AMD ROCm and the Software Gap Explained: A Risk to AMD Stock

Every AI chatbot, image tool and coding assistant runs on chips in a data center, and those chips sit idle until software tells them what to compute. Developers write that software with a chipmaker's own programming tools, and over the years code, courses and habits pile up around whichever tools they learned first.

That pile of existing work is what makes a change of chip supplier expensive. A company that wants a cheaper or faster chip from a second supplier may have to rewrite and retest code that already works. A team facing a launch date often decides that risk is too high. So a chip can win on price and speed and still sell slowly.

For investors in a chipmaker, this is the software gap: the distance between how easily customers can run their work on the market leader's chips and how easily they can run it on a challenger's. It seldom appears as a line in the results, yet it shapes how many chips the challenger can sell and at what price.

In AI chips, the leader's software is Nvidia's CUDA and the challenger's is AMD's ROCm. On BloFin, AMD exposure comes through the AMDUSDT Perpetual, a futures contract that tracks AMD's share price, so a ROCm headline matters to a position there only through its effect on that price.

What is ROCm?

ROCm is AMD's open-source software for programming AMD's graphics processing units (GPUs), the chips that train and run most AI models. It does for AMD's chips what CUDA does for Nvidia's. AMD shipped the first version, ROCm 1.0, in April 2016 (source: AMD ROCm blog, August 2026). ROCm 10.0 followed on August 26, 2026 (source: ROCm release history).

A GPU is a processor designed to run huge numbers of simple calculations at the same time, which suits AI work. Training is the stage where an AI model learns from data, and inference is the stage where the finished model answers requests. Both need a stack of software under them: drivers that talk to the chip, a compiler that turns code a person writes into instructions the chip runs, and libraries of ready-made math routines.

Nvidia describes CUDA as its platform for accelerated computing, the software layer that lets programs use its GPUs (source: NVIDIA CUDA page). ROCm fills that role for AMD's Instinct data center GPUs, its Radeon graphics cards and the graphics built into its Ryzen processors, on Windows and Linux alike (source: AMD ROCm blog, August 2026). Open source means the code is public, so anyone can read it, change it and send improvements back. AMD built ROCm around a C++ compiler and a GPU programming language of its own called HIP (source: AMD ROCm blog, August 2026).

A kernel is a small program that runs on the GPU, such as the routine that multiplies two large grids of numbers. Much of the contest between the two platforms is a race to have fast kernels ready for each new AI model. Most AI developers work higher up, in a framework, a toolkit such as PyTorch that lets them build and train a model in the Python language. A serving engine, such as vLLM or SGLang, is the program that runs a finished model and answers users. The four layers line up the same way on both makers' chips, with different names at the bottom of the stack.

The shared tools matter most for AMD, because a customer whose model runs through PyTorch and a shared serving engine can often move it to AMD with no change to CUDA code. ROCm is also one piece of AMD's push into AI data centers, next to the Instinct chips it programs.

CUDA lock-in and AMD's GPU sales

CUDA lock-in is the cost a customer faces in moving AI code built for Nvidia's CUDA onto another maker's GPUs. Nvidia introduced CUDA in November 2006 (source: Computer Graphics World, November 2006). Most AI tools, courses and code grew up around it in the years before ROCm arrived.

Worked out from the two launch dates, CUDA had a head start of about nine and a half years over ROCm 1.0. In that time researchers wrote their code for CUDA, universities taught it and companies hired people who knew it. A customer who picks AMD also has to check that every piece of that code still runs, and still runs fast, and each check costs engineers' time. Investors call a lasting lead of this kind a moat, and Nvidia built this one over two decades, as the story of Nvidia's moat and the custom-chip threat shows.

