GP104’s Architecture

Looking at an architecture diagram for GP104, Pascal ends up looking a lot like Maxwell, and this is not by chance. After making more radical changes to their architecture with Maxwell, for Pascal NVIDIA is taking a bit of a breather. Which is not to say that Pascal is Maxwell on 16nm – this is very much a major feature update – but when it comes to discussing the core SM architecture itself, there is significant common ground with Maxwell.

We’ll start with the GP104 SM. Simply named the SM for this generation – NVIDIA has ditched the generational suffix due to the potential for confusion with the used-elsewhere SMP – the GP104 SM is very similar to the Maxwell SM. We’re still looking at a single SM partially sub-divided into four pieces, each containing a single warp scheduler that’s responsible for feeding 32 CUDA cores, 8 load/store units, and 8 Special Function Units, backed by a 64KB register file. There are two dispatch ports per warp schedule, so when an instruction stream allows it, a warp scheduler can extract a limited amount of ILP with an instruction stream by issuing a second instruction to an unused resource.

Meanwhile shared between every pair of sub-partitions is 4 texture units and the combined L1/texture cache, again unchanged from Maxwell. Finally, we have the resources shared throughout the whole SM: the 96KB shared memory, the instruction cache, and not pictured on NVIDIA’s diagrams, the 4 FP64 CUDA cores and 1 FP16x2 CUDA core.

Overall then at the diagram level the GP104 SM looks almost identical to the Maxwell SM, but with one exception: the PolyMorph Engine. Although the distinction is largely arbitrary for GP104, the PolyMorph Engine has been moved up a level; it’s no longer part of the SM, but rather part of the newly re-introduced TPC, which itself sits between the GPC and the SM.

The TPC exists because although GP104 still has a 1:1 ratio between PolyMorph Engines and SMs, the Pascal architecture itself allows for different SM configurations, which is in turn used on GP100 to allow it to have multiple smaller SMs of 64 CUDA Cores. For GP100 the TPC allows for multiple SMs to share a PolyMorph Engine, but for GP104 there’s no sharing involved. To that end the TPC as an organizational unit technically exists across all Pascal parts, but it has no real significance for GP104. In fact it doesn’t even have a real name; NVIDIA reused the acronym from earlier DX10 architectures, where the TPC was the name assigned to the Texture Processor Cluster.

Looking at the bigger picture of the complete GP104 GPU, the similarities continue between GP104 and GM204. GP104’s SMs are clustered five-a-piece inside of the GPC, with each cluster sharing a single Raster Engine. Overall there are 4 such GPCs, giving us 20 SMs altogether. Compared to GM204 then, we’re looking at the same number of GPCs, with each GPC having gained 1 SM.

Things get more interesting when we look at the back end of the rendering/execution pipeline, which is comprised of the L2 cache, ROPs, and memory controllers. The ROP/L2 count has not changed relative to GM204 – we still have 64 ROPs paired up with a total of 2MB of L2 cache – however the memory controller count has. And with it the logical configuration of the ROP/L2 blocks have changed as well.

Whereas GM204 had 4 64bit GDDR5 memory controllers, each connected to 2 or 4 memory chips, GP104 breaks that down further to 8 32bit GDDR5X memory controllers, each of which is connected to 1 memory chip on GTX 1080. I’ll go into greater detail on GDDR5X a bit later, but the significance of this backend organizational change has to do with the introduction of GDDR5X. Because GDDR5X reads and writes data in 64B amounts (versus 32B amounts on GDDR5), NVIDIA has reorganized the memory controllers to ensure that each memory controller still operates on the same amount of data. With GDDR5 they teamed up two GDDR5 channels to get 64B operations, whereas with GDDR5X this can be accomplished with a single memory channel.

This in turn is where the ROP reorganization comes from. As there’s a 1:1 relationship between ROP partitions and memory controllers, the 64 ROPs are now broken up into 8 partitions for GP104, as opposed to 4 partitions on GM204. There are some performance tradeoffs that come from having more ROP partitions, but to the best of my knowledge these should not be significant.

