AI Hardware Diversification: The Cross Vendor Abstraction Layer for NVIDIA and AMD

AI Hardware Diversification: The Cross Vendor Abstraction Layer for NVIDIA and AMD

Prepared by Michael Kimball, The Innovation Attorney

July 7, 2026

Executive Summary

Software abstraction layers including PyTorch, Triton, and vLLM run one AI model on NVIDIA or AMD hardware today, and April 2026 MLPerf results show AMD MI355X matching NVIDIA B200. The technical possibility is settled: the question for buyers is cost and payoff. For inference workloads built on PyTorch and vLLM, qualifying a second GPU vendor is an engineering task measured in weeks to a few months, and measured production data show a cost near $22 per million tokens on the AMD Instinct MI300X against $28 on the NVIDIA H100, a savings of roughly 21 percent. Building an independent cross vendor platform from scratch is a different order of expense: Modular raised approximately $380 million over four years to build its Mojo language and MAX engine, reached a $1.6 billion valuation in September 2025, and drew a reported acquisition offer near $3.9 billion from Qualcomm in June 2026.

The MLPerf Inference 6.0 results published in April 2026 recorded the MI355X tying the B200 on Llama 2 70B offline throughput, reaching 97 percent on server workloads, and exceeding the B200 by 19 percent on interactive serving. The commercial consequences arrived before the benchmarks: OpenAI committed to 6 gigawatts of AMD Instinct GPUs in October 2025 alongside 10 gigawatts of NVIDIA systems, and Meta signed its own 6 gigawatt AMD agreement in February 2026. The most significant implication for buyers is negotiating position: a credible second source converts GPU procurement from allocation requests into price competition, and the qualification cost is recovered many times over in the first supply negotiation that follows.

Definition, Scope, and Technical Context

A hardware abstraction layer for AI is systems software that separates the definition of a model from the instruction set of the processor that executes it, so the same model runs on processors from more than one vendor without a rewrite. The layer can sit at several depths in the stack, and the depth determines both the cost of adoption and the performance retained.

This analysis covers datacenter training and inference of large models on NVIDIA GPUs running the CUDA platform, AMD Instinct accelerators running the ROCm platform, and secondarily on Google TPU, AWS Trainium, and Intel Gaudi hardware. Consumer, workstation, and edge deployment are excluded because their software stacks and economics differ materially. The engineering classification is systems software: compilers, runtimes, kernel libraries, and serving engines.

Six distinct technical approaches exist as of mid 2026. Framework level portability runs models through PyTorch or JAX, which dispatch to vendor backends. Compiler level portability writes kernels once in the Triton language or in SYCL and compiles them to each vendor instruction set. Serving level portability runs models through inference engines such as vLLM, SGLang, and ONNX Runtime that implement multiple hardware backends behind one interface. Source translation converts CUDA C++ code to the HIP dialect using the HIPIFY tools, producing code that compiles for both vendors. Binary and compile time translation, represented by ZLUDA and SCALE, executes or recompiles CUDA programs for AMD hardware directly. The sixth approach rebuilds the stack from the kernel language upward, represented by the Mojo language and the MAX inference engine, which deploy one codebase to NVIDIA, AMD, and Apple silicon.

Underlying Scientific and Engineering Principles

Both vendors implement the same computational model: single instruction multiple thread execution over thousands of parallel arithmetic units fed by high bandwidth memory. The differences are contingent engineering choices rather than physics. NVIDIA schedules threads in warps of 32; AMD CDNA hardware schedules wavefronts of 64. NVIDIA tensor cores execute matrix multiply accumulate instructions; AMD matrix cores execute the MFMA instruction family. NVIDIA compiles through the PTX virtual instruction set to device binaries; AMD compiles through LLVM to the AMDGCN instruction set and hsaco binaries. Cache hierarchies, shared memory sizes, and interconnect topologies differ in detail.

The compilation chain explains why portability is achievable. A model written in PyTorch is captured as a computation graph, lowered by the TorchInductor compiler into Triton kernels, translated to LLVM intermediate representation, and only then committed to a vendor instruction set: PTX for NVIDIA or AMDGCN for AMD. Because vendor commitment happens at the final stage, one source program serves both targets. Triton demonstrates this in production: PyTorch, Unsloth, and FlagGems generate kernels through it for NVIDIA, AMD, and Intel backends.

