~/gpu/h100 vs mi300x

H100 80GB manufacturerH100 80GBvsMI300X manufacturerMI300X

Head-to-head for local LLM inference. The honest comparison: VRAM, bandwidth, compute, and which of the 30 catalog models actually fit on each.

The specs.

$ diff specs h100 mi300x
Stat
h100
mi300x
Δ
VRAM
80 GB
192 GB
+140%
Memory bandwidth
3,350 GB/s
5,300 GB/s
+58%
FP16 compute
989 TFLOPS
1300 TFLOPS
+31%
Weights budget at 8K ctx
63 GB
154 GB
+144%

Model fit difference.

$ models that change with the card
Fits on both
27of 30
Only on h100
0
Only on mi300x
0

// showing 12 of 30 models; differing fits first

Model
h100
mi300x
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsQ6_K
fitsFP16/BF16
overQ4_K_M
overQ4_K_M
fitsFP16/BF16
fitsFP16/BF16
fitsQ8_0
fitsFP16/BF16
fitsQ8_0
fitsFP16/BF16
fitsQ6_K
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16
fitsFP16/BF16

Which one wins for…

$ ./recommend --by-workload
More VRAM headroom

MI300X has 112 GB more.

Faster decode (bandwidth)

MI300X by +58%.

Faster prefill (compute)

MI300X by +31% TFLOPS.

Catalog models that fit

Tied: 27 of 30 fit on each.

Drill into either card.

$ ./vrambudget --gpu

Discussion.

$ gh discussion list

// sign in with github to leave a comment. threads live in the repo's discussions tab.