DGX H100 Specs Mastery

Every number you need, why it is what it is, and how exam distractors try to trick you.

1. Warm-up: how much network bandwidth leaves one DGX H100?

A capacity planner asks: "What's the peak bandwidth one DGX H100 can push to the rest of the cluster?" Below are four answers people often give. Only one holds up.

WRONG"Eight 200 Gb/s ConnectX-6 ports, so about 400 GB/s"

This mixes in the wrong generation. ConnectX-6 at 200 Gb/s (HDR) belongs to the DGX A100. The DGX H100 uses ConnectX-7 at 400 Gb/s (NDR). The arithmetic is also wrong: 8 × 200 Gb/s = 1,600 Gb/s = 200 GB/s, not 400 GB/s.

WRONG"7.2 TB/s, because NVSwitch extends between nodes"

7.2 TB/s is a real DGX H100 number, but it's the NVSwitch GPU-to-GPU bandwidth inside the node. It's the right number for the wrong question. Inter-node traffic leaves the box through the network adapters.

CORRECT"About 1 TB/s bidirectional, from ten 400 Gb/s ConnectX-7 ports"

10 × 400 Gb/s = 4,000 Gb/s = 500 GB/s each direction, which is 1 TB/s bidirectional. This is NVIDIA's headline datasheet figure.

WRONG"Count the four OSFP cages on the back: 4 × 400 Gb/s = 200 GB/s"

This miscounts the 4 OSFP cages on the back of the DGX H100. Those 4 physical cages carry 8 ConnectX-7 400 Gb/s ports (two per cage), not 4 adapters.

Three things decide this: the generation (ConnectX-7 / 400 Gb/s, not ConnectX-6 / 200 Gb/s), where the bandwidth applies (NVSwitch is inside the node, the NICs go between nodes), and the units (Gb/s → GB/s, one-way → bidirectional). This lesson trains all three.

2. Units first: the three traps behind most spec questions

Trap 1: bits (Gb) vs bytes (GB)

Network speeds are quoted in gigabits per second (Gb/s, small b). Memory and NVLink bandwidth are quoted in gigabytes per second (GB/s, big B). 1 byte = 8 bits.

400 Gb/s ÷ 8 = 50 GB/s

Trap 2: one direction vs bidirectional

A link can send and receive at full speed at the same time. Marketing figures often add both directions together ("bidirectional"), which doubles the number.

Link	Per direction	Bidirectional (both added)
One ConnectX-7 port	400 Gb/s = 50 GB/s	800 Gb/s = 100 GB/s
One H100's NVLink	450 GB/s	900 GB/s (headline)
All 10 ConnectX-7 in a DGX H100	500 GB/s	1 TB/s (headline)

Trap 3: per GPU vs per node, and which network

Scope	Example
Per GPU	NVLink 900 GB/s · one compute NIC 400 Gb/s · HBM3 3.35 TB/s
Per node, intra-node	NVSwitch total 7.2 TB/s (8 × 900 GB/s)
Per node, compute fabric only	8 × 400 Gb/s = 3.2 Tb/s = 400 GB/s per direction (800 GB/s bidirectional)
Per node, all network adapters	10 × 400 Gb/s = 4 Tb/s = 500 GB/s per direction (1 TB/s bidirectional)

"Small b is for the cable, big B is for the chip." Network = bits. Memory and NVLink = bytes. To go from bits to bytes, divide by 8. For "bidirectional," double it.

3. DGX H100 at a glance: the 10 numbers to memorize

H100 SXM5 GPUs

640 GB

total GPU memory (8 × 80 GB HBM3)

4th-gen NVSwitch chips

7.2 TB/s

NVSwitch GPU-to-GPU bandwidth (bidirectional)

900 GB/s

NVLink per GPU (bidirectional, 18 links)

ConnectX-7 400 Gb/s network adapters

1 TB/s

peak bidirectional network bandwidth

32 PFLOPS

FP8 AI performance (with sparsity)

2 TB

system RAM · 2× Xeon 8480C, 112 cores

~10.2 kW

max power · 8U chassis

Plus storage: 2 × 1.92 TB NVMe M.2 (OS) and 8 × 3.84 TB NVMe U.2 (≈ 30 TB internal data).

