Skip to content
State of GPU Compute 2026

Server compare

Supermicro SYS-822GS-NB3RT HGX vs ASUS ESC N8A-E12

A has 1,664 GB more GPU memory. B costs €24.00 an hour instead of €48.00.

ASupermicro · B300
Supermicro SYS-822GS-NB3RT HGX32 x 96GB RAM
GPU memory
288 GB × 82,304 GB
System memory
32 DIMMs · 32 x 96GB DDR5-6400 ECC RDIMM
€48.00/server·hr
Δ GPU memoryA+1,664 GBmore on Amore on this side
Δ RateB−€24.00less per hour on Bless per server·hr
Δ System memoryB+3,072 GBmore on Bmore on this side
BASUS · H100
ASUS ESC N8A-E12
GPU memory
80 GB × 8640 GB
System memory
6144GB (DDR5 ECC REG)
€24.00/server·hr
Change servers
A
B

Side by side

Every spec, line by line.

Values exactly as each builder lists them. Lines that differ are marked; identical lines are dimmed.

13 lines differ · model name and ID not counted · identical lines folded

ASupermicro SYS-822GS-NB3RT HGX · 32 x 96GB RAM
8× B300 · Supermicro
BASUS ESC N8A-E12
8× H100 · ASUS
Server
Model
Supermicro SYS-822GS-NB3RT HGX
ASUS ESC N8A-E12
Brand(differs)
Supermicro
ASUS
GPU(differs)
1 x NVIDIA HGX B300 NVL8 (8 GPU)
8 x NVIDIA H100
GPU family(differs)
B300
H100
CPU(differs)
2 x Intel Xeon 6768P (64C 336MB Cache 2.40 GHz)
Configurable on request
System memory(differs)
32 x 96GB DDR5-6400 ECC RDIMM
6144GB (DDR5 ECC REG)
Local storage(differs)
8 x 3.2TB NVMe PCIe Gen5 E1.S + 2 x 960GB NVMe PCIe 4.0 M.2
Configurable on request
Form factor(differs)
8U
7U
Model ID
supermicro-sys-822gs-nb3rt-hgx-3tb
asus-esc-n8a-e12
GPU
Architecture(differs)
NVIDIA Blackwell Ultra
NVIDIA Hopper
Memory per GPU(differs)
288 GB HBM3e
80 GB HBM3
GPU memory per server(differs)
2,304 GB
640 GB
GPU interconnect(same on both)
NVLink, all-to-all
NVLink, all-to-all
Rental
On-demand, full server(differs)
€48.00/hr
€24.00/hr
On-demand, per GPU(differs)
€6.00/hr
€3.00/hr
Reserved, 12 months(differs)
€40.00/hr
€16.00/hr
Billing(same on both)
Per second
Per second
Rates in EUR excluding VAT. Fields a builder leaves open are configured to order.

Cost and memory

Price the run before you start.

Both panels read the live price list. A and B mark where your two servers sit.

What a run costs.

Pick a plan, how many GPUs and how long. Totals use the same hourly rates as the price list.

Plan
GPUs
Run length
ASupermicro SYS-822GS-NB3RT HGX · 32 x 96GB RAM
€1,152
8× B300 · 24 hours
BASUS ESC N8A-E12
€576
8× H100 · 24 hours

B costs €576 less for this run (50% lower).

Same run on every family
  • B300A€1,152
  • B200€960
  • H200€768
  • H100B€576
On-demand rates · EUR excl. VAT
GPU memory per server
  • B300A2,304 GB
  • B2001,440 GB
  • H2001,128 GB
  • H100B640 GB
HBM on the GPUs, not system RAM

How much fits.

GPU memory decides which models load without sharding. Compare it per card, per server, or against what you pay.

Metric
Model weights that fit, per server
ServerFP16FP8
A8× B3001,152B2,304B
B8× H100320B640B

Parameters, in billions, whose weights fit in GPU memory at 2 bytes (FP16) or 1 byte (FP8). KV cache and activations need room on top.

More comparisons

See allAll comparisons

Case studies

How teams chose their server.

Write-ups from teams who picked between servers like these. Shown as a sample until approved stories exist.
Sample

Placeholder for a customer story about choosing between these two servers. Real quotes appear here once approved.

Customer nameRole, Company
Before and after · product facts
GPU memory per server
Bbefore640 GB
Aafter2,304 GB
24-hour run, on demand
Bbefore€576
Aafter€1,152

Sample layout: B is the server a team moved from, A the one it moved to.

Ready to pick

Rent the B300 or the H100. Switch any time.

Start on demand, keep the one that fits, and reserve it when the workload settles. Billing is per second, in euros.

On-demand ratesEUR · excl. VAT
  • B300€6.00/GPU·hr
  • B200€5.00/GPU·hr
  • H200€4.00/GPU·hr
  • H100€3.00/GPU·hr