Nvidia announced the RTX 5880 Ada graphics AIB, which aligns with US export policies and adheres to a maximum TPP limit of 4800 points. The AIB has a modified AD102 GPU with 14,080 CUDA cores and 440 Tensor cores, marking a 22% reduction in cores compared to the RTX 6000. The AIB has 48GB GDDR6 memory running at 20 Gbps, offering up to 960 GB/sec bandwidth, similar to the RTX 6000.
What do we think? The RTX 5880 is clearly targeted to satisfy the trade restrictions on high-performance computing devices sold to China. Such restrictions are symbolic and have limited impact of the intended consumers who can find multiple routes and channels to obtain devices, just as Iran, Russia, and other restricted countries have.
Nvidia made the RTX 5880 Ada graphics AIB official after it first showed up in a driver release a month ago, and its core specifications have been slightly reduced, with the rest of the specifications remaining the same. The RTX 5880 is an AIB designed to comply with the US export policies, and it fits within the maximum TPP limit of 4,800 points.
The RTX 5880 Ada has a modified AD102 GPU with 14,080 CUDA cores and 440 Tensor cores—a reduction of 22% cores compared to the RTX 6000. Based on its 69.3 TFLOPs of FP32 performance and its 14,800 cores, the clock is about 2.46 MHz. It also provides 1,108 TFLOPs of Tensor performance and 160.2 TFLOPS of RT Core performance. All that is about a 24% decrease in FP32 and tensor-core performance compared to the RTX 6000 Ada.
The RTX 5880 has 48GB GDDR6 memory running at speeds of 20 Gbps for up to 960 GB/sec bandwidth, the same as the RTX 6000. Power consumption, however, is reduced by 15W (285W) compared to the 300W of the RTX 6000. The AIB is a dual-slot size with active- cooling and offers four DisplayPort 1.4a outputs of 4096×2160 at 120 Hz, or four 5120×2880 at 60 Hz, or two 7680×4320 at 60 Hz refresh. And, like the RTX 6000, it also has triple encore/decode engines.
The Nvidia workstation line now looks like Table 1:
| AIB | RTX 6000 | RTX 5880 | RTX 5000 | RTX 4500 | RTX 4000 | RTX 4000 SFF |
| Architecture | Lovelace | Lovelace | Lovelace | Lovelace | Lovelace | Lovelace |
| GPU Cores | 18,176 | 14,080 | 12,800 | 7,680 | 6,144 | 6,144 |
| Tensor Cores | 568 | 440 | 500 | 300 | 192 | 192 |
| Clock (GHz) | 2.5 | ~2.46 | ~2.60 | ~2.60 | ~2.20 | ~1.70 |
| Single Precision TFLOPS | 91.1 | 69.3 | 65.3 | 39.6 | 26.7 | 19.2 |
| Tensor TFLOPS | 1,458 | 1 108 | 1,044 | 634 | 327 | 306.8 |
| Memory Bus (bits) | 384 | 384 | 256 | 192 | 160 | 160 |
| Memory Bandwidth (GB/sec) | 960 | 960 | 576 | 432 | 360 | 320 |
| VRAM (GDDR6) | 48GB | 48GB | 32GB | 24GB | 20GB | 20GB |
| TDP (W) | 300 | 285 | 250 | 210 | 130 | 70 |
| Fab Process | TSMC 4N | TSMC 4N | TSMC 4N | TSMC 4N | TSMC 4N | TSMC 4N |
| Die Size (mm2) | 608 (AD102) | 608 (AD102) | 608 (AD102) | 295 (AD104) | 295 (AD104) | 295 (AD104) |
| Price | $6,800 US | N/A | $4,000 US | $2,250 US | $1,250 US | $1,250 US |
| Date | Q1 2023 | Q1 2024 | Q3 2023 | Q4 2023 | Q4 2023 | Q1 2023 |
Nvidia workstation AIBs.
Nvidia might make a quick announcement of the RTX 5880 at CES.
The RTX 5880 may be offered only in China and could be more expensive than the RTX 6000 (around $6,800) because of the limited production.