4. Inside the box: a guided tour

5. The H100 GPU itself (SXM5 version, as used in DGX)

Spec	Value	Notes
Architecture	Hopper	Successor to Ampere (A100), predecessor to Blackwell (B200)
Form factor	SXM5	A socketed module on the baseboard (not a PCIe card). Enables full NVLink.
GPU memory	80 GB HBM3	Per GPU
Memory bandwidth	3.35 TB/s	GPU ↔ its own HBM3, the fastest link in the system
NVLink	900 GB/s	4th-gen NVLink, 18 links, bidirectional
PCIe	Gen5 x16	~128 GB/s bidirectional (~64 GB/s each way)
Max power (TDP)	up to 700 W	Per GPU
FP8 Tensor	~3,958 TFLOPS*	≈ 4 PFLOPS, so × 8 GPUs ≈ 32 PFLOPS per DGX
BF16 / FP16 Tensor	~1,979 TFLOPS*	Training precision
TF32 Tensor	~989 TFLOPS*
FP64 / FP64 Tensor	34 / 67 TFLOPS	HPC precision
MIG	up to 7 instances	e.g. 7 × 1g.10gb
Key features	4th-gen Tensor Cores, Transformer Engine (dynamic FP8/FP16), DPX instructions, confidential computing, NVLink SHARP (NVLS) with NVSwitch

* Tensor figures marked with an asterisk are NVIDIA's "with sparsity" peaks. Dense is half.

The H100 PCIe card is a different product: 80 GB HBM2e, ~2 TB/s, 350 W, NVLink only through a bridge. DGX H100 uses SXM5. If an answer quotes PCIe-card specs for a DGX, it's a distractor.

6. Intra-node: NVLink and NVSwitch

Each H100 has 18 NVLink 4 links at 50 GB/s bidirectional each, so 18 × 50 = 900 GB/s per GPU (450 GB/s each way).
The 18 links are spread across 4 NVSwitch chips, so every GPU can talk to every other GPU at full speed at the same time (all-to-all, non-blocking).
Total: 8 GPUs × 900 GB/s = 7.2 TB/s bidirectional GPU-to-GPU bandwidth.
NVSwitch also supports NVLink SHARP: the switch itself can do reductions (NCCL's NVLS algorithm).
Requires Fabric Manager to be running.
In nvidia-smi topo -m, GPU pairs show as NV18.

"4 switches, 18 links, 900 per GPU, 7.2 per box." Check: 18 × 50 = 900 and 8 × 900 = 7,200 GB/s = 7.2 TB/s.

In a standard DGX H100 cluster, NVSwitch bandwidth does not extend across nodes. Answers like "extend the NVSwitch fabric across nodes" or "inter-node NVLink at 7.2 TB/s" are distractors. (Products like GH200 NVL32 / GB200 NVL72 use external NVLink Switch systems, but that's not a DGX H100 BasePOD.)

7. Inter-node: the 10 ConnectX-7 adapters

Group	Count	Speed	Purpose	Physical ports
Compute fabric	8 × single-port ConnectX-7	400 Gb/s each (NDR InfiniBand or Ethernet)	GPU ↔ GPU between nodes (NCCL), one per GPU, rail-optimized	4 OSFP twin-port cages (2 × 400G each)
Storage / in-band	2 × dual-port ConnectX-7	up to 400 Gb/s	Storage fabric and in-band management networks	QSFP112 ports
Management	Onboard NIC + BMC	10 GbE + 1 GbE BMC	Admin, out-of-band	RJ45

Doing the arithmetic

What	Calculation	Result
Per compute NIC (per GPU)	400 Gb/s ÷ 8	50 GB/s each way
Compute fabric per node	8 × 400 Gb/s = 3.2 Tb/s ÷ 8	400 GB/s each way · 800 GB/s bidirectional
All 10 adapters (headline)	10 × 400 Gb/s = 4 Tb/s ÷ 8 = 500 GB/s each way × 2	1 TB/s peak bidirectional

When a question asks for the "total peak network bandwidth of a DGX H100," the expected answer is NVIDIA's datasheet figure: 10 × ConnectX-7 400 Gb/s → 1 TB/s peak bidirectional. When it asks about GPU-to-GPU compute fabric bandwidth specifically, think 8 × 400 Gb/s = 3.2 Tb/s (400 GB/s each way).

"Ten NICs, four hundred each, one terabyte both ways." And "8 for GPUs, 2 for storage."

