A four-system NVIDIA DGX Spark setup is reported to produce 494 tokens per second on code with 32 concurrent requests. That is aggregate throughput across concurrent work—not the speed of one response. In the same report, single-request output is listed at about 96 tokens per second for code and 58 for prose. The figures are attributed to a setup disclosure shared by Patrick Moorhead and reported by Wccftech on October 5, 2026; the opened sources do not provide a reproducible protocol or independent validation of the 494 result.
What the 494 tokens-per-second figure means
The reported 494 tokens per second applies to code generation under a load of 32 concurrent requests. It describes the combined output of the serving setup while handling multiple requests, not the rate a single user should expect to see in one stream. Wccftech also reports about 280 tokens per second for prose at 32 concurrent requests, roughly 96 tokens per second for a single code request, and roughly 58 for a single prose request. These are reported figures, not independently reproduced measurements. Wccftech’s October 5, 2026 report attributes the setup and numbers to Moorhead’s social post.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Dell NVIDIA Tesla V100 GPU SXM2 32GB NWWWX by DELL | $854.96 | Buy on Amazon |
| 2 |
|
Gigabyte NVIDIA GeForce RTX 3060 Gaming OC V2 Graphics Card - 12GB GDDR6, 192-bit, PCI-E 4.0,... | $695.00 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
The article additionally lists approximately 4,764 tokens per second for prompt processing and about 0.2 seconds to first token while idle. Prompt processing and generated-token output measure different parts of inference, so the prompt-processing figure should not be compared directly with the 494 output rate. The idle first-token figure is also a reported setup result, not a guarantee for a busy system or longer prompts.
Why other four-Spark benchmarks show different speeds
A separate benchmark author, posting as tonyd615 on the NVIDIA Developer Forums on September 10, 2026, described a tuned four-Spark vLLM setup using tensor parallelism, DSpark speculative decoding, and CUDA graphs. That post reports 77.2 tokens per second peak on a single stream for counting, 52 on code in the benchmark, 72 on a warm code run, and 214 aggregate tokens per second at six streams, including 143 on code. The author also notes that 203 GB of Engram tables remain on disk. These figures provide context for how a particular serving configuration performed; they do not reproduce or disprove the 494-at-32-stream report. Read the forum benchmark post.
#1 Best Overall
- GPU Chipset: NVIDIA
- Memory: HBM2
- Programming Interface: CUDA
- Memory Capacity: 32GB
- Slot Compatibility: SXM2
The benchmark repository’s September 10 notes report another workload-specific set of results: one-stream decode at 73.8 tokens per second on code, 50.9 on math, 37.8 on reasoning, and 24.4 on prose. Across eight prompt categories, six streams produced 131.9 tokens per second aggregate; the repository also reports a peak aggregate of 225.5 tokens per second on code at six streams. Its author notes a GPU slow-state condition affecting one run. These figures underline that there is no single universal speed for this model on four DGX Sparks. See the benchmark repository notes.
Meaningful comparisons need to match the workload and serving conditions. At minimum, check:
- Whether the result is single-stream or aggregate, and how many requests ran concurrently.
- The task and prompt category, such as code, math, reasoning, or prose.
- Serving software and settings, including quantization, speculative decoding, and graph configuration.
- Context length and prompt-processing workload.
- Whether the number comes from a public, reproducible run or a secondary report of a social post.
What DeepSeek-V4.1-Flash is
DeepSeek’s September 2026 paper describes V4.1-Flash as a multimodal mixture-of-experts model with 552 billion backbone parameters and context lengths up to one million tokens. The paper says it activates 8 billion parameters per token during prefill and 16 billion during decode. The 552-billion figure is the model’s total backbone size; it does not mean all 552 billion parameters are active for each token. Read the DeepSeek model paper.
Free tools Windows power users keep installed
One-click scans. No signup required.
The authors report a global key-value (KV) cache footprint of 890 bytes per token, approximately one quarter of the corresponding DeepSeek-V4-Flash footprint. They attribute the reduction to cross-layer KV reuse in Compressed Sparse Attention 2 and FP4 KV caching, and describe SWA Bounded Replay as reducing persistent KV-cache requirements. These are architecture claims in the authors’ paper, not independent measurements of memory use on the four-Spark system.
Rank #2
- NVIDIA Ampere Streaming Multiprocessors: Building blocks for the world's fastest, most efficient GPUs, the all-new Ampere SM brings twice the FP32 throughput and improved energy efficiency
- 2nd Generation RT Cores - Experience 2x the 1st Generation RT Cores throughput, plus competitive RT and shading for a whole new level of ray-tracing performance
- 【3rd Generation Tensor Cores】Get up to 2X the throughput with structural sparsity and advanced AI algorithms such as DLSS
- Core Clock: 1837MHz
- WINDFORCE 3X Cooler
What four DGX Sparks require
NVIDIA lists each DGX Spark with a 20-core Arm CPU, up to 128 GB of coherent unified memory, 273 GB/s memory bandwidth, and a ConnectX-7 network interface rated at 200 Gbps. Its official specifications also list a 240-watt power supply. Four units therefore represent a multi-system cluster, not one workstation with 512 GB of ordinary shared memory: the reported setup involves serving software and tuning across separate machines. NVIDIA’s published specifications and buying options are on its DGX Spark product page.
The approximately 512 GB pooled-memory description belongs to the reported four-system setup; it should not be read as a single-node memory specification or as proof that every model component resides in pooled memory. The forum benchmark’s note that 203 GB of Engram tables remain on disk illustrates that large-model serving can involve storage and implementation details beyond the nominal combined memory figure.
How to judge the headline against another benchmark
The 494 figure is striking, but the available evidence supports a narrow reading: a reported code workload, 32 concurrent requests, and a particular four-Spark configuration. No reproducible protocol for that exact result is established in the cited materials. The separate forum and repository results use different workloads, stream counts, and serving details, so they are informative comparisons rather than direct checks of the headline number.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteFor a purchasing or deployment decision, first decide whether the target is low-latency single-user generation or high aggregate throughput under concurrency. Then seek benchmarks with the same task, concurrency, prompt lengths, software settings, and measurement method. A peak aggregate number alone cannot tell you how fast a single interactive coding session will feel.
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.
