<\/span><\/h3>\nWith GPT-OSS-120B, the HP ZGX Nano G1n posts its strongest ends in Prefill Heavy, the place throughput climbs from 304.5 tok\/s at batch 1 to 2773.3 tok\/s at batch 64. Equal ISL\/OSL additionally scales steadily, rising from 69.6 tok\/s to 722.9 tok\/s throughout the sweep. Decode Heavy is far lighter by comparability, beginning at 183.7 tok\/s in batch 1, dipping barely in batch 2, then recovering to 262.9 tok\/s by batch 64.<\/p>\n
\u00a0<\/p>\n
<\/span>GPT-OSS-20B<\/span><\/h3>\nWith GPT-OSS-20B, HP\u2019s highest numbers come from Prefill Heavy, however the scaling is much less linear than with the opposite fashions. Prefill begins at 1626.6 tok\/s at batch 1, climbs to 1980.3 tok\/s at batch 2, drops sharply to 1120.3 tok\/s at batch 4, then recovers to 4345.1 tok\/s by batch 64. Equal ISL\/OSL scales extra easily from 92.6 tok\/s to 1550.6 tok\/s, and Decode Heavy rises from 94.4 tok\/s to 670.4 tok\/s.<\/p>\n
<\/span>Qwen3 Coder 30B A3B FP8<\/span><\/h3>\nFor Qwen3 Coder 30B A3B (FP8), HP once more excels in Prefill Heavy, with throughput rising from 432.2 tok\/s at batch dimension 1 to 2069.4 tok\/s at batch dimension 64. Equal ISL\/OSL rises from 104.2 tok\/s to 1274.4 tok\/s, whereas Decode Heavy improves from 55.9 tok\/s to 480.4 tok\/s. That is amongst HP\u2019s stronger general outcomes.<\/p>\n
<\/span>Qwen3 Coder 30B A3B Base<\/span><\/h3>\nOn Qwen3 Coder 30B A3B (Base), HP delivers regular progress throughout all three phases, though the topline stays within the Prefill Heavy section. That section will increase from 258.6 tok\/s at batch 1 to 1629.4 tok\/s at batch 64. Equal ISL\/OSL scales from 60.3 tok\/s to 690.3 tok\/s, whereas Decode Heavy rises from 33.0 tok\/s to 331.8 tok\/s.<\/span>Llama 3.1 8B Instruct FP4<\/span><\/h3>\nWith Llama-3.1-8B-Instruct (FP4), HP reveals a transparent step up in throughput. Equal ISL\/OSL climbs from 76.4 tok\/s at batch 1 to 2774.1 tok\/s at batch 64, making it the strongest of HP\u2019s three phases on this mannequin. Prefill Heavy additionally scales aggressively, rising from 316.8 tok\/s to 2397.1 tok\/s at batch 32 earlier than slipping to 2270.4 tok\/s at batch 64. Decode Heavy will increase from 40.7 tok\/s to 547.6 tok\/s throughout the sweep.<\/p>\n
<\/span>Llama 3.1 8B Instruct (Base)<\/span><\/h3>\nOn Llama-3.1-8B-Instruct (Base), the HP ZGX Nano G1n scales cleanly throughout all three phases. In Equal ISL\/OSL, throughput rises from 28.2 tok\/s at batch 1 to 1298.6 tok\/s at batch 64. In Prefill Heavy, HP will increase from 123.2 tok\/s to 1759.5 tok\/s, with features remaining sturdy all through the sweep earlier than tapering barely on the high finish. Decode Heavy is far lighter by comparability, rising from 15.5 tok\/s at batch 1 to 366.4 tok\/s at batch 64.<\/p>\n
<\/span>GPU Direct Storage<\/span><\/h3>\n<\/span>How GPU Direct Storage Works<\/span><\/h4>\nHistorically, when a GPU processes information from an NVMe drive, the info should first go by the CPU and system reminiscence earlier than reaching the GPU. This course of creates bottlenecks as a result of the CPU acts as a intermediary, including latency and consuming system sources. GPU Direct Storage eliminates this inefficiency by permitting the GPU to entry information immediately from the storage machine over the PCIe bus. This direct path reduces information motion overhead, enabling sooner, extra environment friendly transfers.<\/p>\n
AI workloads, particularly these involving deep studying, are extremely data-intensive. Coaching giant neural networks requires processing terabytes of knowledge, and any delay in information switch results in underutilized GPUs and longer coaching occasions. Accordingly, GPU Direct Storage addresses this problem by delivering information to the GPU as rapidly as attainable, minimizing idle time and maximizing computational effectivity.<\/p>\n