8. The bandwidth ladder: every link from fastest to slowest

This ladder explains much of distributed-training design: keep the most communication-heavy work on the fastest links.

#	Link	Bandwidth	Per direction (GB/s)
1	H100 ↔ its own HBM3 memory	3.35 TB/s	(memory bandwidth)
2	NVSwitch total inside one DGX	7.2 TB/s bidir	3,600 total
3	One GPU's NVLink	900 GB/s bidir	450
4	All 10 NICs of one DGX	1 TB/s bidir	500
5	Compute fabric of one DGX (8 NICs)	3.2 Tb/s	400
6	One GPU's PCIe Gen5 x16	~128 GB/s bidir	~64
7	One GPU's compute NIC	400 Gb/s	50

Per GPU, NVLink (450 GB/s each way) is about 9× faster than its network NIC (50 GB/s). That's why tensor parallelism stays inside a node (on NVLink), while data and pipeline parallelism span nodes (over InfiniBand).

HBM is the desk you work at. NVLink is the hallway to colleagues on your floor. InfiniBand is the road to other buildings. PCIe is the stairwell to the basement. You keep the chattiest teams on the same floor.

9. Storage: NVMe drives

Drives	Capacity	Typical config	Purpose
2 × NVMe M.2	1.92 TB each	RAID 1 (mirror)	Operating system
8 × NVMe U.2	3.84 TB each = 30.72 TB (~30 TB)	RAID 0 (stripe), mounted at `/raid`	Local data cache / scratch: datasets, temporary checkpoints

RAID 0 on the data drives gives maximum speed and capacity, but no redundancy. Treat it as cache, not as permanent storage.
Persistent data lives on shared storage (parallel filesystem / NFS) over the storage fabric, through the 2 dual-port ConnectX-7 adapters.
GPUDirect Storage (GDS) lets data flow from NVMe or network storage directly into GPU memory, bypassing CPU bounce buffers.

"2 small for the OS, 8 big for the data." 1.92 × 2 (mirror) and 3.84 × 8 ≈ 30 TB. The data drives match the GPU count: 8 GPUs, 8 NICs, 8 data NVMe.

10. CPU, memory, power, size

Spec	DGX H100
CPUs	2 × Intel Xeon Platinum 8480C (Sapphire Rapids), 112 cores total (56 each), 2.0 GHz base, up to 3.8 GHz boost
System memory	2 TB DDR5
Max power	~10.2 kW
Rack units	8U
Weight	~130 kg (~287 lb)
Operating temperature	5–30 °C (41–86 °F)
Software	DGX OS (Ubuntu-based) or Red Hat Enterprise Linux; NVIDIA AI Enterprise; Base Command (BCM for cluster management)

10.2 kW per node is a big rack-planning number. Many data-center racks are designed for far less, so DGX H100 deployments often have only a few nodes per rack, or need high-density power and cooling.

11. DGX A100 vs H100 vs H200: spotting the wrong generation

Exams love borrowing a number from the neighboring generation. Learn the pattern of each.

Spec	DGX A100	DGX H100	DGX H200
GPU architecture	Ampere	Hopper	Hopper
GPUs	8 × A100	8 × H100	8 × H200
Memory per GPU	40 or 80 GB (HBM2 / HBM2e)	80 GB HBM3	141 GB HBM3e
Total GPU memory	320 or 640 GB	640 GB	1,128 GB
Memory bandwidth per GPU	~1.6–2.0 TB/s	3.35 TB/s	4.8 TB/s
NVLink per GPU	600 GB/s (NVLink 3, 12 links)	900 GB/s (NVLink 4, 18 links)	900 GB/s
NVSwitch	6 × 2nd-gen, 4.8 TB/s	4 × 4th-gen, 7.2 TB/s	4 × 4th-gen, 7.2 TB/s
Compute NICs	8 × ConnectX-6 200 Gb/s (HDR)	8 × ConnectX-7 400 Gb/s (NDR)	8 × ConnectX-7 400 Gb/s
Storage/in-band NICs	2 × dual-port ConnectX-6	2 × dual-port ConnectX-7	2 × dual-port ConnectX-7
Headline network	-	1 TB/s peak bidirectional	1 TB/s
CPUs	2 × AMD EPYC 7742 (128 cores)	2 × Intel Xeon 8480C (112 cores)	2 × Intel Xeon 8480C
System RAM	1 TB or 2 TB	2 TB	2 TB
AI performance	5 PFLOPS (FP16, sparse)	32 PFLOPS FP8	32 PFLOPS FP8
Max power	~6.5 kW	~10.2 kW	~10.2 kW
Size	6U	8U	8U