Two constraints do not yield to any abstraction layer. First, performance tuning is hardware specific: tile sizes, occupancy targets, and memory access patterns that saturate one chip stall another, so peak efficiency always requires per target tuning even when correctness is portable. Second, numerical results are not bitwise identical across vendors because floating point reduction order differs; validation must therefore be statistical rather than exact, comparing accuracy on evaluation sets rather than checksums. Neither constraint is fundamental in the physical sense, but both are permanent features of heterogeneous computing that budget and schedule must absorb.

System Architecture and Process Flow

The portable stack has five layers. At the top sits the model definition, almost always Python code against a tensor library; industry surveys placed PyTorch above 55 percent of production model share in the third quarter of 2025. Below it sits the graph compiler, TorchInductor in the PyTorch case, which fuses operations and emits Triton kernels. The kernel layer compiles through LLVM to the active vendor target. The runtime layer manages memory, streams, and kernel launches: the CUDA runtime on NVIDIA and the HIP runtime under ROCm on AMD, with intentionally mirrored APIs. The bottom layer is the driver and the collective communication library, NCCL on NVIDIA and RCCL on AMD, which implement the all reduce and all gather operations that multinode training and inference depend on.

Serving engines add a sixth operational layer. vLLM implements a hardware plugin interface with backends for NVIDIA, AMD ROCm, Intel XPU and Gaudi, Google TPU, AWS Trainium, ARM CPUs, and IBM Z, and its TPU backend now unifies JAX and PyTorch model definitions under a single lowering path. As of mid January 2026, 93 percent of the vLLM test groups on AMD hardware pass with daily regression maintenance, with full parity the stated goal.

The translation route serves legacy CUDA codebases. The HIPIFY tools convert CUDA source to HIP source: when the 15,000 line HACC cosmology code was ported, 95 percent of the code converted automatically or passed unmodified. SCALE, released by Spectral Compute, recompiles unmodified CUDA source directly for AMD targets, with version 1.4.2 shipping in October 2025. ZLUDA translates compiled CUDA binaries at load time; it runs against ROCm 7 as of December 2025 but lost its second round of commercial funding and returned to single maintainer status in 2026. The inputs that matter for this technology are not materials but people and test hardware: compiler engineers, kernel authors, and a continuous integration fleet spanning both vendors.

Quantitative Performance Metrics

The table below consolidates the measured figures most relevant to a dual vendor decision. MLPerf figures are independently verified submissions; pricing figures are market observations that vary by volume and contract.

MetricAMDNVIDIA
Llama 2 70B offline throughput, MLPerf Inference 6.0, April 2026MI355X: tieB200: baseline
Llama 2 70B server throughput, MLPerf Inference 6.0MI355X: 97 percent of B200B200: baseline
Llama 2 70B interactive serving, MLPerf Inference 6.0MI355X: 119 percent of B200B200: baseline
GPT OSS 120B offline and server, MLPerf Inference 6.0MI355X: 111 and 115 percent of B200B200: baseline
HBM capacity and bandwidth, prior generationMI300X: 192 GB, 5.3 TB per secondH100: 80 GB, 3.35 TB per second
Typical list price per accelerator, prior generationMI300X: near $15,000H100: $27,000 to $40,000
Measured inference cost per million tokens, batch 1MI300X: $22.22H100: $28.11
Cloud rental, median observed 2026MI300X: $2.72 per hourH100: $1.99 to $4.12 per hour

Three qualifications apply. First, AMD cluster scale submissions exceeded one million tokens per second on Llama 2 70B, demonstrating that the results hold beyond single nodes. Second, several widely repeated figures are vendor reported rather than independently verified: AMD claims up to 1.3 times B200 throughput on Llama 3.1 405B under FP4 quantization in vLLM, and Modular claims up to 53 percent better prefill throughput and 32 percent better decode throughput for its MAX engine; both should be treated as vendor benchmarks until reproduced. Third, community measurements of ROCm on PyTorch and vLLM place general throughput at 90 to 95 percent of equivalent NVIDIA hardware for standard workloads, with the gap widening on workloads that depend on CUDA only libraries such as TensorRT LLM and FlashAttention 3, which have no complete ROCm equivalents. All figures reflect published conditions; production results vary with model, quantization, and batch profile.

Key Advantages and Limitations

The principal advantage is purchasing power. A buyer with one qualified supplier requests allocation; a buyer with two qualified suppliers runs a price competition. The mechanism is visible in market pricing: MI300X rental rates run 15 to 30 percent below comparable H100 rates, and the memory advantage compounds the effect because a single 192 GB MI300X serves models that require two 80 GB H100 units, halving node count for large model inference. Early adopters report per token costs roughly 30 percent below equivalent H100 deployments.

