The RTX Spark Is Not an Apple Silicon Competitor | by Wes Brown | Jun, 2026 | MediumSitemapOpen in appSign up<br>Sign in

Medium Logo

Get app<br>Write

Search

Sign up<br>Sign in

The RTX Spark Is Not an Apple Silicon Competitor

Wes Brown

9 min read·<br>Jun 5, 2026

Listen

Share

Press enter or click to view image in full size

NVIDIA’s new chip is a capable CUDA machine. Comparing it to Apple’s M-series is the wrong comparison, and in this post, I explain why.

NVIDIA announced the RTX Spark at Computex 2026. Several outlets framed it as Windows finally getting its “Apple Silicon moment,” and a few went further and described it as NVIDIA beating Apple Silicon outright.<br>I think the RTX Spark is good engineering and will be a useful machine for the right buyer. The comparison to Apple’s M-series is wrong, though, and this post lays out why, with measured numbers where they exist and modeled estimates where they don’t.<br>What has actually shipped<br>Two products get conflated. The DGX Spark is the GB10 Grace Blackwell desktop running Linux. It has been shipping since around October 2025, and every measured NVIDIA number below comes from it. The RTX Spark is the consumer version: Windows N1X laptops and small desktops, announced at Computex to ship in fall 2026. It is not out. There are no independent shipping-laptop benchmarks yet.<br>The top-bin N1X is the same GB10 die in a lower power envelope, 45 to 80 watts of SoC power instead of the desktop’s 140. So the desktop numbers are a reasonable proxy for the laptop’s silicon. Every NVIDIA laptop figure here is a projection from a shipping desktop, and every Apple figure is a shipping, reviewed product. In the tables, measured values are marked M and derived values D.<br>The numbers<br>All inference figures are gpt-oss-120B in MXFP4, single user, short context, on each platform’s best mainstream runtime. The GB10 desktop row is measured; the N1X laptop rows are modeled from it.

Estimates and methodology. The D values are estimates, not measurements, and were produced as follows.<br>N1X laptop rows: the top N1X is the GB10 die run at 45 to 80 watts instead of 140. Single-core is unchanged, because one X925 core reaches its boost clock inside any of these envelopes. Multi-core is close to unconstrained at 80 watts, where the CPU complex draws roughly 50 to 65 watts with the GPU idle during the test, and is constrained at 45 watts. Prefill takes the compute underclock, roughly 15 to 25% at 80 watts. Decode is unchanged, because LPDDR5X bandwidth does not change with TDP.<br>Apple prefill estimates (M3 Max, M4 Pro, M4 Max, M5 Pro): scaled from the one measured Apple anchor, the M3 Ultra at about 1,000 t/s on 80 GPU cores, by GPU core count and per-core matmul throughput, with the M5 parts given the Neural Accelerator multiplier.<br>Apple decode estimates: scaled by memory bandwidth from the nearest measured anchor.<br>M5 Max prefill (footnote 4): Monte Carlo over the uncertain parameters, 200,000 draws, with the Neural Accelerator multiplier swept 3 to 4x, the GEMM share of prefill 0.60 to 0.85, MLX maturity 0.80 to 1.0, and plus or minus 10% on the anchor.<br>The largest single source of error is the M3 Ultra prefill anchor. It is one measurement, in 8-bit MLX, on a path with a known instability bug. If that number is off, the whole Apple prefill column shifts with it.<br>CPU performance<br>On shipping Linux, the GB10 scores about 3,120 single-core and 18,895 multi-core in Geekbench 6. Single-thread is at M3 level, which is late 2023. Apple has shipped M4 and M5 since, so the Spark is two generations back, about 22% behind the base M4 and 37% behind M5 on single-thread.<br>The cores are Arm Cortex-X925, current client cores co-designed with MediaTek, not the older Neoverse V2 in NVIDIA’s datacenter Grace, despite the shared name. The same X925 ships in phones. A Dimensity 9400 scores about 2,874 single-thread at 3.63 GHz, and the GB10 scores about 3,120 at 3.9 GHz. The per-clock performance is the same as the phone. The desktop’s only single-thread advantage is roughly 7% more clock for a much larger power budget. That is the ceiling of Arm’s best reference core, and it runs about 30 to 44% behind Apple’s custom cores.<br>Multi-core is better because there are 20 cores. About 18,900 puts it near an M3 Max and just under an M4 Pro. Apple reaches 26,675 on the M4 Max and 29,233 on the M5 Max with fewer cores.<br>Power and idle<br>The desktop GB10 is a 140-watt SoC. The laptop N1X bins at 45 to 80 watts. Idle power matters most for a laptop. The DGX Spark idles around 35 to 40 watts as a desktop. The ConnectX-7 NIC accounts for only single digits to about 10 watts of that at idle; the rest of the non-SoC budget is mostly USB-C device provisioning, not the NIC. Backing the NIC out leaves the GB10’s own idle floor in the mid-20s of watts, and that floor comes from the die, so it carries into the laptop.<br>Apple idles at 1 to 2 watts and games at 7 to 8 watts on a fanless M5 MacBook Air. That is a 15x to 25x idle difference....