The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
There is not enough verifiable evidence to say that an Axera AX8860 beats NVIDIA’s DGX Spark in large language model (LLM) performance by 25%. The AX8860’s identity and the associated claims—128 GB of LPDDR5x memory, 273 GB/s bandwidth, 40 W power, and a four-module PCIe card exceeding 1,000 TOPS—are not substantiated by the sources available. NVIDIA does publish DGX Spark specifications, but they are not a head-to-head LLM benchmark.
What is verified—and what is not?
NVIDIA’s official DGX Spark specifications list 128 GB of LPDDR5x unified system memory, 273 GB/s memory bandwidth, up to 1 PFLOP of FP4 tensor performance using sparsity, a 140 W GB10 system-on-chip (SoC) thermal design power (TDP), and a 240 W power supply. These are distinct specifications, not interchangeable measures of LLM speed or power use. NVIDIA’s DGX Spark specifications and its hardware overview do not establish an AX8860 comparison.
The AX8860 figures in the headline are not confirmed by an AX8860 datasheet, product page, or benchmark methodology among the available sources. That leaves the claims unresolved; it does not prove that no such product exists. In particular, the matching 128 GB and 273 GB/s figures are published for DGX Spark, not verified here for an AX8860.
| Claim or specification | What the available evidence supports |
|---|---|
| DGX Spark memory | NVIDIA specifies 128 GB LPDDR5x unified memory. |
| DGX Spark bandwidth | NVIDIA specifies 273 GB/s. |
| DGX Spark compute | NVIDIA lists up to 1 PFLOP FP4 tensor performance with sparsity; its user guide also lists up to 1,000 TOPS inference. |
| DGX Spark power | NVIDIA specifies a 140 W GB10 SoC TDP and a 240 W power supply. These describe different things. |
| AX8860 memory, bandwidth, power, and performance | Not verified by an AX8860 product source or reproducible comparison in the available evidence. |
| Four-module AX8860 PCIe card | Its configuration, performance, and availability are not established. |
Why 273 GB/s or a TOPS figure cannot prove an LLM speed advantage
Memory bandwidth describes how quickly data can move between memory and processors. It can affect performance, but it does not by itself determine how quickly a system generates LLM responses. Likewise, TOPS or FLOPS figures describe theoretical compute capability under a particular precision and methodology. NVIDIA’s 1 PFLOP FP4 figure is qualified with sparsity; it should not be read as a measured tokens-per-second result.
#1 Best Overall
A meaningful claim that one system is “25% faster” needs a direct, reproducible comparison. At minimum, it should disclose the model and version, quantization or precision, prompt and output lengths, batch size, decoding settings, software and runtime versions, and the metric being compared. Time to first token, generation tokens per second, and end-to-end latency are different measures. No such AX8860-versus-DGX-Spark test details are established here.
Why the 40 W claim needs a defined measurement boundary
Power figures are comparable only when they cover the same boundary. A chip or SoC figure, an accelerator-card figure, and the power draw of a complete system are not equivalent. NVIDIA’s 140 W figure is the GB10 SoC TDP; 240 W is the power supply capacity. Neither can be compared fairly with an unspecified 40 W AX8860 claim without knowing what was measured and under what workload.
Check the company and product names carefully
Axera Semiconductor and Axelera AI are separate companies. Their similar names can cause product and benchmark claims to be conflated, but Axelera AI’s accelerator pages are not evidence for an Axera AX8860.
Rank #2
- 900-5G172-2260-000
Axera Semiconductor: a documented AX8850 product
Axera Semiconductor’s October 17, 2025 announcement describes an AX8850-based M.2 2242/2280 computing card with power consumption under 8 W, intended for edge and terminal AI deployments. It says the card supports architectures including DeepSeek, Qwen, and Llama, and describes scaling from an M.2 card to a PCIe card. This is an AX8850 product, not verification of AX8860 specifications or the headline’s comparison. Axera’s AX8850 announcement provides the company’s description.
A separate M5Stack LLM-8850 Card specification identifies its SoC as Axera AX8850 and lists an M.2 M-key 2242 form factor with PCIe 2.0 x2. It is corroboration for an adjacent AX8850 card, not a source for AX8860.
Axelera AI: separate products and vendor-reported results
Axelera AI’s Edge products page describes Edge 130p and Edge 232p PCIe cards, advertising up to 214 TOPS and 629 TOPS, respectively. The page reports 83.4 tokens per second for Qwen 30B on Edge 232p. Axelera labels its comparison data as internal results against publicly available competitor data and says competitor data were current as of August 2026. These vendor-reported results concern Axelera products, not an Axera AX8860 or a DGX Spark head-to-head test.
Rank #3
- Extreme AI Performance: Powered by NVIDIA GB10 Grace Blackwell Superchip delivering 1 petaFLOP of AI performance and 128GB memory for 200B model fine-tuning.
- Developer-Optimized Platform: Designed for AI developers building secure, long-running agentic workflows, with compatibility across frameworks such as OpenClaw and NemoClaw, supporting private on-device inference, sandboxed execution, and governed data access.
- Scalable Architecture: Featuring NVIDIA NVLink-C2C for ultra-fast CPU-GPU memory communication and NVIDIA ConnectX-7 networking to support dual GX10 system stacking, unlocking superior scalability and performance.
- Advanced Thermal Design: Engineered cooling ensures sustained high performance and reliability in an ultra-small form factor.
- Full Stack AI Solution: The GB10 and NVIDIA AI software stack provide a full stack solution for AI development and deployment.
Axelera AI’s corporate site describes a Server 150p with four Europa AIPUs and advertises 2,500+ TOPS. It also reports 83.4 tokens per second for Qwen3-30B-A3B against a stated competitor result of 61 tokens per second. Those are distinct Axelera claims and do not verify the headline’s four-module AX8860 card or its claimed advantage over DGX Spark.
What would substantiate the AX8860 comparison?
Before treating the headline’s performance figures as established, look for an exact AX8860 SKU and manufacturer datasheet, plus a complete test report. A useful comparison would identify:
- The product’s availability and exact configuration, including whether the four-module PCIe card is shipping, announced, or only a proposed design.
- Memory type, organization, capacity, and bandwidth, along with whether memory is shared or dedicated.
- Compute precision and whether advertised TOPS or FLOPS use sparsity.
- LLM results for the same model and version, quantization, prompts, output lengths, batch size, decoding settings, software stack, and metric on both systems.
- Power at a clearly stated boundary—chip, card, or complete system—under a disclosed workload.
- Software framework and runtime support, host requirements, form factor, and price.
Without those details, the 25% advantage is not a reproducible result. It should be treated as an unverified claim, not a demonstrated product comparison.
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.