As well as, GDS advantages workloads that stream giant datasets, similar to video processing, pure language processing, and real-time inference. By decreasing CPU reliance, GDS accelerates information motion and frees CPU sources for different duties, additional enhancing general system efficiency.<\/p>\n
<\/span>GDSIO Learn Throughput 16K<\/span><\/h3>\nTaking a look at GDSIO Learn Throughput 16K, the HP ZGX Nano G1n begins at 0.70GiB\/s with 1 thread, inserting it among the many stronger low-thread performers within the group. It dips to 0.41GiB\/s at 2 threads, then climbs again to 0.86GiB\/s at 4 threads, displaying the identical small early-thread inconsistency seen in a number of of those techniques. From there, scaling turns into far more constant. Throughput rises to 1.6GiB\/s at 8 threads and a pair of.2GiB\/s at 16 threads, then continues upward to three.0GiB\/s at 32 threads. On the greater queue depths, the HP retains gaining floor, reaching 3.9GiB\/s at 64 threads and peaking at 4.6GiB\/s at 128 threads.<\/p>\n
<\/span>GDSIO Learn Common Latency 16K<\/span><\/h3>\nTaking a look at GDSIO Learn Common Latency (16K), the HP ZGX Nano G1n begins at roughly 0.02ms with 1 thread and stays low by 2 threads (0.08ms) and 4 threads (0.07ms). Latency edges up barely at 8 threads (0.08ms) and 16 threads (0.11ms), then will increase extra noticeably at 32 threads (0.16ms) and 64 threads (0.25ms). At 128 threads, latency reaches 0.42ms, nonetheless a bit beneath the best ends in the group whereas monitoring the system\u2019s regular throughput scaling throughout the take a look at.<\/p>\n
<\/span>GDSIO Write Throughput 16K<\/span><\/h3>\nTaking a look at GDSIO Write Throughput 16K, the HP ZGX Nano G1n begins at 0.84GiB\/s on 1 thread, rises to 1.4GiB\/s on 2 threads, and reaches 2.2GiB\/s on 4 threads. Efficiency continues to scale strongly at 8 threads (3.0 GiB\/s) and reaches 3.3GiB\/s at 16 threads, the place it successfully ranges off. From there, throughput stays almost flat at 3.3GiB\/s with 32 and 64 threads, then eases barely to three.2GiB\/s with 128 threads, indicating the platform reaches its write ceiling comparatively early and sustains that degree persistently by the remainder of the sweep.<\/p>\n
<\/span>GDSIO Write Common Latency 16K<\/span><\/h3>\nTaking a look at GDSIO Write Common Latency (16K), the HP ZGX Nano G1n begins at roughly 0.02ms with 1 thread and stays very low by 2 threads (0.02ms) and 4 threads (0.03ms). Latency rises modestly at 8 threads (0.04ms) and 16 threads (0.07ms), then jumps at 32 threads (0.15ms) and 64 threads (0.30ms). At 128 threads, latency reaches 0.61ms, nonetheless pretty nicely managed general, although the upward pattern aligns with the purpose the place write throughput has already flattened at greater thread counts.<\/p>\n
<\/span>GDSIO Learn Throughput 1M<\/span><\/h3>\nTaking a look at GDSIO Learn Throughput 1M, the HP ZGX Nano G1n begins at 3.2GiB\/s on 1 thread and rises to 4.1GiB\/s on 2 threads. Efficiency continues to climb at 4 threads (5.2GiB\/s) and eight threads (5.5GiB\/s), after which the platform successfully reaches its ceiling. Throughput then holds basically flat at 5.5GiB\/s for 16, 32, and 64 threads, earlier than easing barely to five.3 GiB\/s at 128 threads, indicating a powerful early ramp adopted by a really secure high-thread plateau.<\/p>\n
