4× DGX Spark vs M5 Ultra: Why Higher Bandwidth Mac Loses on GLM-5.3-Flash
A developer's benchmark shows four NVIDIA DGX Spark units interconnected via 200Gb/s networking outperform a single 512GB Mac Studio M5 Ultra on GLM-5.3-Flash inference across prefill, code generation, and concurrent workloads, despite the Mac's higher 1.23TB/s memory bandwidth, due to software stack differences, KV cache compression, and batching efficiency.
Origin of the Test
Overseas AI infrastructure developer Wei (@wei.wang) posted a hands-on benchmark on November 14 comparing two hardware configurations running the same GLM-5.3-Flash model (320B total parameters, 18B active, 1M context):
4× NVIDIA DGX Spark (GB10) : each with Grace Blackwell superchip, 128GB LPDDR5x unified memory, 273 GB/s memory bandwidth, ConnectX-7 200 Gb/s NIC. Four units connected in a ring via 200Gb/s cables, model sharded with TensorFold; expert weights 4-bit EXL3 quantized, KV cache stored as FP8, plus a DFlash2 small model for speculative decoding, GPU frequency locked at 2200 MHz.
1× Mac Studio M5 Ultra (512GB) : Apple M5 Ultra, 512GB unified memory, 1.23 TB/s bandwidth (official tier ~1.2 TB/s). Running MLX inference, weights mostly 4-bit (effective ~4.6 bit/expert with scaling), KV cache uncompressed, using model's built-in MTP head for draft tokens.
Notably, the four Sparks' aggregate bandwidth is 4×273 = 1.09 TB/s , 12% lower than the Mac's 1.23 TB/s, and traffic must traverse the network. Conventional wisdom suggests the Mac should win; the results show the opposite.
Speed Benchmarks: GLM-5.3-Flash (2K→1M Context)
All tests used identical prompts. Three metrics, all won by the Spark cluster:
Prefill (prompt ingestion) : at 128K context, Spark 46 seconds to first token vs Mac 60 seconds; at 1M context, Spark 7.2 minutes vs Mac 10 minutes.
Code generation throughput : at 2K context, Spark 85 tok/s vs Mac 52 tok/s; at 1M context, Spark 71 tok/s vs Mac 39 tok/s — gap widens with context length , Mac nearly 2× slower at 1M.
4 concurrent users each sending 8K prompts : aggregate code throughput Spark 109 tok/s vs Mac 59 tok/s — largest disparity.
Why the Higher-Bandwidth Mac Is Slower
The author breaks down per-step latency and tokens-per-step. For plain text generation both produce ~2 tokens/step, so the difference lies entirely in per-step latency: at 2K context 29ms (Spark) vs 41ms (Mac); at 1M context 34ms vs 56ms. Three root causes:
Heavier weights : accounting for MLX scaling factors, Mac expert parameters average 4.6 bits vs Spark's EXL3 4.0 bits , with hundreds of tensors kept at 5–8 bits. More data to move per step.
Uncompressed KV cache : Mac does not compress KV cache, so each cached token requires ~ 2× the bytes of Spark's FP8 KV. Consequently, per-step latency grows ~ 40% from 2K to 1M on Mac, versus only 10–17% on Spark.
Immature concurrent batching : with 4 simultaneous users, Mac total throughput rises only 24–30% over single-user, while Spark scales 71–80% . The single-box Mac is visibly dragged down in multi-user sharing.
Prefill is different — it is compute-bound . Four Blackwell GPUs' aggregate compute exceeds one M5 Ultra, so above 32K context Mac's prefill speed is only 72–79% of the Spark ring.
Author's personal data point: on two Sparks (GPU locked at 1800 MHz) he measured 52.2 tok/s single-stream, 104.6 tok/s aggregate for GLM-5.3-Flash, but methodology differs so not directly comparable.
The author's restrained conclusion: single-user generation gap is primarily software-layer (draft model swap, KV compression, batching improvements can narrow it); but prefill + multi-user sharing is a compute-layer gap that same budget cannot close short-term. Mac's advantage: one box, no networking, runs 1M context with ~268GB memory headroom.
Decision Guide: Four Reader Profiles
Need speed + multi-user sharing : 4× Spark fits better. Similar budget, aggregate bandwidth near Mac, but wins prefill and concurrency — suitable for small-team local inference server.
Want single-box simplicity : 1× M5 Ultra wins. One machine, offline, finishes 1M context with 268GB left, zero hassle — ideal for personal creation/privacy.
Already own a Mac, won't add hardware : Stick with MLX . Single-box MLX already runs GLM-5.3-Flash; no need to buy four boxes for this speed delta.
Budget-sensitive, testing waters : Start with 1× Spark . Official list price ~ $3,999/unit (four ~$16k), far cheaper than maxed Mac Studio, much lower entry barrier.
Honest Boundaries: Don't Take Everything at Face Value
Speed only : no evaluation of answer quality or power draw; Spark side benefited from weeks of extra tuning + two unpublished patches — comparison not fully fair.
Pricing from tweet, may differ from official : tweet claims Mac US $25,799, four GX10 $24,000; but DGX Spark official ~$3,999/unit (four ~$16k), M5 Ultra 512GB China retail ¥229,249. Screenshot's Chinese translation misstates Mac price as $9,999 and GPU frequency as 100GHz — rely on English original and vendor specs. Prices here are order-of-magnitude only.
Model specs per official sources : GLM-5.3-Flash is 320B total / 18B active MoE (not flagship 744B/40B GLM-5.3), native multimodal, 1M context, MIT license; FP8 weights ~328GB, 4-bit fits in 512GB Mac.
Bandwidth figures : 273 GB/s (DGX Spark official), 1.23 TB/s (tweet, official tier ~1.2 TB/s) — both credible.
Individual test, not vendor-endorsed ; replicate with same model in your workflow before committing.
Author's Take
This benchmark punctures a layer of "build philosophy" for local LLMs: bandwidth isn't the sole answer; software stack and interconnect efficiency decide . Mac wins on integrated experience; NVIDIA wins on stackability, tunability, mature community tooling. Whether you want "one box does it all" or "stack small boxes for performance" determines your bet.
Don't let "4 boxes beat 1 box" headlines drive you — that premise assumes same budget, heavy concurrency . For solo creative work, the M5 Ultra's 268GB headroom and zero-friction setup still smell sweet.
Original Tweet & Images (Source)
Source tweet screenshot from @wei.wang, November 14, ~2.8K views / 20 likes.
Image caption (transcribed visible text only, no performance extrapolation): Tweet compares 4× NVIDIA DGX Spark (GB10, 128GB/273GB/s each, 200Gb/s ring + TensorFold) vs 1× 512GB Mac Studio M5 Ultra (1.23TB/s, MLX), both running GLM-5.3-Flash (2K→1M context). Three speed metrics all favor Spark; conclusion: single-user gap mainly software, prefill + multi-user gap is compute. Screenshot low-res; price/frequency numbers per English original and vendor specs.
Data sources: benchmark data from Wei (@wei.wang) 11-14 post; DGX Spark specs (GB10, 128GB, 273GB/s, ConnectX-7 200Gb/s) from NVIDIA official spec sheet; GLM-5.3-Flash specs (320B/18B, 1M context, MIT, 2026-08-26 release) from Zhipu official docs and model card; M5 Ultra 512GB specs and China pricing from Apple official and historical archives. This is technical interpretation, not vendor info; final authority rests with official releases.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
