemiAnalysis, a chip research firm. The note expected engineering samples and small-volume output in the second half of 2026, with the mass-production ramp coming only by the second quarter of 2027 because of manufacturing delays (source: Tom's Hardware).

An AMD executive answered on X. Anush Elangovan, a corporate vice president for software development, wrote that the assessment was "still wrong" and replied "On target for 2H 2026" (source: Tom's Hardware). Both sides are claims made on a date, one by a research firm and one by AMD. AMD's next quarterly report, for the third quarter of 2026, is the first regular document that can test them.

AMD's filings carry the formal version of the risk. The 10-Q gives the second half of fiscal 2026 as the expected start of Helios shipments. It also warns that AMD's newer and more complex products could run into quality or production problems that delay them (source: AMD 10-Q, Q2 2026).

Before you trade on the next MI450 or Helios headline, run it through five questions:

• Is this news an announcement, a launch, production, a shipment, or booked revenue?

• Whose words are these: an AMD filing, an AMD release, a customer, or a research firm?

• Is the figure a peak on paper or a measured result, and compared with what?

• Does the news move a date that AMD gave before, and if so, which one?

• Does it touch a supply bottleneck such as wafers, HBM4 memory, packaging or rack parts?

With those answers in hand, the AMDUSDT Perpetual page shows the live price of BloFin's contract on AMD.

What could slow the MI450 ramp?

Four things could slow the MI450 ramp: new 2nm and 3nm wafers from TSMC, HBM4 memory, packaging and substrates, and rack parts from other companies. AMD's quarterly filing names TSMC as the maker of all its GPU wafers at 7nm or smaller, and warns that missing third-party rack parts could delay rack-scale systems (source: AMD 10-Q, Q2 2026).

AMD's chiplet approach builds one product from several smaller chips. Packaging joins chips and memory into one finished part, and a substrate is the base layer the package sits on.

When an analyst raised tight 3nm supply on the August 4 call, Su pointed to AMD's chiplet approach and said it lets AMD ramp in fewer wafers on the new node (source: AMD Q2 2026 earnings call). The wider risk behind that answer is AMD's dependence on TSMC.

HBM4 memory is the second bottleneck, and on March 18, 2026, Samsung signed an MOU with AMD to align on primary HBM4 supply for the MI455X (source: Samsung and AMD). An MOU is a written agreement to work together, and Samsung's announcement gave no volume or price. On the August call, Su said AMD has very good visibility into HBM allocation for what it expects to deliver in 2027 (source: AMD Q2 2026 earnings call).

Su named back-end capacity, packaging capacity and substrates as parts of the supply chain AMD is watching beyond wafers (source: AMD Q2 2026 earnings call). On May 21, 2026, AMD announced more than $10 billion of investment across Taiwan's chip industry to scale advanced packaging (source: AMD, May 21, 2026).

The last bottleneck is rack parts. The 10-Q warns that a shortage of parts from other companies could delay or reduce customers' ability to build racks, even when AMD delivers its own chips on time (source: AMD 10-Q, Q2 2026).

AMD's unconditional purchase commitments, amounts it has contracted to pay in future periods, were $12,166 million on December 27, 2025 (source: AMD 10-K, 2025). By June 27, 2026 they were $30,276 million, with $17,386 million due in the rest of fiscal 2026 (source: AMD 10-Q, Q2 2026). That is about 2.5 times as much, worked out from the two totals. The total mainly covers wafers, substrates and components, plus multi-year cloud computing capacity, software and technology licenses, with no split given.

Outside the supply chain sits customers' data-center readiness, a bottleneck AMD shares with its buyers. Su said AMD is working very closely with data center operators (source: AMD Q2 2026 earnings call).

After MI450: MI500, MI600 and AMD's yearly cadence

After MI450, AMD plans an MI500 Series in 2027 and an MI600 Series in 2028, each with its own rack. Helios 500 pairs MI500 GPUs with EPYC "Verano" processors, and Helios 600 pairs MI600 with EPYC "Ferrara" (source: AMD, July 23, 2026). AMD says MI500 will use 2nm technology and HBM4E memory (source: AMD, January 5, 2026).

Planned, in AMD's releases, means a year AMD has named for a product it has yet to launch. In January 2026, AMD said MI500 is planned to launch in 2027 on its CDNA 6 architecture, the design generation after the MI455X's CDNA 5 (source: AMD, January 5, 2026). HBM4E is an upgraded version of HBM4 memory.

