tl;dr: Thor’s 20 SMs with tcgen05+tmem beat the Spark’s 48 SMs on tensor
core GEMM. The Spark has 3x the CUDA cores. Similar power under load,
completely different architectures despite both being “Blackwell.”
specs
DGX Spark (GB10) Jetson Thor
---------------- -----------
GPU SM 12.1 SM 11.0
SMs ~48 20
clock DVFS 208-3003 MHz DVFS 315-1575 MHz
tensor cores mma.sync tcgen05.mma + tmem
smem/SM 99 KB 228 KB
CPU 20 cores 14 cores
RAM 128 GB unified 128 GB unified
CUDA 13.2 / 580.142 13.2 / 580.00
Both Blackwell, both aarch64, both 128GB unified. But the Spark’s SMs are
cut down – less shared memory, no tensor memory (tmem), older mma.sync
instead of tcgen05. The Thor has fewer SMs but each one is the (almost) full design.
sustained tensor cores (the number that matters)
cuBLAS FP16->FP32, 4 streams, 30 seconds continuous per size. Both at
DVFS auto, no tuning, stock config.
N Spark TF Thor TF delta Spark W Thor W
---- -------- ------- ----- ------- ------
1024 52.7 74.3 +41% 72 79
2048 77.7 78.3 tie 90 78
3072 81.0 99.5 +23% 86 100
4096 83.9 80.2 -4% 85 86
5120 91.5 104.0 +14% 85 90
6144 84.4 98.6 +17% 85 88
8192 91.9 88.5 -4% 84 88
Thor wins most sizes. Spark wins at 4096 (tile quantization, more below)
and 8192 (more SMs help at scale). Power is similar – both land around
85-100W sustained. The Thor hits 104 TF from 20 SMs. The Spark maxes at
92 TF from 48 SMs. Twice the silicon, less tensor throughput.
isolating the architecture: same clock speed
Locked the Spark to 1575 MHz (Thor’s clock). Removes everything except
per-SM tensor efficiency.
N Spark TF Thor TF Thor/Spark
---- -------- ------- ----------
1024 44.4 74.3 1.7x
2048 60.3 78.3 1.3x
3072 67.9 99.5 1.5x
5120 71.3 104.0 1.5x
8192 73.5 88.5 1.2x
Per SM at 1575 MHz:
Thor: 5.2 TFLOPS/SM
Spark: 1.5 TFLOPS/SM (3.3x less)
tcgen05 + tmem + 228KB smem = 3.3x more tensor work per SM per clock.
The Spark compensates with 2.4x more SMs and higher clocks. Gets close
but never ahead on tensor.
Spark at 1575 is interesting on its own: 73.5 TF at 40W and 67C. Running
tensor cores way below their thermal ceiling. Quiet and efficient.
proving it’s the ISA, not cuBLAS tuning
Wrote the same WMMA kernel (mma.sync + shared memory tiling) for both:
N Spark TF Thor TF ratio
---- -------- ------- -----
1024 10.68 2.98 Spark 3.6x
4096 8.55 2.34 Spark 3.7x
8192 8.14 2.29 Spark 3.6x
Same code, Spark 3.6x. SM count wins when both run the same instruction.
cuBLAS flips it because it dispatches tcgen05.mma on the Thor.
tile quantization: 20 SMs = spiky performance
Thor’s numbers are lumpy. Burst benchmarks (not sustained) make it obvious
N Spark TF Thor TF
---- -------- -------
3584 82 117
4096 81 68 (-42%)
4352 87 117
4864 92 52 (floor)
5120 90 137 (ceiling)
52 to 137 TF, 2.6x swing. Tiles don’t divide evenly across 20 SMs at
certain sizes. Spark with 48 SMs stays in a 76-92 band. Under sustained
load the dips smooth out but don’t disappear. Real model layers will hit
unlucky dimensions.
clock scaling (Spark)
Sustained tensor at each setting:
MHz TFLOPS Watts TF/W Temp
---- ------ ----- ---- ----
1575 73.5 40W 1.84 67C
2000 80.9 42W 1.93 68C
2200 83.5 69W 1.21 70C
DVFS 85.9 88W 0.98 80C
2600 86.0 87W 0.98 81C
Saturation hits hard above 2200. Going from 2000 to 2600 buys 6% tensor
throughput for 2x the power draw. The firmware governor settles around
2200-2400 under sustained load anyway.
2200 is the sweet spot for inference serving: 97% of peak, 20W less,
10C more thermal headroom. Locked clocks also kill DVFS ramp-up jitter
on cold requests.
For power-constrained setups, 2000 MHz at 42W is hard to argue with.
1.93 TF/W is double the efficiency of running wide open.
CPU cores: don’t matter for tensor
Offlined cores on both machines. Thor tensor performance is identical
from 14 cores down to 1. tcgen05+tmem is fully autonomous.
Spark shows ~8% sensitivity on naive CUDA core kernels at large N (CPU
helps manage unified memory traffic) but cuBLAS tensor is flat.
thermals
Spark Thor
idle: 39C / 11W 36C / 8W
sustained: 79C / 85W 76C / 88W
limit: 90C plenty
Both comfortable. Spark runs 3C hotter with 11C to the limit. Not a
problem unless your room is 30C+.
On the Thor, jetson_clocks (locked 1575) sits at a flat 74-78C. DVFS
auto swings 56-76C – 18C cooler at light load, same at heavy. Pick
predictable or pick cool idle.
Spark setup and WiFi (MT7925)
First boot gotcha: the Spark ships with the system clock wrong and WiFi
doesn’t connect until the clock is right (WPA3 auth fails with bad
timestamps). You need to hold ESC during boot to get into the UEFI
menu, set the correct time, then plug in ethernet for initial setup.
WiFi works after the first apt update + reboot.
Once WiFi is up, the MediaTek MT7925 WiFi 7 adapter has a firmware bug
where 6 GHz MLD connections silently hang. No disconnect event, no
logs, just gone. Lost the machine multiple times running headless before
figuring it out. On a mesh network it also kept roaming to a far AP
that had enough signal to associate but not enough to complete DHCP –
NetworkManager would sit there for 5 minutes before retrying.
What fixed it:
- pin to 2.4 GHz (nmcli connection modify … wifi.band bg)
- disable PCI ASPM (modprobe mt7925e disable_aspm=1)
- disable wifi power save
- on a mesh network, pin the BSSID to the nearest AP
Stable since.
tuning script: GitHub - eous/dgx-spark-tune · GitHub
