NVIDIA P100 comes to Quadro
NVIDIA refreshes the Quadro line with new Pascal offerings including a P100 based part with HBM2.
At the start of the SOLIDWORKS World conference this week, NVIDIA took the cover off of a handful of new Quadro cards targeting professional graphics workloads. Though the bulk of NVIDIA’s discussion covered lower cost options like the Quadro P4000, P2000, and below, the most interesting product sits at the high end, the Quadro GP100.
As you might guess from the name alone, the Quadro GP100 is based on the GP100 GPU, the same silicon used on the Tesla P100 announced back in April of 2016. At the time, the GP100 GPU was specifically billed as an HPC accelerator for servers. It had a unique form factor with a passive cooler that required additional chassis fans. Just a couple of months later, a PCIe version of the GP100 was released under the Tesla GP100 brand with the same specifications.
Today that GPU hardware gets a third iteration as the Quadro GP100. Let’s take a look at the Quadro GP100 specifications and how it compares to some recent Quadro offerings.
Quadro GP100 | Quadro P6000 | Quadro M6000 | Full GP100 | |
---|---|---|---|---|
GPU | GP100 | GP102 | GM200 | GP100 (Pascal) |
SMs | 56 | 60 | 48 | 60 |
TPCs | 28 | 30 | 24 | (30?) |
FP32 CUDA Cores / SM | 64 | 64 | 64 | 64 |
FP32 CUDA Cores / GPU | 3584 | 3840 | 3072 | 3840 |
FP64 CUDA Cores / SM | 32 | 2 | 2 | 32 |
FP64 CUDA Cores / GPU | 1792 | 120 | 96 | 1920 |
Base Clock | 1303 MHz | 1417 MHz | 1026 MHz | TBD |
GPU Boost Clock | 1442 MHz | 1530 MHz | 1152 MHz | TBD |
FP32 TFLOPS (SP) | 10.3 | 12.0 | 7.0 | TBD |
FP64 TFLOPS (DP) | 5.15 | 0.375 | 0.221 | TBD |
Texture Units | 224 | 240 | 192 | 240 |
ROPs | 128? | 96 | 96 | 128? |
Memory Interface | 1.4 Gbps 4096-bit HBM2 |
9 Gbps 384-bit GDDR5X |
6.6 Gbps 384-bit GDDR5 |
4096-bit HBM2 |
Memory Bandwidth | 716 GB/s | 432 GB/s | 316.8 GB/s | ? |
Memory Size | 16GB | 24 GB | 12GB | 16GB |
TDP | 235 W | 250 W | 250 W | TBD |
Transistors | 15.3 billion | 12 billion | 8 billion | 15.3 billion |
GPU Die Size | 610mm2 | 471 mm2 | 601 mm2 | 610mm2 |
Manufacturing Process | 16nm | 16nm | 28nm | 16nm |
There are some interesting stats here that may not be obvious at first glance. Most interesting is that despite the pricing and segmentation, the GP100 is not the de facto fastest Quadro card from NVIDIA depending on your workload. With 3584 CUDA cores running at somewhere around 1400 MHz at Boost speeds, the single precision (32-bit) rating for GP100 is 10.3 TFLOPS, less than the recently released P6000 card. Based on GP102, the P6000 has 3840 CUDA cores running at something around 1500 MHz for a total of 12 TFLOPS.
GP100 (full) Block Diagram
Clearly the placement for Quadro GP100 is based around its 64-bit, double precision performance, and its ability to offer real-time simulations on more complex workloads than other Pascal-based Quadro cards can offer. The Quadro GP100 offers 1/2 DP compute rate, totaling 5.2 TFLOPS. The P6000 on the other hand is only capable of 0.375 TLOPS with the standard, consumer level 1/32 DP rate. Inclusion of ECC memory support on GP100 is also something no other recent Quadro card has.
Raw graphics performance and throughput is going to be questionable until someone does some testing, but it seems likely that the Quadro P6000 will still be the best solution for that by at least a slim margin. With a higher CUDA core count, higher clock speeds and equivalent architecture, the P6000 should run games, graphics rendering and design applications very well.
There are other important differences offered by the GP100. The memory system is built around a 16GB HBM2 implementation which means more total memory bandwidth but at a lower capacity than the 24GB Quadro P6000. Offering 66% more memory bandwidth does mean that the GP100 offers applications that are pixel throughput bound an advantage, as long as the compute capability keeps up on the backend.
This also marks the first time we have seen NVLink implemented at an add-in card level, leaving the possibility for the technology to be used for multi-GPU configurations for consumers in 2017. It integrates much like an SLI connection on the Quadro GP100.