Generation fingerprints:
A100 era = ConnectX-6, 200 Gb/s, NVLink 600, 6 NVSwitches, 4.8 TB/s, AMD EPYC, 6U.
H100 era = ConnectX-7, 400 Gb/s, NVLink 900, 4 NVSwitches, 7.2 TB/s, Intel Xeon, 8U.
H200 = the H100 box with bigger, faster memory (141 GB HBM3e, 4.8 TB/s).

"4.8 TB/s" appears twice: as the DGX A100's NVSwitch total and as the H200's per-GPU memory bandwidth. Read what the number is attached to.

12. BasePOD and SuperPOD context

DGX H100 nodes are the building block of NVIDIA's reference architectures. The spec-level facts to tie to this guide:

A DGX SuperPOD scalable unit is 32 DGX H100 nodes = 256 GPUs, with one Quantum-2 QM9700 leaf switch (64 × NDR 400 Gb/s ports) per compute rail: 8 leaves for the 8 compute NICs.
Rail-optimized: compute NIC i of every node connects to leaf i.
One SU of DGX alone is about 32 × 10.2 kW ≈ 326 kW.

BasePOD vs SuperPOD, the four networks, leaf/spine/cable counts per SU, rack power limits, and a sizing calculator are in the AI factory architecture guide (reference architectures, sizing math).

13. Expected vs achieved bandwidth (what tests really show)

Test	Theoretical	Healthy measured (roughly)
Single-node NCCL AllReduce busbw (8 GPUs over NVLink)	450 GB/s per GPU each way	~450+ GB/s; can be higher with NVLS
Multi-node NCCL AllReduce busbw (8 × 400G per node)	400 GB/s per node each way	~350–400 GB/s for large messages
`ib_write_bw` on one NDR link	400 Gb/s	Close to line rate (high 300s Gb/s)
HBM3 memory bandwidth	3.35 TB/s	A large fraction of peak in bandwidth tests

Measured values vary with software versions and message sizes. Exams generally test theoretical datasheet values, and troubleshooting questions test "far below expected."

14. Unit converter

Number of links

Speed per link

Unit of speed

Is the per-link speed already bidirectional?

Try: 10 × 400 Gb/s (the DGX H100 headline) · 8 × 400 Gb/s (compute fabric) · 8 × 200 Gb/s (DGX A100 compute) · 8 × 900 GB/s bidirectional (NVSwitch total).

15. All memory hooks on one page

8-80-640: 8 GPUs × 80 GB = 640 GB.
4-18-900-7.2: 4 NVSwitches, 18 NVLinks per GPU, 900 GB/s per GPU, 7.2 TB/s per box.
10-400-1: 10 ConnectX-7, 400 Gb/s each, 1 TB/s bidirectional.
8 for GPUs, 2 for storage: the split of the 10 NICs. The 8 compute ports go through 4 OSFP cages.
Small b = cable, big B = chip. Divide by 8. Double for bidirectional.
3.35: HBM3 TB/s per GPU ("3-3-5, memory's alive").
2-112-2: 2 Xeon CPUs, 112 cores, 2 TB RAM.
2 small + 8 big NVMe: 2 × 1.92 TB OS, 8 × 3.84 TB ≈ 30 TB data.
32 PF FP8: ~4 PF per GPU × 8.
10.2 kW, 8U: power and size.
SU = 32 nodes = 256 GPUs.
NVLink ≈ 9× the NIC per GPU: tensor parallelism stays in the node.

16. Sample outputs

What the commands in this guide print. Expand a command to see a typical result and what to check in it.