The second advantage is supply resilience. GPU allocation remains the binding constraint on AI capacity growth, and the October 2025 OpenAI commitments across NVIDIA, AMD, and Broadcom custom accelerators, roughly 33 gigawatts in total, show the largest buyer in the market treating multi vendor capability as an operating requirement rather than an experiment.

The limitations are concentrated in three places. Training at frontier cluster sizes remains more mature on NVIDIA because NCCL, the interconnect stack, and years of production hardening at 10,000 GPU and larger scales have no fully equivalent AMD track record yet. Workloads written against CUDA only libraries carry real porting cost, and teams that hand tuned PTX level kernels lose that work entirely on a second vendor. Operationally, a dual vendor fleet doubles the continuous integration surface: every model release must be validated twice, and driver or runtime regressions arrive on two independent schedules. These limitations are contingent rather than intrinsic; each reflects software maturity and staffing rather than any property of the silicon.

Technical Failure-Mode Analysis

Failure modes for an abstraction layer are software failures, and they cluster into five patterns. Performance cliff failures occur when a kernel that saturates one target falls to a fraction of peak on the other because tile sizes or memory access patterns mismatch the hardware; the consequence is degraded throughput rather than incorrect output, and the mitigation is per target autotuning plus continuous benchmark gates in the deployment pipeline. Numerical divergence failures occur because reduction order and fused operation behavior differ across compilers; consequences range from harmless last digit noise to accuracy regressions in quantized models, and the mitigation is statistical acceptance testing against golden evaluation sets on every backend before promotion.

Dependency drift failures are the most frequent in practice: a new model architecture ships with CUDA support on day one while the ROCm path waits for kernel coverage, or a ROCm release regresses a previously passing operator. The vLLM project addresses this with daily regression tracking on AMD hardware, and the 93 percent pass rate recorded in January 2026 is the current measure of residual exposure. Collective communication failures appear only at multinode scale, where NCCL and RCCL differ in topology handling and fault behavior; they present as hangs or throughput collapse during distributed jobs, and they are the main reason training portability trails inference portability.

Translation layer failures deserve separate treatment. Binary translation through ZLUDA depends on undocumented behavior of compiled CUDA artifacts, so any upstream toolchain change can break it, and the project has twice lost commercial sponsorship, leaving no support path for production incidents. This is a cascade risk: a technical failure with no vendor accountable converts directly into an operational outage. Source level routes through HIP, Triton, or native ROCm carry no such dependency. Detection across all five patterns follows the same discipline: two vendor continuous integration, canary deployments, and benchmark regression alarms, budgeted as a permanent engineering cost rather than a one time project.

Real-World Applications and Deployment Context

Dual vendor AI infrastructure moved from pilot to production commitments during late 2025 and early 2026. OpenAI announced a 6 gigawatt AMD Instinct agreement on October 6, 2025, beginning with 1 gigawatt of MI450 systems in the second half of 2026, structured with a warrant for up to 160 million AMD shares vesting against deployment milestones; the same month it committed 10 gigawatts of NVIDIA systems and roughly 10 gigawatts of Broadcom custom accelerators. Meta expanded its AMD agreement to 6 gigawatts on February 24, 2026. Microsoft has run production GPT model inference on Azure MI300X fleets since 2024, which remains the clearest public demonstration that frontier model serving is portable in practice.

In high performance computing, the El Capitan system at Lawrence Livermore National Laboratory reached the top of the Top500 list in November 2024 on AMD MI300A processors, evidence that mission critical national laboratory workloads run at full scale outside the CUDA stack. In commercial serving, vLLM and SGLang deployments across NVIDIA, AMD, TPU, and Trainium backends constitute the mainstream path: most production inference operators now inherit portability from their serving engine rather than building it. The pre commercial frontier is large scale training on AMD hardware, where the OpenAI MI450 deployment beginning in the second half of 2026 will be the deciding demonstration. The threshold each application must meet is consistent: throughput per dollar at equal accuracy, validated on the operator’s own models rather than on published benchmarks.