<\/span>GDSIO Learn Common Latency 1M<\/span><\/h3>\nTaking a look at GDSIO Learn Common Latency (1M), the HP ZGX Nano G1n begins at roughly 0.31ms with 1 thread and stays comparatively low at 2 threads (0.47ms) and 4 threads (0.76ms). Latency will increase with concurrency, rising to 1.4ms at 8 threads, 2.9ms at 16 threads, and 5.9ms at 32 threads. The pattern continues at 64 threads (12.8ms) and reaches 27.2ms at 128 threads, monitoring the upper queue depths although throughput had already flattened a lot earlier within the sweep.<\/p>\n
<\/span>GDSIO Write Throughput 1M<\/span><\/h3>\nTaking a look at GDSIO Write Throughput 1M, the HP ZGX Nano G1n begins at 3.1GiB\/s with 1 thread and rises to three.5GiB\/s with 2 threads, then holds that degree at 4, 8, and 16 threads. Efficiency dips barely to three.3GiB\/s at 32 threads earlier than returning to three.5GiB\/s at 64 threads. At 128 threads, throughput will increase to three.7GiB\/s, indicating a principally flat write profile throughout the sweep with solely minor variation and a small uptick on the highest thread rely.<\/p>\n
<\/span>GDSIO Write Common Latency 1M<\/span><\/h3>\nTaking a look at GDSIO Write Common Latency (1M), the HP ZGX Nano G1n begins at roughly 0.31ms with 1 thread, rising to 0.57ms with 2 threads and 1.1ms with 4 threads. Latency continues to climb as concurrency will increase, reaching 2.2ms with 8 threads, 4.4ms with 16 threads, and 9.4ms with 32 threads. The upward pattern continues at 64 threads (17.7ms) and reaches 37.3ms at 128 threads, reflecting steadily rising queue strain although write throughput itself stays pretty flat by a lot of the sweep.<\/p>\n
<\/span>Conclusion<\/span><\/h3>\nHP\u2019s ZGX Nano G1n carries the DGX Spark platform\u2019s anticipated efficiency profile and provides engineering selections that set it other than the opposite Spark techniques within the subject. In our testing, CPU temperatures peaked at 77.3\u00b0C and GPU temperatures at 69\u00b0C, each on the cooler facet of the Spark models we\u2019ve benchmarked. vLLM efficiency was strongest in Prefill Heavy workloads throughout all six fashions we examined, with scaling that held cleanly by greater batch sizes. GPU Direct Storage learn throughput reached 4.6 GiB\/s at 16K and 5.5 GiB\/s at 1M block sizes, and write throughput plateaued early however held that degree persistently throughout the remaining thread counts.<\/p>\n
The place the ZGX Nano G1n separates itself from the remainder of the Spark subject is within the work HP did across the reference design. The recycled-materials content material, the higher\/lower-chassis cut up that improves inner serviceability, and the acoustic envelope that holds at 27.6 dBA beneath load all replicate deliberate engineering selections past what the GB10 platform itself requires. The safety stack follows the identical sample. TPM 2.0 in FIPS 140-2 mode, Frequent Standards EAL4+, and SED OPAL storage push this unit previous a developer equipment and towards a system that may clear procurement in regulated environments.<\/p>\n
Like different Sparks, this isn’t a general-purpose workstation, and HP doesn’t place it as one. For builders, small groups, and organizations that want native AI compute with credible sustainability and safety tales behind the acquisition, the ZGX Nano G1n is a transparent differentiated choice inside the Spark lineup. For retailers the place these standards don’t apply, the underlying platform is the fixed throughout all 5 OEM techniques we\u2019ve reviewed, and the choice comes all the way down to ecosystem, assist, and worth.<\/p>\n
![\"HP](\"https:\/\/www.storagereview.com\/wp-content\/uploads\/2026\/04\/StorageReview-HP-ZGX-Nano-AI-Station-8.jpg\")
<\/p>\n