GPU hardware

AI servers

From a deskside GB300 to 6U rack servers

Up to 8 NVIDIA RTX 5090 or RTX 4090 GPUs in 6U, a deskside GB300 AI Station with 748 GB of coherent memory, and NVIDIA L40S and AMD Instinct MI210 data center GPUs, all by quote. Size the GPUs to the channels and AI workloads you actually run, with a DVEO engineer.

  • NVIDIA GB300
  • Blackwell Ultra
  • RTX 5090
  • RTX 4090
  • L40S
  • Instinct MI210
  • HBM3e
  • DDR5
DVEO GB300 AI STATION, three-quarter view illustrationPWRNETGPUGB300AI STATION
AI performance, up to, on the GB300 AI Station
20PFLOPS
coherent memory on the GB300 AI Station
748GB
maximum per RTX 5090 or RTX 4090 server
8GPUs

Overview

Match the hardware to the workload

Encoding a linear channel or transcoding at stable volumes runs well on CPU servers. GPUs earn their place when the same operation repeats across many streams or frames at once: full ABR ladders for dozens of channels, or AI subtitling, dubbing, upscaling and restoration running beside the transcode.

DVEO's AI line covers that range. The RTX 4090 and RTX 5090 AI Servers take up to eight GPUs in a 6U rackmount chassis for high channel counts and concurrent AI work. The GB300 AI Station puts 748 GB of coherent memory beside a desk for large models, with no server room. The NVIDIA L40S and AMD Instinct MI210 extend the line into data center deployments.

The same hardware line runs the GPU cloud of 1Legion, DVEO's sister company, for real-time transcoding and AI media processing.

Where it fits

Where AI hardware sits in the media chain

AI processing runs between ingest and delivery, next to or inside the transcode.

AI media workflows: archive files and live streams run on a DVEO GPU server, a GB300 AI Station or RTX AI Server on premises or cloud GPUs, through three parallel pipelines. Subtitles: speech to text, speaker detection, translation into 50+ languages and time alignment. Dubbing: speech to speech translation, voice cloning, tone matching and lip-sync. Upscaling: 2x to 4x resolution, noise and artifact reduction, frame restoration and frame interpolation. Subtitle files, dubbed programs and HD or 4K masters go to playout, OTT and FAST, and Brutus packages WebVTT tracks into HLS.1SOURCE MEDIAArchive filesMP4, MXF, MOVLive streamSRT, HLS2GPU SERVERDVEO GPU serverGB300 AI Station,RTX AI Server,or cloud GPUs3SUBTITLESAI subtitlesSpeech to textSpeaker detectionTranslateTime-alignSRT, VTT, STL4DUBBINGAI dubbingSpeech to speechVoice cloneTone matchLip-syncMP4, MXF, MOV5UPSCALINGAI upscalerUpscale 2x to 4xDenoiseRestore framesInterpolateHD, 4K MASTERS6DELIVERBrutusHLS + WebVTTPlayout, OTTand FASTIPDELIVERY, AIFILESDVEO PRODUCTTHIRD PARTY
Live and file content is ingested, processed on GPUs for subtitles, dubbing and upscaling, transcoded and packaged, and delivered to broadcast, OTT and FAST. The AI media workflows solution shows the full workflow.
  1. Stage 1: Source media

    Live streams and file libraries, including archive hours for batch work.

  2. Stage 2: GPU server

    A deskside GB300 AI Station, RTX 5090 or RTX 4090 AI Servers with up to 8 GPUs in 6U, or NVIDIA L40S and AMD Instinct MI210 cards.

  3. Stage 3: Subtitles

    AI for Media transcribes and translates speech into subtitles on the GPUs, for live streams and recorded video.

  4. Stage 4: Dubbing

    AI dubbing translates speech to speech with cloned voices, and can share the GPUs with the transcode.

  5. Stage 5: Upscaling

    Upscaling and restoration repeat the same operation across every frame, the kind of work where GPUs earn their place.

  6. Stage 6: Deliver

    Brutus or Brutus Cloud transcodes the results to multi-bitrate output for broadcast, OTT and FAST destinations.

Solutions that use AI servers

Features

What the line offers

From a deskside superchip system to 6U multi-GPU servers and data center cards.

Up to eight GPUs
RTX 5090 and RTX 4090 AI Servers take up to eight GPUs each.
Grace Blackwell Ultra
The GB300 AI Station joins a Grace CPU and a Blackwell Ultra GPU over NVLink C2C.
748 GB coherent memory
252 GB HBM3e plus up to 496 GB LPDDR5X as one pool, for models of up to 1 trillion parameters.
800 Gb/s networking
Two 400 GbE QSFP ports on a ConnectX-8 SuperNIC, plus 10 GbE, on the GB300.
Redundant power
Hot-swap CRPS modules from 2000 W to 3200 W with 3+1 redundancy on the RTX servers.
Hot-swap storage
12 hot-swap SATA, SAS or NVMe drive bays on the RTX servers.
Security
TPM 2.0, chassis intrusion alarm and BMC and BIOS redundancy on the RTX servers.
Deskside option
The GB300 AI Station needs no server room, raised floor or dedicated cooling.
Operating systems
GB300: Ubuntu, Red Hat Enterprise Linux, NVIDIA DGX OS. RTX servers: Windows Server, Red Hat Enterprise Linux, SUSE Linux, CentOS, Ubuntu, VMware ESXi.