AMD names the problem in the risk factors of its 10-Q, the quarterly report it files with the US Securities and Exchange Commission. The one for the quarter that ended June 27, 2026 says Nvidia uses its market position in data center GPUs, its financial resources and its proprietary software ecosystem to promote its systems. It adds that Nvidia uses them to influence AMD's customers (source: AMD 10-Q, Q2 2026). That part of the filing also counts a lack of software interoperability among the rival practices that can keep customers from picking AMD's products. That phrase means programs that work on one maker's chips and fail on another's. AMD says this may limit its market share and its margins (source: AMD 10-Q, Q2 2026).

The dates show how late ROCm started and how quickly the pace picked up after 2023. Day-zero support means a new version of a tool worked on AMD's GPUs on the day it launched. AMD reported the first big step for frameworks in June 2023, when it said PyTorch 2.0 had day-zero support on ROCm 5.4.2 across its Instinct GPUs (source: AMD release, June 2023).

For a shareholder, the lock-in is the software gap seen as a risk. It caps how many GPUs AMD can sell and how much it can charge for them, however fast the chips are.

How AMD measures ROCm adoption

AMD measures ROCm adoption by models, launch-day support and outside contributions. On August 4, 2026, chief executive Lisa Su said more than 3 million models run out of the box on AMD. She added that leading open models launch with day-zero support and that outside contributions to ROCm rose more than tenfold in a year (source: AMD Q2 2026 call).

Out of the box means a model runs on AMD's GPUs without changes to its code. An open model is one whose trained settings are published for anyone to download. A contribution is a change that someone outside a project sends in, such as a bug fix or a faster kernel. Su also said ROCm had reached an important inflection point in performance and ease of use for customers running AI in production (source: AMD Q2 2026 call).

AMD gave an earlier reading of a similar count on October 10, 2024. At that point more than 1 million models ran out of the box on its Instinct GPUs, triple the number when its MI300X chip launched (source: AMD release, October 2024). Line up AMD's two figures, more than 1 million in October 2024 and more than 3 million in August 2026, and the count roughly tripled in under two years, worked out from AMD's statements. The bases differ, since the 2024 figure counted models on Instinct GPUs and the 2026 figure counted models on AMD as a whole. At the CES technology show on January 5, 2026, AMD also said ROCm had delivered up to five times better AI performance over the past year (source: AMD release, January 2026).

Each of these is AMD's own count, and each measures reach. A model that runs out of the box can still run slower on an AMD chip than on an Nvidia chip, so speed has to be checked separately. AMD's chief financial officer, Jean Hu, told the same call that operating expenses rose 40% on the year as AMD kept investing in AI chips, systems and software (source: AMD Q2 2026 call).

ROCm.ai and ROCm 10

ROCm.ai is AMD's AI-assisted platform for writing GPU software, launched at its Advancing AI event on July 23, 2026. It lets coding assistants such as Claude, Codex and Cursor understand AMD's chips and ROCm natively, so they can help write, port and tune GPU code (source: AMD release, July 2026). To port code is to move it from one platform to another.

ROCm.ai began rolling out with ROCm 10.0 in three parts. The ROCm CLI is a single command-line tool for installing, checking and serving AI workloads, and it ships as a tech preview, an early version. AMD Skills load AMD-checked ROCm knowledge into coding assistants, and Hyperloom is an open-source system that profiles and tunes AI workloads with no engineer driving each step (source: AMD ROCm blog, August 2026). AMD reports that Hyperloom cuts weeks of manual tuning down to hours (source: AMD ROCm blog, August 2026).

Su told the August 4 call that ROCm.ai delivers more than twice the training performance and more than three times the inference performance of ROCm 7 across a broad range of models (source: AMD Q2 2026 call). Vamsi Boppana, who runs AMD's AI software group, spoke at the July 23 investor roundtable. He said AMD believes ROCm.ai is the biggest leap it has made, perhaps since the early days when it laid out its strategy (source: AMD investor roundtable, July 2026).

These gains are AMD's own measurements of its own software. Outside benchmarks, which test AMD's chips beside Nvidia's with the same models and rules, are the check on them.