Economic and Scalability Considerations

Adoption cost divides into three tiers. The first tier applies to operations already built on PyTorch and vLLM or SGLang: qualification of AMD hardware is a validation and tuning exercise, typically a team of three to ten engineers for one to two quarters, an estimated $500,000 to $5 million depending on model count and accuracy requirements. This figure is an engineering estimate rather than a published benchmark, and it excludes the permanent cost of dual vendor continuous integration, which adds test hardware and roughly one to three sustained engineering positions.

The second tier applies to codebases with custom CUDA kernels. The HIPIFY record suggests 95 percent automatic conversion on well structured code, with the residual manual effort concentrated in PTX assembly, CUDA intrinsics, and performance retuning; published porting experience ranges from hours for simple programs to weeks for complex ones, before performance work. The third tier is building an independent platform, and the market has now priced it: Modular raised approximately $380 million in venture funding over four years, reached a $1.6 billion valuation on a $250 million round in September 2025, and drew a reported $3.9 billion acquisition offer from Qualcomm in June 2026. Qualcomm is separately reported to be committing $14 billion to an open compiler and RISC-V accelerator effort. No buyer needs to replicate that spending; the open stack exists because it is already funded.

The return side is direct. Measured cost per million tokens of $22.22 against $28.11 is a 21 percent saving that compounds across every token served. At fleet size the negotiating effect dominates the arithmetic effect: a buyer committing $10 billion to accelerators that extracts even a 10 percent price concession through a credible second source recovers a $5 million qualification program two hundred times over. The bottlenecks to scaling this practice are people rather than capital: engineers with production ROCm experience remain scarce, kernel library gaps persist for the newest attention and quantization techniques, and every additional supported vendor multiplies the validation matrix. The AI accelerator market grew from roughly $55 billion in 2023 to an estimated $160 billion in 2025 and past $200 billion in 2026, with inference approaching two thirds of spending; because inference is where portability is mature, the addressable share of the market for dual vendor purchasing is the majority and growing.

Regulatory Environment and Compliance Considerations

Three legal domains govern this technology. The first is contract. The CUDA end user license agreement includes a clause prohibiting translation of output generated by the CUDA SDK to target non NVIDIA platforms. Binary translation projects sit squarely against that clause: ZLUDA was funded by AMD until AMD ended the arrangement and its legal department contested the open source release, and the project has since operated without commercial sponsorship. Recompilation approaches such as SCALE, which compile original CUDA source rather than translating SDK output, present a distinguishable posture, and clean room reimplementation of APIs draws support from Google v. Oracle, 593 U.S. 1, decided in 2021, which held API reimplementation to be fair use in the software context. The clause has not been tested in court against a translation layer, and that untested status is itself the deterrent: enterprises will not build production dependence on a legally contested path, which channels adoption toward the uncontested routes through PyTorch, Triton, HIP, and vLLM.

The second domain is trade control. The Department of Commerce published a rule on January 13, 2026, with Federal Register codification on January 15, 2026, permitting export of H200 and MI325X class accelerators to China subject to a 50 percent volume cap relative to United States shipments and a 25 percent tariff arrangement, replacing the broader diffusion framework proposal scrapped in May 2025. Export policy now shifts quarter to quarter, and multi vendor software insulates buyers from single vendor licensing disruptions because workloads can follow whichever silicon is exportable or available. The third domain is competition law: United States and European authorities have had inquiries into GPU market practices open since 2024, and bundling of hardware with proprietary software is a recurring subject of those inquiries. Standards activity runs alongside: the UXL Foundation, formed in September 2023, and the Khronos Group formalized a liaison in 2026 to align SYCL as the open cross vendor kernel standard. None of these frameworks requires portability, but procurement rules in regulated industries and federal acquisition practice have long favored second sources, and counsel drafting GPU supply agreements should treat portability as the technical foundation for every pricing, allocation, and termination lever in the contract.

Future Development Trajectory

The next 18 months contain the deciding milestones. The AMD MI450 generation, built on the CDNA 5 design at the TSMC 2 nanometer node, begins shipping into the first OpenAI gigawatt in the second half of 2026; delivery at volume and competitive training performance would validate the full dual vendor case, while slippage would confirm the view that AMD competes only in inference. On the software side, the measurable markers are vLLM reaching 100 percent test parity on AMD hardware, ROCm closing the attention kernel gap with FlashAttention 3, and the disposition of the reported Qualcomm acquisition of Modular, including whether the platform remains vendor neutral under new ownership.

