DeepX has taped out its second-generation LPDDR5X PIM chip. The company says it supports models with up to 100B parameters and consumes less than 5W during inference. The central question is not whether those specifications sound impressive, but whether independent parties can reproduce them.
Running 100B Models at 5W: What DeepX Actually Claims
According to EETimes, DeepX has taped out its second-generation LPDDR5X PIM chip. The vendor’s headline figures are striking: a 6.7× improvement over the previous generation, support for models with as many as 100B parameters, and inference power consumption below 5W.
The meaning of “supports models” needs to be unpacked. A 100B-parameter model represents an upper bound on model size, not a guarantee that a single chip can process the entire model at a usable speed.
At 4 bits per parameter, 100B weights alone require roughly 50GB of storage; at 8 bits, they require about 100GB. Adding the KV cache, runtime workspace, and the additional overhead created by long contexts pushes actual requirements even higher.
As a result, “supports 100B” and “completes useful inference within 5W” are separated by four major hurdles: capacity, bandwidth, latency, and accuracy.
The sub-5W claim also depends on the test boundary. Does it measure the bare chip or the complete system, including the package, memory, host processor, and cooling? Does it measure sustained generation or a short forward pass?
Without clarity on those conditions, 5W remains an unreproducible number.
Data Movement, Not Model Size, Determines Energy Efficiency
Traditional accelerators generally follow this data path:
LPDDR / HBM → on-chip cache → compute units
As the model processes each layer, weights may move back and forth along this path. Computation itself is not necessarily the largest energy consumer; moving large volumes of bits often is.
PIM, or Processing in Memory, places computational capability near or inside memory. The goal is straightforward: move less data and spend less time waiting for it.
DeepX’s new chip combines LPDDR5X with a Processing in Memory architecture to reduce data movement and energy consumption during model inference. For memory-bound generation tasks, this could deliver lower power use.
PIM is not unique to DeepX. Samsung, SK hynix, and other memory vendors have conducted relevant research for years. DeepX is combining LPDDR5X with its own compute architecture and focusing on edge inference rather than ultra-high-capacity HBM designs.
The strategy is sound, but its commercial viability will depend on shipping products and independent measurements.
A 6.7× Improvement Requires More Rigorous Comparisons
The “6.7×” figure is easy to share but unsuitable for direct comparison without more context.
Which chip generation serves as the baseline? Was the same model used? What was the precision—INT4, INT8, or FP16? Was the batch size 1 or 32?
Latency and throughput are also distinct metrics:
- Latency measures how long it takes to produce one token.
- Throughput measures how many tokens can be processed per second.
- Peak TOPS represents the chip’s theoretical compute ceiling.
LLM inference is often constrained by memory bandwidth rather than peak compute capacity. A chip with twice the TOPS may not deliver twice the end-to-end performance. If it reduces weight movement, energy per token may improve much more dramatically than its TOPS figure suggests.
The following data matter more:
| Metric | Why it must be disclosed |
|---|---|
| Model and precision | Determines data volume and whether results are comparable |
| Prompt and output lengths | Determines prefill and decode workloads |
| Batch size | Determines the tradeoff between throughput and latency |
| TTFT and TPOT | Reflect time to first token and sustained generation speed |
| Memory capacity and bandwidth | Shows whether the model can actually fit |
| Chip area and yield | Indicates whether the cost model is viable |
| Compiler and operator coverage | Determines whether real-world models can run |
If the evaluation uses a proprietary small model instead of a Qwen, Llama, or DeepSeek model, even impressive results only demonstrate that the chip can run that particular small model.
A more convincing evaluation would also disclose the model size, quantization method, power-measurement boundaries, and raw runtime logs.
Bringing 100B Models to the Edge Still Runs Into the Memory and Software Wall
A 5W power budget suits devices at the edge, but it does not automatically make them usable.
First, there is the memory wall. Although LPDDR5X capacity will grow, consumer devices typically reserve far less than 50GB for their NPUs. Even where theoretical capacity is sufficient, packaging, channel counts, and cost would constrain configurations.
Second, there is memory bandwidth. Even with a 100B model quantized to 4 bits, each token requires reading a large volume of weights. If sustained bandwidth at 5W is insufficient, the model may technically run but remain too slow to have product value.
Third, software remains a challenge. Developers will not rewrite an entire inference stack for a niche NPU.
At minimum, the platform needs:
- Support for mainstream operators and quantization formats
- A mature graph compiler
- Integration with llama.cpp, vLLM, or other inference engines
- Well-defined operator fallback behavior
- Reproducible profiling tools
If the compiler falls back most nodes to the CPU, a low chip-level power specification means little. System-level power must include an external CPU, memory, and any accelerated data copies.
Fourth, thermal design matters. Although 5W sounds modest, it is a budget for a sustained workload. Phones, robots, and industrial cameras must also leave headroom for the main SoC, display, wireless communications, and battery.
The more useful question is therefore not “Can it run 100B?” but “How many tokens per second can a real class of models generate within a 5W system budget?”
Wait for Production Hardware Before Judging Edge Models
The news also reached the Hacker News front page, where it received 40 upvotes and three comments.
For people following edge AI, that level of interest is substantial. Most of the Hacker News discussion, however, concerns the direction and potential of the technology rather than purchasing decisions.
“Taped out” means only that the design has entered the manufacturing process. Before volume production, there are still hurdles involving yield, packaging and testing, firmware, the compiler, drivers, and full-system validation.
Three developments would justify confidence:
- Third parties can obtain retail or sample chips rather than relying solely on a vendor demonstration.
- Public models run in the same software environment, accompanied by reproducible benchmarks.
- DeepX provides pricing, power limits, and a usable inference toolchain.
Until then, “100B, 5W, and 6.7×” are best viewed as three technical targets rather than finalized product specifications.
If those targets hold up under third-party testing, competition in edge LLMs will shift from asking who has the highest NPU TOPS to asking who can integrate the model, memory, compiler, and system-level power consumption correctly.
Editorial Perspective
The most significant aspect of DeepX’s new chip is not the “100B” label but its attempt to turn PIM’s efficiency gains from academic results into a path to mass-produced edge hardware. LPDDR5X is more closely aligned with real devices than HBM, while its capacity and cost fit the requirements of robots, cameras, and industrial systems more effectively. If the 5W target survives independent testing, its impact could exceed that of a single benchmark improvement.
The outlook is cautiously optimistic. The logic behind PIM’s reduction in data movement is sound, but the “6.7×” figure likely combines architectural changes, a new memory generation, and software optimizations. It cannot be attributed to the new chip alone.
What distinguishes this announcement from previous rounds of “large edge model” launches is the simultaneous presentation of two demanding targets: very large model capacity and very low power consumption, rather than a single benchmark score.
The uncertainties are equally clear. DeepX has not specified whether its 100B configuration keeps all weights resident in memory or allows tiled loading, nor whether the 5W figure covers the chip or the complete system. Those details could materially change the conclusion. The next milestone is not another launch event but a production chip that others can test independently.