Which AI frameworks run on ROCm in 2026

The main AI frameworks and serving engines run on ROCm in 2026. The ROCm 10.0 release of August 26, 2026 supports PyTorch 2.13.0, JAX 0.11.0, vLLM 0.27.0 and SGLang 0.5.15, in place of the older versions ROCm 7 had supported (source: ROCm 10.0 release notes). The gaps sit in the newest code built for Nvidia's chips.

That list is the practical test for a customer. When the tools a team already uses run on ROCm, moving to AMD is mostly a matter of testing. When the team depends on code written only for Nvidia's chips, the move means rewriting. PyTorch's own install page offers its stable 2.14.1 release with a ROCm 7.14 build, from the ROCm 7 line, alongside its CUDA builds (source: PyTorch). vLLM lists AMD ROCm among the GPU types it supports (source: vLLM docs). SGLang keeps a separate page on running it on AMD GPUs (source: SGLang docs).

The newest, most tuned code is where AMD trails. FlashAttention is a popular kernel library that speeds up attention, the step where an AI model weighs how the words in a request relate to each other. Its second version runs on AMD GPUs through two backends, versions of its code written for AMD's chips, while its third version is written for Nvidia's Hopper GPUs and its fourth for Hopper and Blackwell (source: FlashAttention on GitHub). Hopper and Blackwell are generations of Nvidia GPUs, such as the H100 and the B200.

A rack is a cabinet of stacked servers, and rack-scale software makes a whole rack run as one machine, the way AMD's Helios racks link 72 GPUs into one system. Serving one very long request across many GPUs needs extra software. SemiAnalysis, a chip research firm, wrote in August 2026 that AMD's version of two such methods, which it calls DCP and PCP, was yet to be tuned (source: SemiAnalysis report, August 2026). The main open tools run on ROCm today, while the newest code tuned for Nvidia and some multi-GPU serving methods still lag.

ROCm's production record at Microsoft and Meta

ROCm has run large AI services in production since 2024. AMD said on May 21, 2024 that its MI300X GPUs and ROCm were powering Microsoft's Azure OpenAI Service for GPT-3.5 and GPT-4 (source: AMD release, May 2024). Meta later said MI300X served all live traffic on its Llama 405B model (source: AMD release, October 2024).

A production deployment is stronger evidence than a benchmark, because a customer's own service depends on the chips every hour of the day. On the same day in May 2024, Microsoft made its Azure ND MI300X v5 virtual machines generally available, with Hugging Face as the first customer. A virtual machine is a slice of a server that a cloud provider rents out. Kevin Scott, Microsoft's chief technology officer, said using ROCm and MI300X had given Microsoft's AI customers and developers excellent price-performance (source: AMD release, May 2024).

Meta's statement came at AMD's Advancing AI event on October 10, 2024, as AMD's release reports it. By June 12, 2025, Meta said MI300X was broadly deployed for inference on its Llama 3 and Llama 4 models (source: AMD release, June 2025). Meta has since signed its own GPU deal with AMD, sized in gigawatts, a measure of the power the chips draw, and both companies sit on the list of who buys AMD's AI chips.

The Microsoft and Meta deployments came before AMD's gigawatt-scale GPU deals with OpenAI in October 2025 and Meta in February 2026.

Anthropic's July 2026 deal to put as much as 2 gigawatts of MI455X GPUs to work in Helios racks also covers software: the two companies said they would use Claude to tune workloads for Instinct GPUs and speed up ROCm development (source: AMD release, July 2026).

Independent benchmarks for ROCm against CUDA

Independent benchmarks show ROCm close to CUDA on some AI models and far behind on others. In SemiAnalysis tests from late August 2026, AMD led on cost for parts of some workloads. Nvidia was up to five times cheaper per unit of output on one model and delivered over 20 times the performance on another (source: SemiAnalysis report, August 2026).

A benchmark is a standard test that runs the same work on different chips so the results can be compared. MLCommons, an industry group, runs MLPerf Inference, which it calls an industry-standard benchmark. Its version 6.1 results, published on September 16, 2026, drew a record 30 submitting organizations, AMD and Nvidia among them (source: MLCommons, September 2026). MLCommons named no overall winner in its announcement.

