Test report DSG-2566 · Rev A · tested November 18, 2026

Processors & AcceleratorsDevice under test

CUDA 13.5 vs ROCm 10.1: 20 Years Meets a Six-Week Clock

CUDA 13.5 hit GA November 4 with Vera Rubin support; AMD claims MI455X beats it by up to 30% in MLPerf — all vendor numbers, no independent test.

Read
5 min
Words
952
Node
10nm
Operator
Amara Osei

Spec summary

  1. CUDA Toolkit 13.5 reached general availability November 4, 2026, with production-ready Vera Rubin support.
  2. AMD claims the MI455X reached 21,333 tokens/s per GPU on DeepSeek-R1, up to 30% above NVIDIA's published Vera Rubin result.
  3. NVIDIA reported $107.6 billion Data Center revenue in fiscal Q3 2027 versus AMD's $8.5 billion in Q3 2026.
  4. AMD ships ROCm roughly every six weeks; ROCm 10.0 arrived August 27, 2026.
  5. Jensen Huang resigned as NVIDIA CEO on November 12, 2026; Colette Kress became interim CEO.
CUDA 13.5 vs ROCm 10.1: 20 Years vs 6-Week Cycles [2026]
Fig. ACUDA 13.5 vs ROCm 10.1: 20 Years vs 6-Week Cycles [2026] — AI-generated

NVIDIA shipped CUDA Toolkit 13.5 to general availability on November 4, 2026, with production-ready support for its Vera Rubin architecture; eight days later AMD launched the Instinct MI450X and claimed its MI455X beats Vera Rubin on a published MLPerf inference result by up to 30%. Both numbers come from the vendors selling the hardware. No independent, same-workload test of ROCm against CUDA exists in the reviewed record.

The core question for GPU buyers is durability: which software stack will still serve them in 2028. CUDA turned 20 on November 16, according to NVIDIA material dated that day. ROCm 10.0 arrived on August 27 as the first major version change since the 7.x series, and AMD says ROCm 10 and its minor releases will land roughly every six weeks. ROCm 10.1 followed on October 5 with data-movement improvements, per AMD's ROCm blog.

What shipped, and when?

The last 90 days compressed a year of news:

  • August 18: AMD announced the ROCm Certification Program for developers and platform engineers.
  • October 27: AMD and Meta announced an expanded partnership; AMD delivered Day-0 support for Meta's Muse Spark models on Instinct and Muse Glimmer on Ryzen AI PCs and Radeon GPUs.
  • November 3: AMD reported $8.5 billion in Q3 2026 Data Center revenue, up 97% year over year, crediting EPYC demand and the MI450 launch.
  • November 4: CUDA 13.5 reached GA with Vera Rubin support.
  • November 12: AMD introduced the MI450X for Helios. The same day, Jensen Huang resigned immediately as NVIDIA president, CEO and board member; Colette Kress became interim CEO while keeping the CFO role, and Harvey Jones became the company's first-ever board chair.
  • November 16: NVIDIA announced Rubin HPC, published its first Vera Rubin MLPerf Training and Inference results, and marked 20 years of CUDA. AMD's Lux supercomputer opened to scientific users at Oak Ridge National Laboratory.
  • November 17: AMD completed its acquisition of Taalas, an AI-inference silicon company, planning to fold the technology into its accelerator roadmap.

No NVIDIA material reviewed ties the leadership change to any toolkit roadmap shift.

What do the benchmark claims actually prove?

AMD reported that an MI455X running in the Helios rack hit 21,333 tokens per second per GPU on DeepSeek-R1 in an AMD-tested MLPerf Inference Offline configuration — up to 30% above NVIDIA's latest published Vera Rubin result, by AMD's own comparison. AMD also claims up to 20% higher HBM bandwidth for Helios versus Vera Rubin.

NVIDIA, on November 16–17, claimed 10 times the throughput per watt on certain agentic-inference workloads and described Vera Rubin as delivering leading performance across training, post-training and agentic inference — in its own submission. MLCommons published MLPerf Training v6.1 on November 16, including Llama 2 70B LoRA, Llama 3.1 8B, Flux.1 and the newly added DLRM v4. The vendors measured different things on different days; the 30% and the 10x cannot be netted against each other.

Where does each stack run today?

Seven deployments from the past month show real trust:

  • Meta runs Instinct GPUs in production for training, ranking and recommendation, and plans a custom Instinct GPU.
  • Oak Ridge's Lux, built by HPE from EPYC CPUs, MI355X GPUs, Pensando networking and ROCm, opened to scientists November 16.
  • AMD's blog documents GLM-5.3-Flash serving on one eight-GPU MI300X node: a ~306 GB FP8 checkpoint inside 1,536 GB of HBM (192 GB per GPU), leaving most memory for KV cache.
  • AMD Life Science 26.08, MONAILabel 0.8.5 and hipCIM 26.06 package medical imaging on Instinct.
  • NVIDIA's DGX Spark HPC Developer Kit, introduced November 16, links up to four systems for 512 GB of shared memory.
  • CUDA 13.5 extends Windows-on-Arm to DGX Station for Windows — and drops macOS support entirely.
  • NVIDIA announced an expansion of Korea's national AI factory with Naver and Brookfield; scale unconfirmed.

What does the money say?

NVIDIA reported fiscal Q3 2027 total revenue of $115.8 billion, up 103%, with $107.6 billion from Data Center, up 110%. AMD's comparable segment figure is $8.5 billion. The quarters do not align exactly and AMD's segment includes EPYC processors, but an order-of-magnitude revenue gap explains why CUDA's library depth — CUDA Python cuda.core 1.2, CCCL 3.5, CUDA Tile, Nsight — is hard to match. AMD's column reads as platform plumbing: TheRock build system (in production since ROCm 7.14), GPU Operator v1.5.0 with DRA support, the certification program, and Day-0 support for PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM and Triton on MI400.

AMD's Anush Elangovan, Corporate Vice President of AI Software and Solutions, told The Register in July 2026: "For every generation of AMD GPUs, we have published not just the ISA spec. We actually publish the machine-readable ISA." The bet: AI coding models reading that spec close the kernel gap faster than human engineers.

What is the verdict?

Maturity still cuts both ways. NVIDIA's own November 5 notes document a cuBLASLt grouped-GEMM bug producing incorrect results on Blackwell and Rubin GPUs; a six-week ROCm cadence means more version bumps and regression testing per year. Neither vendor publishes prices for its newest datacenter hardware — quotation only.

The decision rule: choose CUDA 13.5 for ecosystem depth, hiring ease and one toolchain from desk to rack; choose ROCm 10 when memory capacity — 288 GB of HBM4 on the MI450X — vendor diversification or negotiating leverage matters, and code lives at framework level.

Every performance claim on the record is vendor-submitted; independent proof belongs to nobody yet.

via shattered.io (Source)

Filed under

  • cuda-13-5
  • rocm-10-1
  • nvidia-vera-rubin
  • amd-instinct-mi450x
  • mlperf
Share this article:

More from Amara Osei

Amara Osei

Show full bio

Staff writer covering money, funding and market shifts around hardware layer of artificial intelligence at Die Signal.

698 articles

Same lot · LOT-362E

« Previous articleNext article »