AMD has made big claims for MI500. In January 2026 it said MI500 is on track for up to a 1,000x increase in AI performance over the MI300X, a GPU it introduced in 2023 (source: AMD, January 5, 2026). On the August 4 call, Su said AMD expects MI500 to be the largest generational leap in Instinct history. She said the 2027 platform adds larger scale-up domains, meaning more GPUs linked as one, and both copper and optical links, which carry data as light (source: AMD Q2 2026 earnings call).

The CPUs in each rack follow AMD's EPYC server roadmap, which brings EPYC chips on its "Zen 7" CPU design in 2028 (source: AMD, July 23, 2026).

AMD is also buying companies whose work can feed later generations. On June 15, 2026, AMD announced it was acquiring MEXT. Its technology uses AI to make flash storage behave more like DRAM, the fast working memory in a computer, so AI systems get more usable memory (source: AMD, June 15, 2026). AMD's August results release then listed MEXT as acquired (source: AMD Q2 2026 results). On August 6, 2026, AMD agreed to acquire Taalas, a Toronto company founded in 2023 that designs chips for AI inference, and said it plans to fold that technology into its accelerator roadmap (source: AMD, August 6, 2026).

Every generation also runs on AMD's ROCm software. At the July 2026 event, AMD introduced ROCm.ai, a platform that uses AI to help developers build and tune GPU software faster (source: AMD, July 23, 2026).

Advancing AI, AMD's roadmap event

Advancing AI is the event AMD used in both 2025 and 2026 to unveil its next Instinct generation. At Advancing AI 2025 on June 12, 2025, AMD unveiled the MI350 Series and previewed Helios as its next AI rack, to be built on MI400 Series GPUs (source: AMD, June 12, 2025). At Advancing AI 2026 on July 23, 2026, it launched the MI400 Series and Helios (source: AMD, July 23, 2026).

Between the two events, AMD also set roadmap dates in a November 2025 release, which put MI500 in 2027 (source: AMD, November 11, 2025). At CES in January 2026, it kept that year and announced the MI440X (source: AMD, January 5, 2026).

AMD's shares closed at $552.33 on July 22, 2026, the day before the 2026 event, and at $539.69 on the day of it (source: Stock Analysis price history). One event can bring launches, new dates and new comparisons with Nvidia all at once, so the useful question afterward is which dates AMD moved.

After each Advancing AI, set the new roadmap beside the old one. Look for three changes: a year that moved, a chip that changed name, and a claim that changed its basis. Each of those changes a date or a basis you may already be relying on.

AMD's roadmap dates and the AMDUSDT Perpetual

The AMDUSDT Perpetual is BloFin's perpetual contract on AMD's share price, settled in USDT, so roadmap news reaches it through AMD's stock. The contract offers up to 20x leverage (source: BloFin AMDUSDT contract information). The next check on the roadmap is AMD's third-quarter 2026 report, the first that can show whether Helios shipments began.

The AMDUSDT Perpetual has traded on BloFin since May 6, 2026. A perpetual works like a futures contract that never reaches a settlement date, so you can hold it open while your margin still covers it. Funding is a regular payment between long and short holders, and on this contract it settles every eight hours (source: BloFin AMDUSDT contract information). BloFin's stock futures follow the referenced shares' price and give you no ownership interest, so a position on the contract holds no AMD shares (source: BloFin notice).

Several of AMD's roadmap releases have landed outside regular US trading hours. Its November 11, 2025 release came out at 4.30PM EST, after the US close (source: AMD, November 11, 2025). Its May 21, 2026 release came out at 1.35AM EDT, hours before the open (source: AMD, May 21, 2026). Outside regular weekday trading hours, and on weekends and market holidays, BloFin may restrict its stock futures, the AMDUSDT Perpetual included, to reduce-only orders (source: BloFin notice). Those orders can only cut back or close out an open position.

Say a report of a Helios delay comes out on a Saturday while you are long the AMDUSDT Perpetual. Until regular trading resumes, BloFin may let you shrink or close that position and hold back any order that adds to it.

Leverage raises the stakes of a surprise date. BloFin may reduce or liquidate a position once its margin ratio falls to 100%, meaning your collateral plus any profit has dropped to the maintenance margin plus closing fees (source: BloFin liquidation rules). Maintenance margin is the minimum needed to keep a position open, so a position can be cut before all the margin you posted is gone.