SemiAnalysis runs InferenceX, an open benchmark that tests chips all the time and posts every run's log on GitHub for anyone to check (source: SemiAnalysis report, August 2026). In August 2026 the firm added AgentX, a test that replays long coding sessions of the kind AI agents run, across more than 1,000 chips (source: SemiAnalysis report, August 2026). Performance per dollar is the output a chip delivers for each dollar it costs to buy and run. AI output is measured in tokens, short chunks of text that are often part of a word.

The results split by model and by software, and they pit MI355X, the Instinct generation before AMD's July 2026 MI400 series, against Nvidia's B200 GPU and its GB300 NVL72, a rack of 72 GPUs.

With AMD's own ATOM engine, MI355X beat the GB300 NVL72 running vLLM on performance per dollar for Kimi K3, between 40 and 60 seconds of response time (source: SemiAnalysis, Kimi K3). ATOM also gave AMD better performance per dollar than the GB300 NVL72 over part of the range for GLM 5.3 (source: SemiAnalysis, GLM 5.3 on ATOM). DeepSeek V4 Pro was the closest race. Until August 21, 2026, MI355X on SGLang matched a B200 on vLLM for performance per dollar, though a B200 on SGLang was already ahead, and vLLM updates then put the B200 on vLLM ahead too (source: SemiAnalysis, DeepSeek V4 Pro). For Qwen3.5, with SGLang on both, SemiAnalysis found over 20 times better performance for the GB300 NVL72 at 90 tokens a second per user (source: SemiAnalysis, Qwen3.5).

For GLM 5.3 at 150 tokens a second per user, SemiAnalysis wrote that Nvidia's lead was so large that cost per token would still be lower on Nvidia even if AMD's chips were free. It tied that result to the current state of AMD's software (source: SemiAnalysis, GLM 5.3 on SGLang).

How to read a ROCm benchmark

A ROCm benchmark result depends on four things: the software that ran on the AMD chip, the kind of work tested, the date of the run and who ran it. The same chip can look strong or weak depending on those four, so before you treat one headline as the answer on the software gap, check each of them.

• The software on the AMD chip matters, because MI355X running the open vLLM engine trailed AMD's own ATOM engine in the firm's tests. It urged AMD to move those gains into vLLM, the engine more customers use (source: SemiAnalysis report, August 2026).

• The kind of work matters. vLLM on MI355X worked on day zero for short, single-question requests with Kimi K3, yet it was unusable in the first week on long sessions of many turns (source: SemiAnalysis report, August 2026).

• The date matters, because the DeepSeek V4 Pro ranking flipped within one week in August 2026, and a memory-copy function vLLM needs on AMD chips was missing until ROCm 7.14 (source: SemiAnalysis report, August 2026).

• The tester matters, because a figure from AMD's own lab, such as the ROCm.ai gains, measures AMD's software on AMD's terms, while an outside tester sets its own rules.

The trend in these reports is toward faster support. The firm wrote in August 2026 that AMD's first public recipe for splitting a model across machines on MI355X had reached InferenceX in January, months behind Nvidia. For the MiniMax M3 model the same kind of support landed on day zero, which it called an improvement on the DeepSeek-R1 period, when matching Nvidia took months (source: SemiAnalysis report, August 2026). Even so, wide gaps remain on some models. SemiAnalysis's September 14, 2026 report on Vera Rubin NVL72, Nvidia's next rack, also measured DeepSeek V4 Pro at 100 tokens a second per user, with SGLang on both MI355X and a B200. MI355X reached 2.01 million tokens a second for each megawatt of power, against 6.95 million for the B200 (source: SemiAnalysis, September 2026).

AMD's software acquisitions since 2023

Since August 2023, AMD has added at least seven AI software businesses or teams. It bought Mipsology and Nod.ai in 2023, Silo AI in 2024, and Brium, Untether AI's engineers and MK1 in 2025; the FastFlowLM team joined it in 2026. Only Silo AI has a published price, about $665 million in cash (source: AMD release, July 2024).

