Embench-IoT Results on GaZmusino

Sep 8, 2025 2 min read

The performance of a microcontroller or RISC-V core can be measured in many ways, but Embench-IoT has become a reference benchmark for embedded systems and IoT applications. In this post, we compare the performance of our core GaZmusino, evaluated via simulation with Verilator, against the results reported by Antmicro on their Embench-Tester platform.

Quick summary: GaZmusino is competitive in most benchmarks, but operations requiring heavy division show delays due to the lack of a hardware divider.

🔧 Methodology

The comparison was based on the following approaches:

GaZmusino: simulations performed using Verilator to obtain execution times for each benchmark.
Antmicro: public results available on the Embench Tester platform, reporting the performance of various open-source cores.
Compiler and flags: riscv64-unknown-elf-gcc 14.2.0 with -O2 optimization. Different versions may slightly affect execution times.

📊 Results

The table below is scrollable both horizontally and vertically, to allow viewing all benchmarks and cores.

Benchmark	GaZmusino	cv32e40p	CVA5	VexRISCV	serv	picorv32	Ibex	Rocket	microwatt
cubic	10	24	46	239	2	8	15	23	26
nbody	14	45	90	1597	2	14	39	52	108
st	18	43	86	907	2	11	21	45	62
minver	24	38	82	334	2	13	27	62	77
nsichneu	57	39	67	111	6	12	31	57	55
wikisort	65	65	166	242	14	23	33	85	98
crc32	77	41	92	92	25	15	40	72	68
nettle-aes	77	38	90	71	27	17	29	47	93
ud	77	36	79	73	11	15	27	59	65
aha-mont64	75	46	82	131	7	14	31	74	71
nettle-sha256	82	40	93	62	39	5	33	60	41
picojpeg	86	37	94	94	15	17	25	61	77
edn	89	42	90	90	2	10	24	77	52
matmult-int	90	39	90	91	3	13	21	62	44
primecount	112	60	93	127	9	18	36	81	85
sglib-combined	121	45	101	98	19	21	33	74	63
qrduino	123	54	115	113	21	22	35	79	80
slre	125	55	131	139	23	25	39	82	87
huffbench	126	46	115	119	44	28	35	96	76
md5sum	159	64	153	137	56	32	65	115	115
tarfind	228	38	143	77	16	20	33	88	70
statemate	298	52	181	203	19	24	40	101	95

Benchmarks ordered by GaZmusino performance

📝 Conclusions

Strengths of GaZmusino
- Very competitive in md5sum, qrduino, slre, huffbench, tarfind, and statemate.
- Consistently outperforms lightweight cores such as serv, picorv32, and Ibex.
Detected limitations
- Low performance in cubic, nbody, st, and minver.
- Main reason: absence of a hardware divider, divisions are handled in software, increasing latency.
Overall comparison
- Medium-high performance level, approaching more complex cores such as VexRISCV and CVA5.
- Solid in integer, memory, and logic workloads; weak in division-heavy tasks.