B300 vs H200 vs RTX PRO 6000: three GPUs for three different jobs
288 GB HBM3e, 141 GB HBM3e or 96 GB GDDR7. How the B300, H200 and RTX PRO 6000 differ in memory, FP4, NVLink and power, with NVIDIA specs and rental prices.
The NVIDIA B300, H200 and RTX PRO 6000 Blackwell Server Edition all run large language models, but they come from three different product lines and are sold in very different machines. This article compares them on the numbers NVIDIA publishes: memory, bandwidth, precision, links between GPUs and power.
Which one for which job
| If you need to... | GPU | Why | Rental, median per GPU-hour |
|---|---|---|---|
| Serve models up to about 70B to many users, at the lowest cost per hour | RTX PRO 6000 | 96 GB per card, native FP4, a standard PCIe card | $2.32 |
| Run one very large model fast, or train across many GPUs | B300 | 288 GB per GPU, 8 TB/s, NVLink at 1.8 TB/s | $9.00 |
| Run FP8 or FP16 models, or need double precision (FP64) | H200 | 141 GB, 4.8 TB/s, 67 TFLOPS of FP64 | $4.88 |
Rental medians are on-demand prices from the getdeploying tracker on 22 October 2026 (42, 15 and 33 providers with a priced offer). Our own server with eight RTX PRO 6000 cards rents from $1.70 per GPU-hour, as a whole machine on a monthly contract.
The specifications side by side
| B300 (Blackwell Ultra) | H200 SXM | RTX PRO 6000 Server Edition | |
|---|---|---|---|
| GPU memory | 288 GB HBM3e | 141 GB HBM3e | 96 GB GDDR7 |
| Memory bandwidth | 8 TB/s | 4.8 TB/s | 1,597 GB/s |
| FP4 | Yes: 15 PFLOPS dense, 20 sparse | No | Yes: 4 PFLOPS |
| FP8 | 5 PFLOPS dense, 10 sparse | 3,958 TFLOPS, sparse | 2 PFLOPS |
| GPU-to-GPU link | NVLink 5, 1.8 TB/s | NVLink, 900 GB/s | PCIe 5.0, 128 GB/s |
| Maximum power | up to 1,400 W | up to 700 W | up to 600 W |
| How it is sold | 8-GPU HGX and DGX systems, GB300 racks | 8-GPU HGX systems; H200 NVL as a PCIe card | PCIe card, air or liquid cooled |
All values are from NVIDIA's product pages and its Blackwell Ultra architecture post. NVIDIA does not mark the RTX PRO 6000 compute figures as dense or sparse on the product page, so compare them with care. The memory and bandwidth rows are the ones that matter most for inference, and those are exact.
Memory capacity: the B300 holds the most on one GPU
The first question for any model is whether its weights and its KV cache fit. As a rule of thumb, FP8 needs about one byte per parameter and FP4 about half a byte.
- A 70B model in FP8 needs about 70 GB for its weights. It fits on one GPU of each kind, but on the RTX PRO 6000 it leaves only about 26 GB for the cache, and on the B300 more than 200 GB.
- A 405B model in FP8 needs about 405 GB. That is two B300 GPUs, at least three H200 GPUs with little room left for the cache, or five or more RTX PRO 6000 cards.
- The same 405B model in FP4 needs about 200 GB. That fits on one B300, or on three RTX PRO 6000 cards. The H200 has no hardware FP4, so it stays at FP8 or higher.
The more memory each GPU has, the fewer GPUs a model has to be split across. Splitting costs speed, as the section on links below shows.
Memory bandwidth: the B300 writes one answer fastest
When a model generates text for a single user, it reads its weights again for every new token. Memory bandwidth sets that pace. The B300 has about 1.7 times the bandwidth of the H200 and five times that of the RTX PRO 6000.
With many users at once, each read is shared by the whole batch, so compute and memory capacity count for more and the gap narrows. This is why a card with less bandwidth can still give good throughput per dollar when it serves a busy chat service or an API.