Most of these buys added people who write the lower layers of the stack, where AMD trails most. A compiler team, for example, builds the software that turns a model's code into fast instructions for the chip. When AMD announced Brium, it called the deal the latest in a series after Silo AI, Nod.ai and Mipsology. AMD said the series aims to support open-source software and faster performance on AMD hardware (source: AMD blog, June 2025). Mipsology, based in Palaiseau, France, made AI inference software whose Zebra tool supports TensorFlow, PyTorch and ONNX Runtime (source: AMD blog, August 2023). Nod.ai brought open-source compiler tools, including its SHARK software (source: AMD release, October 2023). The Silo AI deal closed on August 12, 2024 (source: AMD release, August 2024). The Untether AI engineers joined to work on AI compilers and kernels (source: CRN, June 2025). MK1's Flywheel software was serving over 1 trillion tokens a day when AMD completed the deal in November 2025 (source: AMD blog, November 2025). The FastFlowLM team, which built software for running AI models on AMD-powered PCs and workstations, joined AMD's AI group, as AMD announced on July 17, 2026 (source: AMD blog, July 2026).

AMD's spending on people points the same way. Its research and development costs were $2.5 billion in the quarter that ended June 27, 2026, up 33% on a year earlier, mainly from hiring for its AI strategy (source: AMD 10-Q, Q2 2026). On September 8, 2026, Jean Hu said AMD's acquisitions really focus on the software capabilities it wants to add, and that investors should expect that focus to continue (source: Citi conference transcript). Its two most recent deals widen that lens beyond the stack. On August 6, 2026 AMD agreed to buy Taalas, a maker of chips built for AI inference, subject to approvals (source: AMD release, August 2026). Then on September 28, 2026 it agreed to acquire World Labs, the spatial-intelligence AI lab founded by Fei-Fei Li, in an all-stock deal worth about $8.2 billion, its largest acquisition since Xilinx; Li is to join AMD as chief scientist, and the deal is expected to close by the end of 2026 (source: AMD release, September 2026).

Custom AI chips and AMD's software gap

Custom AI chips are processors that large cloud and AI companies design for their own data centers, and they compete with AMD for the same role as an alternative to Nvidia. Amazon, Google, Microsoft, Meta and OpenAI all have one, so work that leaves CUDA can go to those chips as well as to AMD's GPUs.

That is the second half of the software question. The open tools AMD depends on also make it easier to move work onto chips that AMD does not sell. Microsoft's software kit for its Maia 200 chip includes PyTorch support and a Triton compiler (source: Microsoft blog, January 2026). Triton is OpenAI's framework for writing GPU kernels, and AMD and OpenAI use it with ROCm to tune GPT-class models on AMD's MI455X GPUs (source: AMD release, July 2026). So code that leaves CUDA through Triton can land on a custom chip as well as on an AMD GPU.

Several of these chips sit inside AMD's own customers. Microsoft, an early ROCm user, says Maia 200 will serve OpenAI's GPT-5.2 models, among others (source: Microsoft blog, January 2026). Meta and OpenAI have both signed GPU deals with AMD sized in gigawatts, including AMD's OpenAI deal, and both also build chips of their own with Broadcom. Amazon made its Trainium3 chip generally available on December 2, 2025 (source: AWS, December 2025). Meta widened its work with Broadcom on MTIA chips on April 14, 2026, with a commitment of more than 1 gigawatt (source: Meta, April 2026). Google also builds its own AI chips. The dated steps for each chip show how fast these rivals are moving.

SemiAnalysis claimed on August 25, 2026 that OpenAI's first chip beat every Nvidia, AMD and Google chip it had been able to test on several top open models (source: SemiAnalysis, August 2026). That is one firm's test of a first-generation chip, and it is the kind of result that can change after the next round.