Illustrative output, not captured from a real system. Hostnames, serial numbers, GUIDs, MAC and IP addresses are invented, and columns and wording vary between versions. Learn the shape of each output and the fields called out under it.

nvidia-smi topo -mConfirm NV18 between all GPU pairs

        GPU0  GPU1  GPU2  GPU3  GPU4  GPU5  GPU6  GPU7  NIC0  NIC1  NIC2  NIC3  NIC4  NIC5  NIC6  NIC7  NIC8  NIC9  NIC10 NIC11 CPU Affinity      NUMA Affinity  GPU NUMA ID
GPU0     X    NV18  NV18  NV18  NV18  NV18  NV18  NV18  PIX   NODE  NODE  NODE  NODE  NODE  SYS   SYS   SYS   SYS   SYS   SYS   0-55,112-167      0              N/A
GPU1    NV18   X    NV18  NV18  NV18  NV18  NV18  NV18  NODE  NODE  NODE  PIX   NODE  NODE  SYS   SYS   SYS   SYS   SYS   SYS   0-55,112-167      0              N/A
GPU2    NV18  NV18   X    NV18  NV18  NV18  NV18  NV18  NODE  NODE  NODE  NODE  PIX   NODE  SYS   SYS   SYS   SYS   SYS   SYS   0-55,112-167      0              N/A
GPU3    NV18  NV18  NV18   X    NV18  NV18  NV18  NV18  NODE  NODE  NODE  NODE  NODE  PIX   SYS   SYS   SYS   SYS   SYS   SYS   0-55,112-167      0              N/A
GPU4    NV18  NV18  NV18  NV18   X    NV18  NV18  NV18  SYS   SYS   SYS   SYS   SYS   SYS   PIX   NODE  NODE  NODE  NODE  NODE  56-111,168-223    1              N/A
GPU5    NV18  NV18  NV18  NV18  NV18   X    NV18  NV18  SYS   SYS   SYS   SYS   SYS   SYS   NODE  NODE  NODE  PIX   NODE  NODE  56-111,168-223    1              N/A
GPU6    NV18  NV18  NV18  NV18  NV18  NV18   X    NV18  SYS   SYS   SYS   SYS   SYS   SYS   NODE  NODE  NODE  NODE  PIX   NODE  56-111,168-223    1              N/A
GPU7    NV18  NV18  NV18  NV18  NV18  NV18  NV18   X    SYS   SYS   SYS   SYS   SYS   SYS   NODE  NODE  NODE  NODE  NODE  PIX   56-111,168-223    1              N/A
(NIC0-NIC11 rows trimmed)

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_5
  NIC6: mlx5_6
  NIC7: mlx5_7
  NIC8: mlx5_8
  NIC9: mlx5_9
  NIC10: mlx5_10
  NIC11: mlx5_11

What to look at: This is how the spec shows up on a real node: every GPU pair reads NV18 (18 NVLink 4 links through the 4 NVSwitches, 900 GB/s per GPU), and each GPU has one PIX NIC, which is its own 400 Gb/s ConnectX-7 compute port. The 12 NIC entries are the 8 compute ports plus the 4 ports of the 2 dual-port storage adapters; a GPU pair showing SYS or a lower NV count means NVLink is not fully up.

ib_write_bw -d mlx5_0 --report_gbits dgx02RDMA write bandwidth on one NDR link

---------------------------------------------------------------------------------------
                    RDMA_Write BW Test
 Dual-port       : OFF          Device         : mlx5_0
 Number of qps   : 1            Transport type : IB
 Connection type : RC           Using SRQ      : OFF
 PCIe relax order: ON
 ibv_wr* API     : ON
 TX depth        : 128
 CQ Moderation   : 1
 Mtu             : 4096[B]
 Link type       : IB
 Max inline data : 0[B]
 rdma_cm QPs     : OFF
 Data ex. method : Ethernet
---------------------------------------------------------------------------------------
 local address: LID 0x11 QPN 0x0123 PSN 0x1a2b3c RKey 0x1fff00 VAddr 0x007f2a00000000
 remote address: LID 0x12 QPN 0x0456 PSN 0x4d5e6f RKey 0x2fff00 VAddr 0x007f3b00000000
---------------------------------------------------------------------------------------
 #bytes     #iterations    BW peak[Gb/sec]    BW average[Gb/sec]   MsgRate[Mpps]
 65536      5000             391.85             391.72               0.747146
---------------------------------------------------------------------------------------

What to look at: The client run against a server already listening on dgx02 reports about 392 Gb/s, close to the 400 Gb/s NDR line rate of one ConnectX-7 port, which is about 50 GB/s each way. A result near 200 Gb/s suggests the link trained at HDR or half width, and far lower numbers point to cabling, PCIe placement, or a congested fabric; without --report_gbits the same figures appear in MB/sec.