Precision: FP4 is a Blackwell feature
Both Blackwell products, the B300 and the RTX PRO 6000, run FP4 in hardware. FP4 halves the memory a model needs compared with FP8 and halves the data read per token. NVIDIA's Blackwell Ultra post lists 15 PFLOPS of dense FP4 per B300 GPU, up from 10 on the first Blackwell.
The H200 is a Hopper GPU. Its smallest hardware format is FP8. If your models are already tuned for FP8 or FP16, that is no handicap. If you want to run FP4 checkpoints, it is.
Blackwell Ultra is built for AI formats. NVIDIA lists 67 TFLOPS of FP64 for the H200 SXM. If your work includes double-precision simulation, check the FP64 figure on the B300 datasheet before you buy: NVIDIA's Blackwell Ultra architecture post does not list it.
Links between GPUs: NVLink decides how far one model scales
The B300 uses fifth-generation NVLink at 1.8 TB/s per GPU, the H200 SXM NVLink at 900 GB/s. The RTX PRO 6000 talks to other cards over PCIe 5.0 at 128 GB/s in both directions combined.
This matters only when one model is split across several GPUs. In a public benchmark we covered in RTX PRO 6000 vs H100, one large model split across eight RTX PRO 6000 cards ran at about a third of the speed of eight H100 GPUs. Eight separate copies of a smaller model, one per card, scaled almost perfectly. On PCIe cards, run each model on one, two or four GPUs and add copies. On NVLink systems you can split one model across all eight.
Power and the room around it
The three GPUs need very different data centre space:
- B300: up to 1,400 W per GPU. NVIDIA rates a DGX B300 with eight GPUs at about 14 kW, in a 10U chassis.
- H200: up to 700 W per SXM GPU. The PCIe version, H200 NVL, is a dual-slot air-cooled card at up to 600 W.
- RTX PRO 6000 Server Edition: up to 600 W per card, a dual-slot PCIe card with air cooling, or single-slot with liquid cooling.
Our own server with eight RTX PRO 6000 cards draws about 6.5 kW from the wall under full GPU load. We published the full power and temperature measurements. A rack position that can take 7 kW of air-cooled load is far easier to find than one for 14 kW per server.
How we run it
Our server in Sofia has eight RTX PRO 6000 Server Edition cards with 768 GB of GPU memory in total, rented as a whole machine. It is built for serving models that fit on one to four cards, many users at once, and FP4. See the server specifications or the dedicated GPU server page for how renting the whole machine works.
Questions
Is the B300 the same as the GB300?
No. The B300 is the Blackwell Ultra GPU. The GB300 combines Blackwell Ultra GPUs with NVIDIA Grace CPUs: a GB300 NVL72 rack holds 72 Blackwell Ultra GPUs and 36 Grace CPUs in one NVLink domain.
Can I buy a single B300 card?
NVIDIA's HGX and DGX pages list the B300 only as SXM modules on 8-GPU boards, and in GB300 rack systems. The H200 has a PCIe version, the H200 NVL. The RTX PRO 6000 Server Edition is a PCIe card that fits standard GPU servers.
What is the difference between the H200 and the H100?
Both are Hopper GPUs. The H200 has more and faster memory: 141 GB of HBM3e at 4.8 TB/s, against 80 GB at 3.35 TB/s on the H100.
How much memory does an eight-GPU server of each have?
Eight B300 GPUs have about 2.3 TB (NVIDIA lists 2.1 TB for the HGX B300), eight H200 GPUs 1.1 TB, and eight RTX PRO 6000 cards 768 GB.
Sources
- NVIDIA: Inside NVIDIA Blackwell Ultra, August 2025, updated September 2025 (288 GB, 8 TB/s, FP4 and FP8, NVLink, 1,400 W, Hopper comparison).
- NVIDIA: HGX platform and DGX B300, checked 22 October 2026 (2.1 TB per HGX B300, about 14 kW and 10U for DGX B300).
- NVIDIA: H200, checked 22 October 2026.
- NVIDIA: RTX PRO 6000 Blackwell Server Edition, checked 22 October 2026.
- getdeploying: rental prices for the B300, H200 and RTX PRO 6000, 22 October 2026.
- NEOXIS: our own power measurements, July to October 2026.
