amd

Bilingual

Running Kimi K3 on MI355X at Better Performance per Dollar Than B300
2 months ago
- AMD MI355X achieves ~952 tok/s/node for Kimi K3, outperforming B200 on performance per dollar due to lower cost and higher HBM capacity.
- Kimi K3 is a 2.8T-parameter open-source model that cannot fit on a single B200 node, requiring two nodes or a B300 node with 288GB VRAM per GPU.
- AMD MI355X offers comparable specs to B300 at ~2.4× lower cost per GPU, making it a cost-efficient alternative despite software challenges.
- Software issues on ROCm were resolved with simple fixes (e.g., adding top-k renorm function for speculative decoding, padding head count for MLA prefill), not requiring custom kernels.
- Speculative decoding improved single-stream throughput by ~2.2× and peak aggregate by +18%, while prefill optimizations cut time-to-first-token by 2–3×.
- The MI355X's large HBM capacity gives it a practical edge over B200 for models like Kimi K3 that require >1.5TB VRAM for weights and KV cache.
AMD Posts Linux Patches for HDMI 2.1 Auto Low-Latency Mode "ALLM" & VRR
2 months ago
- AMD has released Linux patches for HDMI 2.1 Auto Low-Latency Mode (ALLM) and Variable Refresh Rate (VRR) in the AMDGPU driver.
- HDMI ALLM allows the computer to trigger a display's 'Game Mode', reducing latency and lag for gaming.
- HDMI 2.1 VRR support is now added, complementing the existing DisplayPort VRR support for AMD Radeon graphics on Linux.
- The patches are under review and may or may not be merged into the upcoming Linux v7.3 kernel cycle.
Can AMD Break the CUDA Moat? AMD Advancing AI 2026
2 months ago
- AMD's software and hardware have significantly improved, shifting from a 0% to a high chance of success against Nvidia's CUDA moat, but key risks remain.
- Two major risks: 1) Helios rack production challenges due to cable design and Ethernet retimer requirements; 2) Lack of stable GPU clusters for internal software development and CI testing, slowing progress.
- AMD's MI455X silicon leads in logic and memory integration (2nm, 12 HBM4 stacks, 432GB), but microarchitecture and software still lag behind Nvidia's Rubin.
- AMD's software stack (ROCm, vLLM, SGLang) is improving rapidly, especially in single-node inference and documentation, but distributed inference (disaggregation, WideEP) needs more work.
- AI agents and open-source tools like ROCm.ai/GEAK are helping AMD close the software gap faster than traditional methods.
- AMD needs to prioritize stable CI/GPGPU clusters for internal teams, upstream distributed inference support, and ensure composability of optimizations across models.
- Anthropic and Microsoft/OpenAI have committed to AMD's MI455X Helios, with financial incentives like equity rebate discounts making TCO highly attractive.
- The CUDA moat is eroding as AI agents reduce the need for manual kernel engineering, benefiting AMD's open-source approach.
Instella-Moe: An Open Mixture-of-Experts Language Model
2 months ago
- AMD introduces Instella-MoE, a fully open Mixture-of-Experts language model with 16 billion total parameters and 2.8 billion active parameters per token.
- The model is trained from scratch on AMD Instinct MI300X and MI325X GPUs using the AMD ROCm software stack, incorporating architectural innovations like Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective connectivity.
- Instella-MoE undergoes a multi-stage training pipeline including pre-training, mid-training, long-context extension, supervised fine-tuning, direct preference optimization, and reinforcement learning, with all artifacts openly released.
- The model achieves competitive performance on benchmarks against dense and MoE models of comparable or larger active parameter counts, with the final base model averaging 76.7 and the post-trained Think model averaging 73.22.
- AMD releases model weights, training configurations, data mixtures, and code for all stages to promote reproducibility and collaboration in the open AI community.
AMD and Cerebras Launch AI Inference Solution
2 months ago
- AMD and Cerebras Systems announced a technical partnership to deliver a new disaggregated AI inference solution.
- The solution combines AMD Helios rackscale solutions with the Cerebras Wafer-Scale Engine for ultra-low-latency and high throughput.
- The joint solution is expected to deliver up to 5x higher tokens per second per watt (T/s/W) compared to a Cerebras WSE-only configuration.
- AMD Helios provides high-throughput prompt processing, while Cerebras WSE accelerates memory-bandwidth-intensive token generation with ultra-low latency.
