A hardware transcoder is a dedicated video processor, such as a GPU, an ASIC or an FPGA. It converts video faster, and with less power, than a CPU. Comparing one comes down to three numbers: real-time streams per card, quality at a given bitrate, and cost per stream.
This guide shows how to read each number for any transcoding workload. It also draws on a Streaming Media buyers’ guide, a November 2025 academic paper and NVIDIA’s own tables. Finally, it covers what changes when the card sits in a rented dedicated server.
📖 New to the term?
Read What Is Video Transcoding? A Complete Beginner’s Guide for how transcoding works before you compare hardware.
What Counts as a Hardware Transcoder?
A hardware transcoder is any dedicated video chip that encodes faster than a CPU. It comes in three forms: a GPU, an FPGA, or an ASIC built for video. Streaming Media’s buyers’ guide, written by Jan Ozer, groups high-volume transcoders into the same three classes.
The guide leaves out Intel’s Quick Sync engine inside the CPU, because that approach is hard to scale. By contrast, you can pack 10 or more GPU or ASIC devices into one server.
Four ways to transcode video
| Where the work runs | Examples | Cards in one server (Ozer’s quotes) |
|---|---|---|
| GPU encoder, such as NVENC | NVIDIA T4 | Up to 6 in a 4RU server |
| Video ASIC or VPU | AMD MA35D, NETINT Quadra | 6 in a 2RU server (MA35D), up to 10 in NETINT’s Quadra video server |
| Engine inside Intel graphics | Intel Quick Sync | Not built to scale |
| Software on the CPU | x264, x265 | Not applicable |
Table of four ways to transcode video: a GPU encoder such as NVENC on an NVIDIA T4 with up to six cards in a 4RU server, a Video ASIC or VPU such as the AMD MA35D or NETINT Quadra with six or ten cards per server, the Intel Quick Sync engine that is not built to scale, and software on the CPU.
Quick Sync also runs on Intel’s discrete Arc GPUs, which the November 2025 paper tests. Ozer discloses that he worked at NETINT until March 2024 and made an AMD MA35D training course in 2024.
How Many Streams Can a Hardware Transcoder Handle?
Count real-time streams, not frames per second. For example, Ozer notes that a transcoder that outputs 1080p30 at 1,200 frames per second does not necessarily produce 40 streams. Some devices cap simultaneous streams.
The same rule also applies to GPU transcoding. In addition, a vendor figure without a named preset probably used a fast, low-quality one, Ozer warns. However, published numbers vary a lot. His T4 figures give 16 streams per card, while a Wowza benchmark of a T4-powered AWS instance gives 60.
The same GPU appears as 16 or 60 streams, depending on the source
Published 1080p30 streams per device. Presets and settings differ between sources.
NVIDIA T4 card (Ozer’s tests and sources)
AMD MA35D card (H.264, HEVC or mid-quality AV1)
AWS g4dn.xlarge instance, one T4 (Wowza benchmark)
Figures as reported in Streaming Media’s buyers’ guide. Ozer does not explain the gap, and his throughput figures are for H.264.
Horizontal bar chart of published 1080p30 stream counts per device: an NVIDIA T4 card at 16 streams in Ozer’s tests and sources, an AMD MA35D card at 32 streams, and an AWS g4dn.xlarge instance with one T4 at 60 streams in a Wowza benchmark.
How Many Cards Does a 100-Channel Service Need?
A 100-channel service needs six MA35D cards or twelve T4 cards. Ozer’s example ladder equals 1.87 streams of 1080p per channel, so 100 channels need 187 streams of 1080p30. At 32 streams per card, that is six cards. At 16 per card, however, it is twelve.
Cards needed for 100 live channels (187 streams of 1080p30)
| Card | Streams per card | Cards needed | Servers, per Ozer’s quotes |
|---|---|---|---|
| AMD MA35D | 32 | 6 | One 2RU server, $18,500 turnkey |
| NVIDIA T4 | 16 | 12 | Two 4RU servers, cards at about $750 each |
Table comparing the AMD MA35D and NVIDIA T4 for a 100-channel live service of 187 streams of 1080p30: the MA35D handles 32 streams per card and needs six cards in one 2RU server at 18,500 dollars, and the T4 handles 16 per card and needs twelve cards in two 4RU servers.