The software gap as an AMD stock risk

The software gap matters to AMD stock because it helps decide how many GPUs AMD can sell and at what price. AMD reports no separate figure for ROCm, so its progress shows up in data center results, in customer deployments and in outside benchmarks, each arriving on a date you can mark (source: AMD 10-Q, Q2 2026).

AMD's own view is upbeat, with a caveat from the same speaker. Matt Ramsay, who leads financial strategy and investor relations, said on August 11, 2026 that the gap had narrowed to the point where it was hardly a topic with AMD's top customers. He added that software is a battle every day (source: KeyBanc conference transcript). The outside tests say the answer depends on the model and the software stack. Both views feed into comparing Nvidia and AMD stock, where the software moat is one factor among several.

A software setback reaches the share price through expectations. If a benchmark or a customer report suggests ROCm is falling behind on a popular model, investors may lower what they expect AMD to sell, and the price can move before any sales figure changes. If you hold AMD shares elsewhere, a short position on the AMDUSDT Perpetual gains when AMD's price falls. It can offset part of that kind of drop, which is how hedging with the AMD perpetual works.

Before you trade on ROCm news, put these dated items on your calendar:

1. New ROCm releases, and the framework versions each one supports in its release notes.

2. SemiAnalysis's next AgentX update, which it said on August 24, 2026 would come in three to four weeks with new AMD and Nvidia results (source: SemiAnalysis report, August 2026).

3. The next MLPerf results from MLCommons, and whether AMD submits its newest GPUs.

4. AMD's quarterly results, and whether management repeats or updates its ROCm figures, such as the 3 million models it cited on August 4, 2026.

5. Helios deliveries: on August 4, AMD said Helios was in production, with first shipments on track to begin later in its third quarter and ramp through the fourth (source: AMD Q2 2026 call).

6. The time of day each item comes out, since AMDUSDT may be restricted to reduce-only orders, the kind that can only make a position smaller or close it, outside regular weekday trading hours.

Ahead of any of those dates, the AMDUSDT Perpetual page shows the contract's live price.

Trading AMD's software news on BloFin

The AMDUSDT Perpetual is BloFin's futures contract on AMD stock, with leverage up to 20x, so a position can be up to 20 times the money behind it. ROCm news moves it only through AMD's price, and outside regular weekday trading hours it may take only orders that trim or end a position (source: BloFin stock futures notice).

Software news runs on its own schedule. ROCm releases, AMD blog posts and outside benchmark reports come out when they are ready, so a headline about the software gap can land on a weekend. BloFin lists the contract with 24/7 trading hours, yet different rules apply outside the stock market's regular trading hours (source: BloFin listing notice). Nasdaq's regular session runs from 9.30AM to 4PM New York time, so pre-market and after-hours trading fall outside it (source: Nasdaq, stock market schedule). BloFin's notice on stock futures lets it restrict them to reduce-only orders at those times (source: BloFin stock futures notice).

Each contract stands for 0.01 AMD shares and has no end date, which is what perpetual means. Its price tracks an index that outside data vendors build from AMD quotes, and the index keeps its last value when no outside quote arrives (source: BloFin listing notice). The position gives you AMD's price movement and nothing else, with no share, no dividend and no vote attached (source: BloFin stock futures notice). For scale, AMD closed at $611.76 on September 30, 2026 (source: Stock Analysis).

Say an outside benchmark report comes out on a Saturday and shows ROCm well behind on a popular model, while you hold a long position of 1,000 contracts opened at $630. A long position gains when AMD's price rises. Yours carries about $6,300 of AMD exposure, worked out as 1,000 × 0.01 × $630. At 10x leverage you would have put up about $630 of margin, the money that backs the position, so a 10% slide before Monday would use up all of it. BloFin may trim the position or close it sooner, once your margin, with any profit or loss counted, falls to the maintenance margin plus the fees for closing out (source: BloFin liquidation rules). Maintenance margin is the least amount of margin that keeps a position alive. During reduce-only hours you could trim that position, and any addition has to wait until regular weekday trading hours resume.