Pricing hasn’t been disclosed as NVIDIA only says that its partners and system integrators will be offering up options starting in March. The Quadro P6000 sells for $4500 on Amazon today, so I expect the GP100 to be above that, likely in the $8000-10k range. It is interesting to see NVIDIA’s strategy with GP100 and the P100 GPU pan out. With nearly a year on the market already in the form of HPC-class Tesla products, my guess is that NVIDIA was waiting for that high-margin, high-dollar market to saturate out some before offering the same performance on a lower priced, more easily accessible product like the Quadro GP100. This means that any wafer capacity still rolling through a 16nm fab facility can be utilized, even if it happens to be at a lower ASP. Will this product ever be priced at a point that makes sense for consumers and the GeForce product line? I doubt it; if we see any other Pascal-based GeForce product (non-Titan) it will be from the GP102 lot, a P6000/Titan X (Pascal) priced down to take on AMD’s upcoming Vega GPU.
Complete Quadro Pascal Lineup
Along with the flagship and highly interesting Quadro GP100, NVIDIA announced a line of Quadro cards from the P4000 to the P400, with varying levels of performance, memory capacity and features.
Quadro P4000 | Quadro P2000 | Quadro P1000 | Quadro P600 | Quadro P400 | |
---|---|---|---|---|---|
GPU | GP104 | GP106 | GP107 | GP107 | GP107 |
FP32 CUDA Cores / GPU | 1792 | 1024 | 640 | 384 | 256 |
Base Clock | ? | ? | ? | ? | ? |
GPU Boost Clock | 1480 MHz | 1470 MHz | 1400 MHz | 1430 MHz | 1170 MHz |
FP32 TFLOPS (SP) | 5.3 | 3.0 | 1.8 | 1.1 | 0.6 |
Texture Units | 112 | 64 | 32 | 32 | 32 |
ROPs | 64 | 40 | 40 | 24 | 16 |
Memory Interface | 256-bit | 160-bit | 128-bit | 128-bit | 128-bit |
Memory Size | 8GB | 5GB | 4GB | 2GB | 2GB |
TDP | 105 W | 75 W | 47 W | 40 W | 30 W |
Manufacturing Process | 16nm | 16nm | 16nm | 16nm | 16nm |
All of these cards are single slot designs, with off-the-shelf GPUs that we have seen before. They replace Kepler and Maxwell cards that currently sit at these price points in the Quadro stack, but with relative performance increases ranging from 1.3x to 2.0x.
I will be taking a closer look at many of these P-series Quadro cards in the upcoming weeks, so check back for full reviews.
When does the Titan GP100
When does the Titan GP100 come out?
As soon as you come up with
As soon as you come up with the $8000-10k, they will be happy to place you on the list. Do not expect the drivers that come with this SKU to be tuned for gaming workloads as the drivers for this SKU will be tuned for accuracy not FPS. This is a Quadro SKU and not a gaming SKU!
I know how their product
I know how their product lineup works.
The real selling point for the original Titan, Titan Z and Titan Black was the cheap DP FLOPS, something the last two Titans didnt deliver because Nvidia was stuck on 28nm longer than they wanted to be.
They used the BS marketing smokescreen of “its a machine learning card” for Titan Maxwell and Titan GP102. Anyone who knows anything about semiconductor design, engineering and manufacturing could see through that lie though. It was about low yield on a new process node with expensive HBM on an interposer after being stuck with no space for FP64 cores on a 28nm planar chip.
Quadros are typically ~5x the price of Titan cards with the same chip, so we will have to wait until the margin on the whole GP100 interposer package can be determined.
The Titan Z was for
The Titan Z was for professional DP compute non graphics workloads with its lack of more expensive Quadro pro graphics driver features but still having more DP performance than a consumer gaming GPU. Even AMD took that route with it’s Radeon Pro Duo with developers getting the professional level drivers that allowed them to develop applications on the Radeon Pro Duo that could be directly transferred to the more expensive Radeon Pro WX SKUs that cost 2+ times as much.
So the Titan Z and Radeon Pro Duo have their lower cost development/compute uses but they can not be used for actual production runs as that requires the full Quadro/Radeon Pro WX(Formally Firepro) SKUs that have all the error correction bells and whistles required for actual production work.
The Radeon Pro Duo has greatly come down in price to $799 and is a great deal for software developers needing the pro grade drivers to develop for the professional Radeon Pro WX market. The Titan Z is still running in the $1500 price range down from $3000.
The Bullshit marketing is well marketing by definition and that “profession” has its roots traced back directly to the snake oil salesmen and the first lie ever told.
I have no problems with both AMD and Nvidia making full use of their production yields by binning defective parts into useful SKUs to serve specialized markets at lower costs to end users with a need for these products. It’s not all about gamers and gaming needs and both AMD and Nvidia are in business to make a resonable profit for their investors.
As I stated, I really don’t
As I stated, I really don't think NVIDIA is interested in putting out an HBM2 based GP100 part for GeForce gamers this year.
i guess we will have to wait
i guess we will have to wait for the gpu tech con in may. who knows they might announce it or save it for volta.
Titans were always a better
Titans were always a better value for DP FLOPS for real 3D work than they were for games though.
I guess it really depends on what their actual margins for a GP100 including the interposer and HBM stacks is.
I distinctly remember Titan Z being $3k and then dropping to $1.5k as well.
Pardon my ignorance, but what
Pardon my ignorance, but what programs still prefer quadro over geforce to a significant degree.