Meanwhile, NVIDIA’s support matrix lists the encoder units on each card. Extra units raise total throughput only across parallel streams, according to NVIDIA’s engineers.
📖 Planning the viewer side as well?
Read Dedicated Server for Streaming for the bandwidth maths behind concurrent viewers.
Which Hardware Transcoder Gives the Best Quality?
No hardware transcoder wins on quality everywhere, because the ranking changes with the codec, the preset and the test. Three published benchmarks agree on HEVC and split on AV1.
A November 2025 paper on arXiv compared Intel Quick Sync and NVIDIA NVENC with software encoders. It used the slowest software presets that still run in real time. In that test, hardware still matched software. Quick Sync saved about 5% of bitrate, and NVENC needed about 10% more bits than Quick Sync.
The authors also add that hardware from the past six years shows negligible quality differences. Instead, the gains came from newer codecs: HEVC beat H.264 clearly, while AV1’s edge over HEVC was small. However, they split on AV1, as the table shows.
Three benchmarks, one disagreement
| Benchmark | What it tested | What it found |
|---|---|---|
| Moscow State University test (2023), as summarised by Ozer | Gaming GPUs: RTX 4070 Ti, Arc A380, RX 6800 XT | Intel led HEVC by about 23% in bitrate. NVIDIA led AV1 by about 7.5% |
| Tom’s Hardware test (March 2023) | Consumer cards: RTX 4090, Arc A770, RX 7900 XTX | Arc led HEVC. NVIDIA’s Ada led AV1. AMD lagged |
| arXiv paper, November 2025 | Intel Arc and NVIDIA GPUs against the slowest real-time software presets | Quick Sync saved about 5%. NVENC needed about 10% more bits than Quick Sync |
Table comparing three hardware encoder benchmarks: a 2023 Moscow State University test where Intel led HEVC and NVIDIA led AV1, a March 2023 Tom’s Hardware test where Intel’s Arc led HEVC, NVIDIA’s Ada led AV1 and AMD lagged, and a November 2025 arXiv paper where Quick Sync saved about 5 percent and NVENC needed about 10 percent more bits.
None of the three tested data centre cards such as the T4. The arXiv test also capped software at what a laptop CPU could run in real time, and a many-core server could run slower presets. So test your own content.
What About AMD’s Encoders?
AMD’s gaming GPUs lagged in the one independent test, but the MA35D is a different chip. Tom’s Hardware found AMD’s consumer encoders behind NVIDIA and Intel in 2023. The MA35D is a video-specific ASIC, and AMD describes its quality only as close to x264 medium, according to Ozer.
How Do You Compare the Cost of a Hardware Transcoder?
Compare cost per stream: add up cards, servers, power and rack space, then divide by the real-time streams they deliver. Ozer prices the same 100-channel service several ways, so the numbers are comparable.
In the cloud, the T4-powered g4dn.xlarge is the cheapest option. It delivers 60 streams for $0.21 per hour. Overall, the 100-channel service costs $36,792 over three years with a three-year commitment. However, bigger instances do not scale. The g4dn.12xlarge costs 7.4 times as much for 3.5 times the throughput. The g4dn.16xlarge costs 8.3 times as much for about 10% more.
Bigger T4 cloud instances cost far more than they deliver
Price and throughput as multiples of the g4dn.xlarge (1x).
g4dn.12xlarge
g4dn.16xlarge
Figures from Streaming Media’s buyers’ guide, based on Wowza’s AWS benchmarks and a three-year commitment.
Bar chart comparing price and throughput of larger T4 cloud instances as multiples of the g4dn.xlarge: the g4dn.12xlarge costs 7.4 times as much for 3.5 times the throughput, and the g4dn.16xlarge costs 8.3 times as much for 1.1 times the throughput.