A stop-loss is an order that closes a position at a price you set in advance. Size, leverage and a stop-loss are worth settling before a ROCm release or a benchmark day, the same choices behind leveraged AMD trades of any kind.

Looking to trade the AMDUSDT Perpetual on ROCm news? 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

Is ROCm open source?

ROCm is open source, and since ROCm 7.14 every release has come out of TheRock, an automated open-source system that builds and checks the whole stack in one pipeline before it ships. AMD is also moving its ROCm packages, its GPU driver and its public GPU tools onto a single download site, repo.amd.com. The site can hold several ROCm versions side by side, so teams can choose when to upgrade (source: AMD ROCm blog, August 2026).

Does ROCm work on Windows?

ROCm runs on Windows as well as Linux, and both now draw on one software kit. With ROCm 10.0, AMD retired its separate Windows HIP software development kit (SDK) and moved both systems onto the same ROCm Core SDK and the same release schedule. It says native Windows installers will follow later in 2026 (source: AMD ROCm blog, August 2026). At the CES technology show on January 5, 2026, AMD said ROCm 7.2 would extend support for its new Ryzen AI 400 Series PC processors across Windows and Linux (source: AMD release, January 2026).

Which AMD GPUs does ROCm support?

ROCm supports AMD's Instinct data center GPUs, its Radeon graphics cards and the graphics built into Ryzen processors. The ROCm 10.0 release of August 26, 2026 added the Radeon RX 9050. It also added virtual machine setups for the Instinct MI355X, MI350X, MI350P, MI325X, MI300X and MI210, which let one physical GPU be passed to a virtual server. AMD keeps the full, current list on its ROCm hardware support page, which changes with each release (source: ROCm 10.0 release notes).

How often does AMD release ROCm updates?

AMD says ROCm 10 and its minor releases will arrive about every six weeks (source: AMD ROCm blog, August 2026). Before ROCm 10 the gaps varied: 7.1 came about six weeks after 7.0, while 7.2 on January 21, 2026 and 7.14 on July 15, 2026 were months apart. ROCm 10.0 followed on August 26, 2026, six weeks after 7.14, and AMD still patches the older line, with 7.14.1 on September 2, 2026 (source: ROCm release history). Matt Ramsay said on August 11, 2026 that another ROCm version will launch with AMD's MI500 GPUs in 2027 (source: KeyBanc conference transcript).

What is MLPerf and who runs it?

MLPerf is a set of AI benchmarks run by MLCommons, an open engineering group backed by more than 130 members and affiliates, which began the MLPerf tests in 2018. Its inference round measures how fast systems run trained AI models under fixed rules. In the version 6.1 round of September 16, 2026, the best DeepSeek R1 result per chip was 5.7 times better than a year earlier. MLCommons also said a newer test, MLPerf Endpoints, will replace the inference round for data centers (source: MLCommons, September 2026).

Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include AMD's August 4, 2026 earnings call, its quarterly report for the three months to June 27, 2026, and its press releases and blog posts on ROCm and its purchases from 2023 to 2026. They also include AMD's ROCm release notes and release history and transcripts of AMD's July, August and September 2026 investor events, and AMD's September 28, 2026 announcement of its agreement to acquire World Labs. Outside sources include Nvidia's CUDA page, the PyTorch, vLLM, SGLang and FlashAttention project pages, MLCommons' MLPerf results and SemiAnalysis's InferenceX reports. The rest are posts from Microsoft, Amazon Web Services and Meta, CRN, Computer Graphics World, Stock Analysis, Nasdaq's stock market schedule and BloFin's AMDUSDT notices and liquidation rules, current as of October 1, 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 ROCm adoption figures here are AMD's own counts, and outside benchmark results can reverse within weeks as software versions change, so one release or one report says little about how many GPUs AMD will sell. Software news often breaks outside regular weekday trading hours, and in those hours BloFin may accept only orders that cut or close an AMDUSDT position. A leveraged position that moves against you then can fall to its maintenance margin and be reduced or liquidated before you can add margin. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.