AMD’s COMPUTEX 2024 keynote previewed the Instinct MI325X as a challenger to NVIDIA’s H200, but the numbers changed between AMD’s June preview and its October product announcement. In October, AMD specified 256 GB of HBM3E and 6.0 TB/s of memory bandwidth for MI325X, and claimed advantages over H200 in memory, theoretical compute, and selected inference tests. Those performance comparisons are AMD-reported results, not an independent head-to-head review.
What AMD announced, and when
AMD chair and CEO Lisa Su delivered the COMPUTEX 2024 opening keynote on June 2. AMD used the event to preview MI325X as part of an expanded annual cadence for Instinct accelerators. Its June announcement projected up to 288 GB of HBM3E and Q4 2024 availability; a separate roadmap release projected 6 TB/s of memory bandwidth. The memory figures were projections based on specifications and estimates available at the time. AMD’s June 2 announcement and roadmap release provide the dated context.
On October 10, 2024, AMD announced MI325X with 256 GB of HBM3E and 6.0 TB/s of memory bandwidth. That is the later, specified configuration and should not be conflated with the June projection of up to 288 GB. AMD’s October release also set out its comparisons with H200. Read AMD’s October MI325X announcement.
MI325X vs. H200: published specifications
The figures below reflect AMD’s October 2024 product announcement, which cited H200 specifications for comparison. They are vendor-published figures rather than independent measurements.
#1 Best Overall
- HP Q1K38A AMD Radeon Instinct MI25 - GPU Computing Processor - Radeon Instinct MI25-16 GB HBM2 - for ProLiant XL270d Gen9
| Comparison | AMD Instinct MI325X | NVIDIA H200, as cited by AMD | AMD’s comparison |
|---|---|---|---|
| Memory capacity | 256 GB HBM3E | 141 GB | MI325X: 1.8× the capacity |
| Memory bandwidth | 6.0 TB/s | 4.8 TB/s | MI325X: 1.3× the bandwidth |
| Peak theoretical FP16 compute | AMD did not give a standalone value in the cited comparison | AMD did not give a standalone value in the cited comparison | AMD claimed 1.3× H200’s peak theoretical performance |
| Peak theoretical FP8 compute | AMD did not give a standalone value in the cited comparison | AMD did not give a standalone value in the cited comparison | AMD claimed 1.3× H200’s peak theoretical performance |
The theoretical compute ratios describe peak performance, not the speed a user should expect from every model or application. Memory capacity and bandwidth are more directly comparable specifications, but by themselves they do not establish which accelerator will perform better in a particular system.
AMD’s inference comparisons
AMD also published model- and precision-specific inference comparisons against H200. Its October 2024 announcement reported:
- Up to 1.3× on Mistral 7B at FP16.
- 1.2× on Llama 3.1 70B at FP8.
- 1.4× on Mixtral 8x7B at FP16.
These are AMD’s results for the named models and precisions; they should not be generalized to other workloads or deployments. AMD’s disclosed Llama 3.1 70B test used 2,048 input tokens and 2,048 output tokens. Its MI325X system ran vLLM, while the H200 system ran TensorRT-LLM.
How to interpret the benchmark setup
The comparisons were not a matched independent review. AMD disclosed a 1,000 W MI325X reference platform with an AMD Ryzen 9 7950X CPU, Ubuntu 22.04, and ROCm 6.3 prerelease. Its H200 comparison platform was a Supermicro system with accelerators rated at 700 W, Ubuntu 22.04, and CUDA 12.6. AMD’s test notes say results may vary with server manufacturer, software version, drivers, and optimizations. AMD’s announcement includes the disclosed test conditions and notes.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Differences in accelerator power ratings, platform configuration, and software frameworks matter when interpreting the reported ratios. The figures are useful as AMD’s account of selected tests, but they do not establish a universal performance ranking or predict results on a customer’s own system.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Availability: a dated forecast, not a stock check
In October 2024, AMD said MI325X production shipments were on track for Q4 2024 and forecast broad system availability beginning in Q1 2025 through providers including Dell Technologies, Eviden, Gigabyte, Hewlett Packard Enterprise, Lenovo, and Supermicro. That was a forward-looking company statement at the time; it does not verify current inventory, delivery dates, configurations, or pricing.
What COMPUTEX adds to the comparison
NVIDIA’s June 2 COMPUTEX announcement focused on Blackwell-powered systems and data-center infrastructure for cloud, on-premises, embedded, and edge deployments. That explains the broader event context, but it does not validate or independently confirm AMD’s MI325X-versus-H200 performance claims. NVIDIA’s COMPUTEX 2024 announcement describes its event focus.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Recommended Free Tools