On premises, the T4 has the lowest purchase cost, but it needs two 4RU servers. Still, Ozer found that housing and powering them for five years more than makes up the difference. Overall, he puts owned hardware about 22% below the cloud, before the time value of money.
Quality gaps also cost real money. For example, in Ozer’s 100-channel model at $0.04 per GB, a 23% bitrate saving is worth about $170,000 over five years. A 7.5% saving, by contrast, is worth about $55,300.
How Much Power Does Software Add?
Software encoding at the best real-time preset roughly doubles the power draw, according to the arXiv paper. For example, on one laptop encoding 1080p60 H.264, x264’s slow preset used 38.3% of the CPU and 85 W. Quick Sync used 5.2% and 28 W.
Real-time 1080p60 H.264 on one laptop
| Encoder | CPU load | System power |
|---|---|---|
| Intel Quick Sync (13th generation) | 5.2% | 28 W |
| NVIDIA NVENC (Ada) | 4.9% | 37 W |
| x264, medium preset | 20.8% | 77 W |
| x264, slow preset | 38.3% | 85 W |
Table of CPU load and system power for real-time 1080p60 H.264 encoding on one laptop: Intel Quick Sync used 5.2 percent CPU and 28 watts, NVIDIA NVENC 4.9 percent and 37 watts, x264 medium 20.8 percent and 77 watts, and x264 slow 38.3 percent and 85 watts.
One rule of thumb ties cost to quality. Above 200 to 300 viewers per channel, though, the highest-quality preset pays for itself, Ozer argues. The bandwidth saved outweighs the extra transcoding cost.
Not Sure How Many Cards Your Channels Need?
Send us your channel count, codecs and ladder. We will tell you what a server for it looks like, and quote a GPU if it makes sense.
→ Talk to Us About Your WorkloadHardware vs Software Transcoding: Which Fits Which Job?
Software transcoding fits video on demand and low-volume live, while most high-volume live work needs hardware. That is Ozer’s conclusion. The arXiv paper reaches a similar one for real-time work, although it tested only consumer hardware.
How Fast Is Hardware Against Software at 4K?
At 4K, the arXiv paper timed H.264 on a Ryzen 9 5900X and on GPUs. Hardware beat x264’s medium and very fast presets, but not the two fastest.
At 4K, hardware outruns x264’s medium and very fast presets, not the two fastest
H.264 encoding speed at 2160p, frames per second. Faster software presets give lower quality.
x264 ultra fast (CPU)
x264 super fast (CPU)
NVIDIA Ada, NVENC (GPU)
Intel Arc, Quick Sync (GPU)
x264 very fast (CPU)
x264 medium (CPU)
Source: arXiv paper, November 2025, Table VIII. Software ran on a Ryzen 9 5900X.
Horizontal bar chart of H.264 encoding speed at 4K: x264 ultra fast at 157 frames per second, x264 super fast at 105, NVIDIA Ada NVENC at 69.5, Intel Arc Quick Sync at 62, x264 very fast at 39.9, and x264 medium at 29.3.
Which Job Suits Which Encoder?
Match the encoder to the job: software for libraries and low volume, hardware for dozens of live channels.
Hardware vs software transcoding, job by job
| Job | Better fit | Why |
|---|---|---|
| A video-on-demand library | Software | Slow presets compress best, and there is no real-time deadline |
| A few live channels | Software or one GPU | Low volume does not justify extra hardware, but a GPU frees the CPU |
| Dozens of live channels | Hardware | Streams per rack unit and power decide the cost |
| Several codecs, AV1 included | Check the card | Ozer advises planning for H.264, HEVC and AV1 over three to four years |
Table matching four transcoding jobs to the better fit: software for a video-on-demand library, software or one GPU for a few live channels, hardware for dozens of live channels, and checking the card when several codecs including AV1 are needed.
For the CPU-against-GPU choice on a rented machine, the next guide goes deeper.
📖 Choosing the machine, not just the card?
Read Video Transcoding Server: CPU, GPU, and What to Check Before You Rent for the NVENC card table, the GeForce licence clause and how to measure capacity.