Models

Five AI hardware products

All five are sold by quote, configured to your workload.

  • Quote only
    DVEO GB300 AI STATION, front panel close-up illustrationPWRNETGPUGB300AI STATION

    DVEO GB300 AI Station

    Deskside AI supercomputer on NVIDIA DGX Station architecture

    Price on requestSend inputs, outputs, codecs and channel count for a quote.

    • NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip
    • Up to 20 PFLOPS AI, 748 GB coherent memory
    • 2 x 400 GbE QSFP plus 10 GbE
    • 1600 W 80 PLUS Titanium, 245 x 528.4 x 595 mm
  • Quote only
    DVEO RTX 5090 AI SERVER, front panel close-up illustrationRTX 5090AI SERVERPWRNETGPUGPU 1GPU 2GPU 3GPU 4GPU 5

    DVEO RTX 5090 AI Server

    6U server for up to 8 NVIDIA RTX 5090 GPUs

    Price on requestSend inputs, outputs, codecs and channel count for a quote.

    • Up to 8 NVIDIA RTX 5090 GPUs
    • 1 or 2 Intel Xeon Scalable, 4th or 5th Gen, up to 350 W
    • 32 DDR5 slots, up to 5600 MHz
    • 12 hot-swap bays, 13 PCIe slots, 3+1 redundant power
  • Quote only
    DVEO RTX 4090 AI SERVER, front panel close-up illustrationRTX 4090AI SERVERPWRNETGPUGPU 1GPU 2GPU 3GPU 4GPU 5

    DVEO RTX 4090 AI Server

    6U server for up to 8 NVIDIA RTX 4090 GPUs

    Price on requestSend inputs, outputs, codecs and channel count for a quote.

    • Up to 8 NVIDIA RTX 4090 GPUs
    • 1 or 2 Intel Xeon Scalable, 4th or 5th Gen, up to 350 W
    • 32 DDR5 slots, up to 5600 MHz
    • 12 hot-swap bays, 13 PCIe slots, 3+1 redundant power
  • Quote only
    L40S NVIDIA, front panel close-up illustrationL40SNVIDIA

    NVIDIA L40S

    Data center GPU on NVIDIA Ada Lovelace architecture

    Price on requestSend inputs, outputs, codecs and channel count for a quote.

    • Fourth-generation Tensor Cores with structural sparsity
    • TF32 for mixed-precision training
    • Generative AI, LLM training, rendering and video processing
  • Quote only
    MI210 AMD INSTINCT, front panel close-up illustrationMI210AMD INSTINCT

    AMD Instinct MI210

    PCIe accelerator on AMD CDNA2

    Price on requestSend inputs, outputs, codecs and channel count for a quote.

    • 6,656 stream processors, 104 compute units
    • 64 GB HBM2e at 1.6 TB/s with full-chip ECC
    • 181 TFLOPS FP16, 45.3 TFLOPS FP64 matrix
    • PCIe 4.0 x16, passive cooling, double slot

Compare models

Model comparison
SpecGB300 AI StationRTX 5090 AI ServerRTX 4090 AI ServerNVIDIA L40SAMD Instinct MI210
Form factorDeskside, 245 x 528.4 x 595 mm6U rackmount6U rackmountData center GPUPCIe card, full height, double slot
GPUs1 NVIDIA Blackwell UltraUp to 8 RTX 5090Up to 8 RTX 40901, Ada Lovelace1, CDNA2
CPUNVIDIA Grace1 or 2 Intel Xeon Scalable1 or 2 Intel Xeon ScalableHost serverHost server
Memory748 GB coherent32 DDR5 slots, up to 5600 MHz32 DDR5 slots, up to 5600 MHzNot listed64 GB HBM2e
Networking or bus2 x 400 GbE QSFP, 1 x 10 GbE13 PCIe slots13 PCIe slotsNot listedPCIe 4.0 x16
Power1600 W 80 PLUS Titanium2000 to 3200 W CRPS, 3+12000 to 3200 W CRPS, 3+1Not listedPassive cooling

All models: quoted per project. Ask us for NVIDIA L40S memory, bus and power details.

Request a quote

What to send for a quote

Send these details with your request so an engineer can match GPUs and memory to the work.

  1. Step 1: Workloads

    Transcoding, subtitling, dubbing, upscaling, restoration or inference, and which run together.

  2. Step 2: Channels and jobs

    How many channels or jobs run at once, and the library hours for batch work.

  3. Step 3: Codecs and resolutions

    Input and output codecs and resolutions, for example SD archives to 4K.

  4. Step 4: Models and data

    Model sizes, languages and data sets you plan to run.

