Slashdot Mirror
NVIDIA Unveils Next-Gen Turing Quadro RTX Professional Graphics Cards (hothardware.com)
MojoKid shares a report from Hot Hardware: We been hearing a lot about NVIDA's next-generation GPU architecture since late last year, and today NVIDIA is announcing the first products based on Turing. NVIDIA is targeting the professional graphics market first with its new Quadro RTX 8000, RTX 6000 and RTX 5000 GPUs. Turing GPU architecture replaces Pascal, which has served both the consumer and professional markets since 2016. But as its 8th generation GPU architecture, NVIDIA is ushering in a number of advances with Turing and it's billed as the "world's first ray-tracing GPU." When it comes to content creators, NVIDIA claims that with the power of Turing, "applications can simulate the physical world at 6x the speed of the previous Pascal generation."
Getting down to brass tacks, the entry-level Quadro RTX 5000 has 3,072 CUDA cores, 384 Tensor cores, and will come with 16GB of 14Gbps GDDR6 memory. Its ray-tracing performance is dialed in at 6 GigaRays/sec, according to NVIDIA. Both the Quadro RTX 6000 and RTX 8000 have 4,608 CUDA cores and 576 Tensor cores; the only difference between the two is that the former has 24GB of GDDR6, while the latter doubles that to 48GB. Ray-tracing performance for both of these GPUs tops out at 10 GigaRays/sec. NVIDIA is also claiming up to 16TFLOPs compute performance for the Quadro RTX 8000. NVIDIA's new Quadro GPUs will also be among the first to support both USB-C and VirtualLink for next-generation virtual reality and mixed reality headsets. Other VirtualLink backers include AMD, Oculus, Microsoft and Valve. The Quadro RTX 5000, RTX 6000 and RTX 8000 will all be available during the fourth quarter of 2018 priced at $2,300, $6,300 and $10,000 respectively.
Getting down to brass tacks, the entry-level Quadro RTX 5000 has 3,072 CUDA cores, 384 Tensor cores, and will come with 16GB of 14Gbps GDDR6 memory. Its ray-tracing performance is dialed in at 6 GigaRays/sec, according to NVIDIA. Both the Quadro RTX 6000 and RTX 8000 have 4,608 CUDA cores and 576 Tensor cores; the only difference between the two is that the former has 24GB of GDDR6, while the latter doubles that to 48GB. Ray-tracing performance for both of these GPUs tops out at 10 GigaRays/sec. NVIDIA is also claiming up to 16TFLOPs compute performance for the Quadro RTX 8000. NVIDIA's new Quadro GPUs will also be among the first to support both USB-C and VirtualLink for next-generation virtual reality and mixed reality headsets. Other VirtualLink backers include AMD, Oculus, Microsoft and Valve. The Quadro RTX 5000, RTX 6000 and RTX 8000 will all be available during the fourth quarter of 2018 priced at $2,300, $6,300 and $10,000 respectively.
Note that if you do RTFA, it's still badly explained, wrongly written but you get 48GB on the top end Quadro (not in the article : it's up from the older 24GB on GDDR5, because GDDR6 is available with twice the density)
If you have two Quadro RTX 8000 linked with a bridge, you get a real and genuine 96GB memory space in the same way it happens on modern dual CPU systems. Coherent, seamless but bandwidth is much slower if you go over the border of course. This feature was already available on Quadro GP100 (Pascal with fast double precision floating point ; 16GB HBM2 on each of two cards, combined 32GB)
It is a NUMA system, much like Opteron and dual Xeon E5 before.
Accessorily the article says the bridge feature (NVLink) is available on all three new boards, so two RTX 5000 is 16GB + 16GB, two RTX 6000 is 24GB + 24GB.
On the picture with the boss showing two artistic (stylized, unrealistic) representations of Pascal GP102 vs Turing, it's written : 24GB for the GP102, 48GB + 48GB for the large Turing.
(GP102 is using on GTX 1080 Ti with 11GB GDDR5X RAM, on Titan Xp with 12GB RAM, on Quadro P6000 with 24GB RAM)