The third-quarter shipments check will come with AMD's earnings for that quarter, when a roadmap update and a results surprise can hit the price together. If you plan on trading AMD around earnings, size the position for both at once.

Before you trade AMD with leverage into a roadmap date, decide how far the price could move against you on that day. Then set your margin so the position can survive that move.

Looking to trade the AMDUSDT Perpetual around AMD's next roadmap date? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and go to the AMDUSDT Perpetual trading page.

Frequently asked questions

When will the AMD MI450 be released?

AMD launched its MI400 Series on July 23, 2026, and the MI455X, the Helios GPU that AMD has also placed in its MI450 Series, lists that as its launch date (source: AMD MI455X specs). Delivery is a separate step, and AMD's Helios product page says the Helios reference design is being shared with partners now, with volume deployments expected in the second half of 2026 (source: AMD Helios page). AMD also said in July that Meta had begun testing and validating workloads on Helios racks as it prepares to deploy at scale (source: AMD, July 23, 2026).

Is the AMD MI450 the same as the MI455X?

In practice both names refer to the GPU in AMD's Helios rack, and AMD itself has switched between them. Its July 23, 2026 release described Anthropic's plan to deploy up to 2 gigawatts of MI455X GPUs in Helios racks (source: AMD, July 23, 2026). A day earlier, AMD's announcement with Anthropic called the same plan up to 2 gigawatts of MI450 Series GPUs (source: AMD and Anthropic). AMD's May 21, 2026 release used a third spelling, "MI450X", for the GPUs in Helios-based systems (source: AMD, May 21, 2026).

How much does an AMD Helios rack cost?

AMD lists no price for a Helios rack or an MI455X, and partners build and sell the finished systems. What AMD publishes is a cost-efficiency estimate. Its figure of up to 30% more tokens per dollar than an Nvidia Vera Rubin NVL72 rack came from AMD Performance Labs in July 2026. It used the Kimi K2 Thinking model with 32K input tokens and 8K output tokens, and a projection of hourly GPU prices. AMD's footnote adds that system setups may vary by manufacturer and may produce different results (source: AMD, July 23, 2026).

What memory does the AMD MI455X use?

The MI455X uses 432 GB of HBM4 arranged in 12 memory stacks, which works out to 36 GB per stack. It also carries 192 MB of L2 cache, a small, very fast memory on the chip itself (source: AMD MI455X specs). Samsung, the planned primary HBM4 supplier, says its HBM4 is built on its sixth-generation 10nm-class DRAM process, called 1c. Samsung rates it at processing speeds of up to 13 gigabits per second (source: Samsung and AMD).

What did AMD announce at Advancing AI 2026?

At Advancing AI 2026 on July 23, 2026, AMD launched its 6th Gen EPYC server processors alongside the Instinct MI400 Series GPUs and Helios racks. It also launched Ryzen AI Embedded X100 processors and the Kria AI SOM and Robotics Developer Platform, part of what it called a physical AI portfolio (source: AMD, July 23, 2026). Anthropic, OpenAI, Meta, Cerebras, AT&T and Cisco each described how they work with AMD, and AMD shared new CPU and GPU roadmap details reaching 2030.

Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include AMD's Form 10-Q for the second quarter of 2026 and Form 10-K for fiscal 2025, AMD's second-quarter 2026 results release and earnings call transcript, AMD's product pages for the Instinct MI355X, MI400 Series, MI455X and Helios, AMD announcements of June 12, 2025, October 14, 2025, November 11, 2025, January 5, 2026, May 21, 2026, June 15, 2026, July 23, 2026, August 6, 2026 and August 31, 2026, the AMD and Anthropic announcement, Samsung and AMD's announcement of March 18, 2026, the transcript of AMD's September 8, 2026 Citi conference appearance, reporting by Tom's Hardware, Stock Analysis price history, and BloFin's AMDUSDT contract information, stock futures notice and liquidation rules, current as of September 28, 2026.

Nothing in this article constitutes financial advice, and nothing in it is a recommendation to buy, sell or hold AMD in any form. The MI450 and Helios dates are AMD's own guidance, so a late first shipment, a shortfall in HBM4 memory or packaging, or racks that fall short of AMD's comparisons with Nvidia could jolt AMD's share price when the news lands, sometimes outside regular weekday trading hours. In those hours BloFin may limit the AMDUSDT Perpetual to reduce-only orders, and a leveraged position can be reduced or liquidated once its margin falls to the maintenance level, before all the margin you posted is used up. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.