  5. Step 5: Deployment

    Deskside, rack or data center; power, cooling and rack space; on-prem, cloud or hybrid.

  6. Step 6: Timeline

    When the hardware has to be in service.

Specifications

AI hardware specifications

Compute

GB300 AI StationNVIDIA GB300 Grace Blackwell Ultra Desktop Superchip: 1 Blackwell Ultra GPU and a Grace CPU over NVLink C2C; up to 20 PFLOPS AI; models up to 1 trillion parameters
RTX 5090 AI ServerUp to 8 NVIDIA RTX 5090 GPUs; 1 or 2 Intel Xeon Scalable (4th or 5th Gen), up to 350 W TDP
RTX 4090 AI ServerUp to 8 NVIDIA RTX 4090 GPUs; 1 or 2 Intel Xeon Scalable (4th or 5th Gen), up to 350 W TDP
NVIDIA L40SAda Lovelace; fourth-generation Tensor Cores with structural sparsity; TF32; DLSS
AMD Instinct MI210CDNA2; 6,656 stream processors, 104 compute units, 1700 MHz; 181 TFLOPS FP16 and bfloat16; 45.3 TFLOPS FP32 and FP64 matrix; 22.6 TFLOPS FP32 and FP64 vector; 181 TOPS INT8 and INT4

Memory and storage

GB300 AI Station252 GB HBM3e at 7.1 TB/s plus up to 496 GB LPDDR5X at up to 396 GB/s, 748 GB coherent; NVMe: 2 x M.2 PCIe 5.0, 2 x M.2 PCIe 6.0
RTX 5090 and RTX 4090 AI Servers32 DDR5 slots, up to 5600 MHz; 12 hot-swap SATA, SAS or NVMe drives
AMD Instinct MI21064 GB HBM2e, 4096-bit, 1.6 GHz, 1.6 TB/s, full-chip ECC

Transport and protocols

GB300 AI Station2 x 400 GbE QSFP (ConnectX-8 SuperNIC), 1 x 10 GbE RJ45
AMD Instinct MI210PCIe 4.0 x16; 3 Infinity Fabric links at up to 100 GB/s

Control and management

GB300 AI StationAspeed AST2600 BMC; Ubuntu, Red Hat Enterprise Linux, NVIDIA DGX OS; NVIDIA AI software stack and DVEO AI for Media
RTX 5090 and RTX 4090 AI ServersWindows Server, Red Hat Enterprise Linux, SUSE Linux, CentOS, Ubuntu, VMware ESXi; TPM 2.0, chassis intrusion alarm, BMC and BIOS redundancy
AMD Instinct MI210AMD ROCm; RAS with page retirement

Physical

GB300 AI StationDeskside, 245 mm W x 528.4 mm H x 595 mm D; 1600 W 80 PLUS Titanium ATX; 1 x double-wide and 2 x single-wide PCIe 5.0 x16
RTX 5090 and RTX 4090 AI Servers6U rackmount, Intel Eagle Stream; 13 PCIe slots; CRPS 2000, 2200, 2600 or 3200 W hot-swap power, 3+1 redundancy; 12 hot-swap 6056 and 4 hot-swap 8038 fans
AMD Instinct MI210PCIe add-in card, full height, double slot, 10.5 in (267 mm), passive cooling

Warranty

StandardOne year on every DVEO product
Extended coverageAsk sales

Rows are grouped by product where the line differs. Specifications not listed by DVEO are left out.

FAQ

AI hardware questions

AI hardware is sold by quote. Send your workload and DVEO replies with a configuration and a price.

Not always. GPUs pay off at high channel counts, for multi-resolution live transcoding and when AI work shares the hardware. A single low-density channel with strict quality targets can be cheaper, and just as fast, on CPU. Size by channel count, codec and concurrency first.

No. It is a deskside unit, 245 x 528.4 x 595 mm, with no server room, raised floor or dedicated cooling required.

The NVIDIA AI software stack for fine-tuning, inference and data science, and DVEO AI for Media for dubbing, subtitling, upscaling and restoration.

GB300 AI Station: Ubuntu, Red Hat Enterprise Linux, NVIDIA DGX OS. RTX 5090 and RTX 4090 AI Servers: Windows Server, Red Hat Enterprise Linux, SUSE Linux, CentOS, Ubuntu, VMware ESXi.

After the sale

How support works

  • One-year warranty

    Every DVEO product carries a one-year warranty.

  • DVCare, 1 to 4 years

    Extended warranty with priority phone and email support, advance replacement units shipped within 24 hours and firmware notifications with release notes.

  • Repairs and RMA

    Email rma@dveo.com (or info@dveo.com) for an RMA number and ship the unit to San Diego. Repairs usually come back within a week unless parts are short.

  • Spare parts

    Spare parts are available for DVEO products. Ask for the part with the model and serial number.

From the DVEO blog

Further reading

All news and articles

Size AI hardware to your workload

Tell us what you process, how much runs at once and where it runs, and DVEO recommends a configuration.