What Changes When You Rent a Dedicated Server?
On a rented dedicated server, a hardware transcoder is a card in a slot. The questions therefore become slots, power, software and licence. Ozer’s configurations show the scale: six cards fill a 2RU or 4RU server.
- Slots and power: ask how many free PCIe slots the server has, and what power budget each card needs.
- Software support: almost every device supports FFmpeg, and GStreamer is another popular option, but check that your streaming application supports the card.
- Licence: a GeForce card carries a driver clause about data centres, which our server guide explains.
- Quick Sync: Intel lists no integrated graphics for the Xeon Gold 5215 and 6138, so do not count on it with them.
Swify’s standard plans use those two CPUs, so they are CPU-only. Meanwhile, GPU servers are quoted to order. In practice, the cheapest first step is to measure your ladder on CPU cores. Then ask for a card once you know the number.
📖 Measuring on CPU first?
Read CPU Cores Explained: Cores, Threads, and vCPUs for what the core count of a server means for encoding speed.
Hardware Transcoder: The Short Version
A hardware transcoder pays off when you run many live channels. Compare one on streams, quality and cost per stream, and test with your own footage, because published figures vary widely.
- Count real-time streams per card, not frames per second.
- Expect bitrate gaps of roughly 7% to 23% between vendors, and test your own content.
- Compare cost per stream, including servers, rack space and power.
Find Out What Your Footage Needs Before You Buy a Card
Swify dedicated servers start at €120 a month, with up to 40 physical cores, 1 Gbps unmetered bandwidth and full root access. Run your ladder on CPU first, read FFmpeg’s speed figure, and ask for a GPU quote once you know the number.
→ Explore Swify Dedicated ServersFrequently Asked Questions
What is a hardware transcoder?
A hardware transcoder is a dedicated video processor that converts video faster and with less power than a CPU. It can be a GPU with a built-in encoder, an FPGA, or an ASIC built for video. For live streaming, however, its capacity is counted in real-time streams, not frames per second.
Read What Is Video Transcoding? A Complete Beginner’s Guide for how transcoding works.
Is hardware transcoding better than software transcoding?
It depends on the job. Software transcoding suits video on demand and low-volume live, because slow presets compress best. Hardware transcoding, by contrast, suits high-volume live, where streams per server and power draw decide the cost. In a November 2025 test, modern GPU encoders also matched software at real-time presets.
Read Video Transcoding Server: CPU, GPU, and What to Check Before You Rent for the choice on a rented server.
How many streams can a GPU transcode?
It depends on the card, the codec and the preset. Streaming Media’s buyers’ guide lists 16 streams of 1080p30 for a T4 card. However, a T4 cloud instance reached 60 in a vendor benchmark. Count real-time streams, not frames per second, and test your own ladder.
Read Dedicated Server for Streaming for how many viewers one port can serve.
Does a hardware transcoder lower video quality?
Not necessarily. At slow presets, software compresses better. However, at the presets that run in real time, modern hardware encoders matched software in a November 2025 academic test. Between vendors, bitrate gaps of roughly 7% to 23% appeared, and the leader changed with the codec.
Read What Is Video Transcoding? A Complete Beginner’s Guide for why each re-encode can lose a little quality.
What is the difference between NVENC and Quick Sync?
NVENC is NVIDIA’s hardware encoder, and Quick Sync is Intel’s. Quick Sync runs in Intel’s integrated graphics and, in addition, in its Arc GPUs. In the November 2025 test, Quick Sync saved about 5% of bitrate against real-time software. NVENC, however, needed about 10% more bits than Quick Sync.
Read Video Transcoding Server: CPU, GPU, and What to Check Before You Rent for the NVENC card table.
Can I add a hardware transcoder to a dedicated server?
Yes, but plan for slots, power, drivers and licence. Swify’s standard plans are CPU-only, and GPU servers are quoted to order. Intel lists no integrated graphics for the Xeon Gold 5215 and 6138, so Quick Sync is not an option on those plans.
Read Swify Dedicated Server Plans: Which Is Right for You? for the standard configurations.