Market projections, labeled as estimates, place NVIDIA near 80 percent of AI accelerator revenue in 2026 trending toward 75 percent as AMD and custom silicon grow, within a market exceeding $200 billion. Custom accelerators from Google, AWS, and Broadcom collectively take more share than AMD, and every one of them depends on the same portability layers this report describes, which makes the abstraction software the common beneficiary of every diversification path. The risk that cross vendor software fails to reach commercial scale is low for inference: it is already commercial, embedded in the serving engines most operators run. For frontier training the risk is moderate: collective communication maturity and cluster operations experience remain to be demonstrated on AMD at the largest scales, and the capital committed by OpenAI and Meta is the strongest available evidence that sophisticated buyers judge that risk acceptable.

Want more industry insights? I publish daily analysis on the intersection of deep tech and venture law.

Subscribe to The Innovation Attorney on Substack

Open Questions

Five questions will determine how far hardware diversification runs. First, does training portability arrive at frontier scale: RCCL and AMD cluster operations have no public track record at 50,000 GPU and larger jobs, and the OpenAI MI450 deployment through 2027 will supply the first real answer, which matters because training remains roughly one third of accelerator spending. Second, will the CUDA license clause against translation ever be enforced or adjudicated: a ruling either way would reset the legal boundary for every compatibility project, and until one arrives enterprises will keep avoiding binary translation regardless of its technical merit. Third, does Qualcomm keep Modular vendor neutral: a $3.9 billion acquisition by a chip vendor could either fund the strongest independent layer or fold it into a single vendor stack, and the answer will be visible in roadmap and licensing decisions within a year of closing. Fourth, does quantization format fragmentation rebuild lock in one level up: FP4 and FP8 implementations differ across vendors, and if model weights become format captive the portability won at the kernel layer is lost at the checkpoint layer; standards work through 2027 will decide this. Fifth, who bears the permanent cost of dual stack validation: today it is absorbed by open source projects and hyperscaler engineering budgets, and whether mid market operators can afford the same discipline determines whether diversification stays a hyperscaler practice or becomes the industry default over the next five years.

Source List

1. MLCommons, MLPerf Inference v6.0 results, April 2026. https://mlcommons.org/2026/04/mlperf-inference-v6-0-results/. Independently verified benchmark data comparing MI355X and B200 across Llama 2 70B and GPT OSS 120B scenarios.

2. AMD press release, AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs, October 6, 2025. https://ir.amd.com/news-events/press-releases/detail/1260/. Primary source for the 6 gigawatt commitment, MI450 timing, and the 160 million share warrant structure.

3. AMD press release, AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs, February 24, 2026. https://www.amd.com/en/newsroom/press-releases/2026-2-24-amd-and-meta-announce-expanded-strategic-partnersh.html. Primary source for the Meta commitment.

4. AMD blog, AMD Delivers Breakthrough MLPerf Inference 6.0 Results, 2026. https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-inference-6-0-results.html. Vendor commentary and cluster scale token throughput figures; treated as vendor reported where not verified by MLCommons.

5. vLLM project blog, Beyond Porting: How vLLM Orchestrates High-Performance Inference on AMD ROCm, February 27, 2026. https://vllm.ai/blog/2026-02-27-rocm-attention-backend. Technical detail on the ROCm backend and the 93 percent test group pass rate.

6. ROCm documentation, Porting CUDA code to HIP. https://rocm.docs.amd.com/projects/HIP/en/latest/how-to/hip_porting_guide.html. Authoritative porting guidance and tool behavior.

7. ROCm blog and AMD GPUOpen, Application portability with HIP. https://rocm.blogs.amd.com/software-tools-optimization/hipify/README.html. Source of the HACC 15,000 line, 95 percent automatic conversion figure.

8. Triton distributed compiler paper, arXiv 2504.19442, 2025. https://arxiv.org/html/2504.19442v2. Peer reviewed description of Triton compilation to PTX and AMDGCN through LLVM.

9. UXL Foundation and Khronos Group, collaboration announcement on the SYCL open standard, 2026. https://www.khronos.org/blog/uxl-foundation-khronos-liaison-on-the-sycl-and-safety-critical-systems. Primary source for standards alignment.

10. Modular, Modular and AMD announcement and MAX platform documentation. https://www.modular.com/blog/modular-x-amd-unleashing-ai-performance-on-amd-gpus. Vendor source for single codebase deployment claims and reported throughput advantages.

11. CNBC, OpenAI looks to take 10 percent stake in AMD through AI chip deal, October 6, 2025. https://www.cnbc.com/2025/10/06/openai-amd-chip-deal-ai.html. Financial press coverage of the warrant economics.