- The solution addresses the growing diversity of AI inference workloads with heterogeneous infrastructure.
- Cerebras plans to deploy AMD Helios in its data centers, with availability through Cerebras Cloud in the second half of 2026.
AMD attacks the rack with Helios systems that rival Nvidia's
2 months ago
- AMD launched Helios, its first rack-scale AI platform, competing directly with Nvidia's Vera Rubin.
- Helios claims 50% more HBM4 memory, 15-25% higher AI training performance, and 30% better performance per dollar than Nvidia's system.
- The platform uses 72 MI455X GPUs with CDNA 5 architecture, featuring 2nm compute dies and 3D packaging.
- Helios matches Nvidia's NVL72 layout with 18 liquid-cooled blades, but uses Ultra Accelerator Link over Ethernet (UALoE) and Broadcom switches.
- Real-world performance may reach only half of peak theoretical FLOPS, but AMD claims it still leads the industry.
- AMD also announced MI440X for enterprise, MI430X for HPC, and Venice Epyc CPUs with up to 256 cores.
AMD calls CUDA a 'non-event'
2 months ago
- Nvidia's dominance in AI hardware is largely due to its CUDA software platform, but AMD claims CUDA is becoming irrelevant for many companies.
- AMD's executive says customers are now programming at higher levels of abstraction, bypassing CUDA entirely, and AI agents are helping optimize platforms without it.
- If CUDA's moat erodes, cheaper alternatives like AMD's ROCm could lure AI companies away from Nvidia hardware, threatening Nvidia's market lead.
- However, skepticism remains because Nvidia's established ecosystem and proven infrastructure may still be favored, especially for high-stakes AI data centers.
AMD Ryzen 7 7700X3D Review: 3D V-Cache Gaming Performance for Less – HotHardware
2 months ago
- The Ryzen 7 7700X3D is a more affordable 3D V-Cache processor for gamers, sharing an 8-core design with the 7800X3D but with lower clocks and a $329 MSRP.
- It offers a cost-effective entry to AM5 platform with future compatibility for next-generation Zen 6 processors.
- The chip features 96MB total L3 cache (including 64MB 3D V-Cache), which enhances gaming performance by keeping data closer to cores.
- Compared to the 7800X3D, the 7700X3D has a 200MHz lower base clock and 500MHz lower boost clock, likely reducing performance in lightly threaded tasks but less impact in multi-threaded workloads.
- It requires a BIOS/UEFI update for motherboard compatibility, but most board manufacturers have already released updates.
Scrying the AMD GFX1250 LLVM Tea Leaves
2 months ago
- AMD's new GFX1250 accelerator, part of MI455X for machine learning, shows similarities to RDNA4 including WGP structure and WMMA support, but lacks dynamic VGPR allocation.
- GFX1250 allows up to 1024 VGPRs per wave, a major improvement over CDNA's 512 and RDNA's 256, and features a merged 448KB WGP Cache for LDS and vector L0.
- Graphics support is minimal in GFX1250, with no rasterizer, texture, or raytracing instructions, making it a pure compute accelerator.
- Tensor operations in GFX1250 combine RDNA4's programming model with CDNA4's performance, supporting various data formats and sparse WMMA, but excluding fp64.
- New features include clusters for thread block coordination, cluster-level barriers and loads, cooperative atomics, and hardware support for tanh and explicit prefetching.
- Data dependency handling is redesigned with more granular counters (e.g., ASYNCcnt, TENSORcnt) to reduce stalls, improving efficiency over GCN/CDNA.
AMD pivots from HBM to LPDDR5X for new Versal Gen 2 memory on package chips
3 months ago
- AMD announces Versal Premium Gen 2 Memory on Package SoCs with on-package LPDDR5X for integrated, turn-key solutions.
- The pivot from HBM to LPDDR5X is driven by HBM supply shortages due to AI boom demand, leading to Versal HBM discontinuation.
- On-package memory strategy shift addresses long-term availability issues, as HBM has shorter production cycles than DDR technologies.

Running Kimi K3 on MI355X at Better Performance per Dollar Than B300

AMD Posts Linux Patches for HDMI 2.1 Auto Low-Latency Mode "ALLM" & VRR

Can AMD Break the CUDA Moat? AMD Advancing AI 2026

Instella-Moe: An Open Mixture-of-Experts Language Model

AMD and Cerebras Launch AI Inference Solution

AMD attacks the rack with Helios systems that rival Nvidia's

AMD calls CUDA a 'non-event'

AMD Ryzen 7 7700X3D Review: 3D V-Cache Gaming Performance for Less – HotHardware

Scrying the AMD GFX1250 LLVM Tea Leaves

AMD pivots from HBM to LPDDR5X for new Versal Gen 2 memory on package chips