Nvidia unveils RTX 4000, 4500 and 5000 GPUs
Nvidia has announced the RTX 4000, RTX 4500 and RTX 5000, the three latest workstation GPUs to use its Ada Lovelace architecture.
The firm pitches the $4,000 32GB RTX 5000, available today, as a “huge performance leap” over its predecessor, the Ampere generation RTX A5000.
The $2,250 24GB RTX 4500, due in October, is described as a “balanced performer” – while the $1,250 20GB RTX 4000, due next month, is the “most powerful single-slot GPU on the planet”.
The new cards were announced at Nvidia’s keynote at SIGGRAPH 2023.
The latest professional GPUs based on Nvidia’s Ada Lovelace architecture
All three of the new cards are based on Nvidia’s current Ada Lovelace GPU architecture, intended to provide “revolutionary performance for ray tracing and AI-based neural graphics”.
It features iterative improvements to all of the cards’ key hardware core types: CUDA cores for general GPU computing, Tensor cores for AI operations, and RT cores for hardware-accelerated ray tracing.
According to Nvidia, all three core types see improvements of “up to 2x” in raw performance over its previous Ampere architecture.
The architecture also features a new Optical Flow Accelerator, used by DLSS 3.0, the latest version of Nvidia’s Deep Learning Super Sampling technology.
Initially a render upscaling system, DLSS is used in real-time visualisation tools like D5 Render to improve viewport frame rates by rendering frames at lower resolution, then up-resing them to display resolution.
| Nvidia Ada Generation workstation GPUs | ||||||
|---|---|---|---|---|---|---|
| RTX 6000 | RTX 5000 | RTX 4500 | RTX 4000 | RTX 4000 SFF | ||
| Architecture | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace | |
| CUDA cores | 18,176 | 12,800 | 7,680 | 6,144 | 6,144 | |
| Tensor cores | 568 | 400 | 240 | 192 | 192 | |
| RT cores | 142 | 100 | 60 | 48 | 48 | |
| Compute performance FP32 (Tflops) |
91.1 Tflops | 65.3 Tflops | 39.6 Tflops | 26.7 Tflops | 19.2 Tflops | |
| GPU memory | 48GB GDDR6 |
32GB GDDR6 |
24GB GDDR6 |
20GB GDDR6 |
20GB GDDR6 |
|
| Memory bandwidth | 960 GB/s | 576 GB/s | 432 GB/s | 360 GB/s | 280 GB/s | |
| NVLInk | No | No | No | No | No | |
| Graphics bus | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | |
| TDP | 300 W | 250 W | 210 W | 130 W | 70 W | |
| Display connectors | 4 x DP 1.4 | 4 x DP 1.4 | 4 x DP 1.4 | 4 x DP 1.4 | 4 x MiniDP 1.4 | |
| Size (H x L) | 4.4” x 10.5” Dual slot |
4.4” x 10.5” Dual slot | 4.4” x 10.5” Dual slot | 4.4” x 9.5” Single slot | 2.7” x 6.6” Dual slot |
|
| Launch date | 2022 | 2023 | 2023 | 2023 | 2023 | |
| Launch price | $6,799 | $4,000 | $2,250 | $1,250 | $1,250 | |
Technical specifications
The three cards fill out the mid range of Nvidia’s Ada Lovelace generation of workstation GPUs, between the RTX 6000 and the RTX 4000 SFF.
Of the three, the RTX 4000 is noteworthy in being the first single-slot card in the line-up – and, according to Nvidia, the “most powerful single-slot GPU on the planet”.
It’s certainly a step up from the RTX 4000 SFF, its small form factor sibling, providing close to a 40% increase in single precision (FP32) compute performance, but almost doubling power consumption.
The change of form factor also provides space for full DisplayPort ports rather than Mini DisplayPort.
Its other key specs – core counts and on-board memory – are identical to the RTX 4000 SFF, aside from a significant increase in memory bandwidth.
The RTX 4500 and RTX 5000 are both dual-slot cards, and have more or less the specs you would expect from their relative positions in the product range.
Core counts, compute performance, memory capacity and memory bandwidth all increase roughly in line with price, as does power consumption.
Performance benchmarks
Unlike with some previous GPU launches, Nvidia hasn’t provided detailed performance figures for the new cards with DCC or CAD software.
The closest we have to a benchmark comparison is that the RTX 5000 is “3x” faster than the previous-gen RTX A5000 for GPU rendering with RTX Renderer, the native renderer in Nvidia’s Omniverse platform.
For third-party GPU renderers, the speed boost is lower, with Nvidia quoting an average figure of “2x”, although we don’t have any details on which applications were tested.
Prices and release dates
The RTX 4000 has a MSRP of $1,250 and will be available in September 2023. The RTX 4500 has a MSRP of $2,250 and will be available in October. The RTX 5000 has a MSRP of $4,000, and is available this week.
Find full specifications for the RTX 4000, RTX 4500 and RTX 5000 on Nvidia’s website