12. SemiAnalysis, AMD vs NVIDIA Inference Benchmark: Performance and Cost Per Million Tokens. https://newsletter.semianalysis.com/p/amd-vs-nvidia-inference-benchmark-who-wins-performance-cost-per-million-tokens. Independent measurement source for per token cost comparisons.

13. Tom’s Hardware, CUDA emulator for AMD GPUs ZLUDA loses funding with v6 release, 2026. https://www.tomshardware.com/pc-components/gpu-drivers/. Reporting on ZLUDA sponsorship history and status.

14. Congressional Research Service, U.S. Export Controls and China: Advanced Semiconductors, report R48642. https://www.congress.gov/crs-product/R48642. Authoritative summary of the export control regime.

15. RunPod engineering blog, AMD MI300X vs NVIDIA H100 SXM: Performance Comparison on Mixtral 8x7B Inference. https://www.runpod.io/blog/amd-mi300x-vs-nvidia-h100-sxm-performance-comparison. Independent cloud operator benchmark and pricing data.

16. Computer Weekly, CUDA at 20: From billion dollar gamble to agentic AI, 2026. https://www.computerweekly.com/news/366640614/. Source for CUDA developer base figures cited at GTC 2026.

Bibliography

Advanced Micro Devices. 2025. AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs. October 6, 2025. https://ir.amd.com/news-events/press-releases/detail/1260/.

Advanced Micro Devices. 2026. AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs. February 24, 2026. https://www.amd.com/en/newsroom/press-releases/2026-2-24-amd-and-meta-announce-expanded-strategic-partnersh.html.

Advanced Micro Devices. 2026. AMD Delivers Breakthrough MLPerf Inference 6.0 Results. https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-inference-6-0-results.html.

Advanced Micro Devices. 2026. Porting CUDA Code to HIP. ROCm Documentation. https://rocm.docs.amd.com/projects/HIP/en/latest/how-to/hip_porting_guide.html.

Advanced Micro Devices. 2026. Application Portability with HIP. ROCm Blogs. https://rocm.blogs.amd.com/software-tools-optimization/hipify/README.html.

CNBC. 2025. OpenAI Looks to Take 10 Percent Stake in AMD Through AI Chip Deal. October 6, 2025. https://www.cnbc.com/2025/10/06/openai-amd-chip-deal-ai.html.

Computer Weekly. 2026. CUDA at 20: From Billion Dollar Gamble to Agentic AI. https://www.computerweekly.com/news/366640614/.

Congressional Research Service. 2026. U.S. Export Controls and China: Advanced Semiconductors. R48642. https://www.congress.gov/crs-product/R48642.

Google LLC v. Oracle America, Inc., 593 U.S. 1 (2021).

Khronos Group. 2026. UXL Foundation and Khronos Collaborate on the SYCL Open Standard. https://www.khronos.org/blog/uxl-foundation-khronos-liaison-on-the-sycl-and-safety-critical-systems.

MLCommons. 2026. MLPerf Inference v6.0 Results. April 2026. https://mlcommons.org/2026/04/mlperf-inference-v6-0-results/.

Modular. 2026. Modular and AMD: Unleashing AI Performance on AMD GPUs. https://www.modular.com/blog/modular-x-amd-unleashing-ai-performance-on-amd-gpus.

RunPod. 2026. AMD MI300X vs NVIDIA H100 SXM: Performance Comparison on Mixtral 8x7B Inference. https://www.runpod.io/blog/amd-mi300x-vs-nvidia-h100-sxm-performance-comparison.

SemiAnalysis. 2026. AMD vs NVIDIA Inference Benchmark: Performance and Cost Per Million Tokens. https://newsletter.semianalysis.com/p/amd-vs-nvidia-inference-benchmark-who-wins-performance-cost-per-million-tokens.

Tom’s Hardware. 2026. CUDA Emulator for AMD GPUs ZLUDA Loses Funding with v6 Release. https://www.tomshardware.com/pc-components/gpu-drivers/.

Triton Distributed Authors. 2025. Triton-Distributed: Programming Overlapping Kernels on Distributed AI Systems with the Triton Compiler. arXiv 2504.19442. https://arxiv.org/html/2504.19442v2.

vLLM Project. 2026. Beyond Porting: How vLLM Orchestrates High-Performance Inference on AMD ROCm. February 27, 2026. https://vllm.ai/blog/2026-02-27-rocm-attention-backend.

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/