all_reduce_perf -b 8 -e 8G -f 2 -g 8Single-node NCCL AllReduce over NVLink

# nThread 1 nGpus 8 minBytes 8 maxBytes 8589934592 step: 2(factor) warmup iters: 5 iters: 20 agg iters: 1 validation: 1 graph: 0
#
# Using devices
#  Rank  0 Group  0 Pid  51234 on      dgx01 device  0 [0x18] NVIDIA H100 80GB HBM3
#  Rank  1 Group  0 Pid  51234 on      dgx01 device  1 [0x2a] NVIDIA H100 80GB HBM3
#  Rank  2 Group  0 Pid  51234 on      dgx01 device  2 [0x3a] NVIDIA H100 80GB HBM3
#  Rank  3 Group  0 Pid  51234 on      dgx01 device  3 [0x5d] NVIDIA H100 80GB HBM3
#  Rank  4 Group  0 Pid  51234 on      dgx01 device  4 [0x9a] NVIDIA H100 80GB HBM3
#  Rank  5 Group  0 Pid  51234 on      dgx01 device  5 [0xab] NVIDIA H100 80GB HBM3
#  Rank  6 Group  0 Pid  51234 on      dgx01 device  6 [0xba] NVIDIA H100 80GB HBM3
#  Rank  7 Group  0 Pid  51234 on      dgx01 device  7 [0xdb] NVIDIA H100 80GB HBM3
#
#                                                              out-of-place                       in-place          
#       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)       
           8             2     float     sum      -1    30.21    0.00    0.00      0    29.87    0.00    0.00      0
   ...
     1048576        262144     float     sum      -1    30.50   34.38   60.16      0    30.12   34.81   60.93      0
   ...
   134217728      33554432     float     sum      -1    620.0  216.48  378.84      0    618.4  217.04  379.82      0
   ...
  1073741824     268435456     float     sum      -1   4100.0  261.89  458.30      0   4096.2  262.13  458.73      0
   ...
  4294967296    1073741824     float     sum      -1    16000  268.44  469.76      0    15987  268.65  470.14      0
  8589934592    2147483648     float     sum      -1    31400  273.56  478.74      0    31386  273.69  478.96      0
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 152.418 
#

What to look at: Read the busbw column for the largest messages: about 479 GB/s at 8 GB matches the guide's healthy single-node figure of roughly 450 GB/s or more over NVLink (higher with NVLS). A large-message busbw near 50 GB/s means traffic is falling back to PCIe or the network instead of NVLink, and any nonzero #wrong value means corrupted results.

df -h /raidCheck the local NVMe data RAID

Filesystem      Size  Used Avail Use% Mounted on
/dev/md1         28T  1.2T   27T   5% /raid

What to look at: The 8 x 3.84 TB U.2 drives in RAID 0 total 30.72 TB, which df -h shows as 28T because it counts in binary units (TiB). A size near 3.5T would mean only one data drive is in the array, and a missing /raid mount means the data RAID was not assembled at boot; remember that it is scratch with no redundancy.

17. Distractor patterns

Wrong generation: ConnectX-6, 200 Gb/s, HDR, NVLink 600 GB/s, 6 NVSwitches, AMD EPYC, 4.8 TB/s NVSwitch → that's DGX A100.
Wrong scope: 7.2 TB/s (NVSwitch, intra-node) offered as network bandwidth. 900 GB/s (per GPU) offered as per-node.
Wrong units: "400 GB/s per NIC" (it's 400 Gb/s = 50 GB/s). Mixing up Tb/s and TB/s.
Wrong direction: 500 GB/s (one way) vs 1 TB/s (bidirectional). Read whether the question says "bidirectional."
Wrong count: "4 ConnectX-7" (confusing the 4 OSFP cages with 8 ports), "8 NICs total" (forgetting the 2 storage adapters).
Wrong form factor: H100 PCIe specs (HBM2e, 2 TB/s, 350 W) for a DGX (SXM5: HBM3, 3.35 TB/s, 700 W).
Wrong memory: 141 GB / HBM3e / 4.8 TB/s per GPU → that's H200.
"Extend NVSwitch across nodes" in a DGX H100 BasePOD → no. Inter-node traffic uses InfiniBand.
Sparse vs dense FLOPS: 32 PF FP8 is the sparse figure. Dense is about 16 PF.

18. Flashcards (click to flip)

19. Practice quiz (25 questions)

Score: 0 / 0