Meanwhile the new GDDR5X memory controllers are also backwards compatible with traditional GDDR5, which in turn is used to drive the GTX 1070 with its 8Gbps GDDR5. The difference in operation between GDDR5 and GDDR5X does make the ROP situation a bit trickier overall for NVIDIA’s architects – now they need to be able to handle two different memory access patterns – though for NVIDIA this isn’t a wholly new problem. Previous generation architectures have supported both GDDR5 and DDR3, the two of which have their own differences in memory access patterns.

In a by-the-numbers comparison then, Pascal does not bring any notable changes in throughput relative to Maxwell. CUDA cores, texture units, PolyMorph Engines, Raster Engines, and ROPs all have identical theoretical throughput-per-clock as compared to Maxwell. So on a clock-for-clock, unit-for-unit basis, Pascal is not any faster on paper. And while NVIDIA does not disclose the size/speed of most of their internal datapaths, so far I haven’t seen anything to suggest that these have radically changed. This continuity means that outside of its new features, GP104 behaves a lot like GM204. Though it should be noted that real world efficiency isn’t quite as cut and dry, as various factors such as the increased SM count and changes in memory technology can greatly influence this.

GP104: The Heart of GTX 1080 FP16 Throughput on GP104: Good for Compatibility (and Not Much Else)
Comments Locked

200 Comments

View All Comments

  • patrickjp93 - Wednesday, July 20, 2016 - link

    That doesn't actually support your point...
  • Scali - Wednesday, July 20, 2016 - link

    Did I read a different article?
    Because the article that I read said that the 'holes' would be pretty similar on Maxwell v2 and Pascal, given that they have very similar architectures. However, Pascal is more efficient at filling the holes with its dynamic repartitioning.
  • mr.techguru - Wednesday, July 20, 2016 - link

    Just Ordered the MSI GeForce GTX 1070 Gaming X , way better than 1060 / 480. NVidia Nail it :)
  • tipoo - Wednesday, July 20, 2016 - link

    " NVIDIA tells us that it can be done in under 100us (0.1ms), or about 170,000 clock cycles."

    Is my understanding right that Polaris, and I think even earlier with late GCN parts, could seamlessly interleave per-clock? So 170,000 times faster than Pascal in clock cycles (less in total time, but still above 100,000 times faster)?
  • Scali - Wednesday, July 20, 2016 - link

    That seems highly unlikely. Switching to another task is going to take some time, because you also need to switch all the registers, buffers, caches need to be re-filled etc.
    The only way to avoid most of that is to duplicate the whole register file, like HyperThreading does. That's doable on an x86 CPU, but a GPU has way more registers.
    Besides, as we can see, nVidia's approach is fast enough in practice. Why throw tons of silicon on making context switching faster than it needs to be? You want to avoid context switches as much as possible anyway.

    Sadly AMD doesn't seem to go into any detail, but I'm pretty sure it's going to be in the same ballpark.
    My guess is that what AMD calls an 'ACE' is actually very similar to the SMs and their command queues on the Pascal side.
  • Ryan Smith - Wednesday, July 20, 2016 - link

    Task switching is separate from interleaving. Interleaving takes place on all GPUs as a basic form of latency hiding (GPUs are very high latency).

    The big difference is that interleaving uses different threads from the same task; task switching by its very nature loads up another task entirely.
  • Scali - Thursday, July 21, 2016 - link

    After re-reading AMD's asynchronous shader PDF, it seems that AMD also speaks of 'interleaving' when they switch a graphics CU to a compute task after the graphics task has completed. So 'interleaving' at task level, rather than at instruction level.
    Which would be pretty much the same as NVidia's Dynamic Load Balancing in Pascal.
  • eddman - Thursday, July 21, 2016 - link

    The more I read about async computing in Polaris and Pascal, the more I realize that the implementations are not much different.

    As Ryan pointed out, it seems that the reason that Polaris, and GCN as a whole, benefit more from async is the architecture of the GPU itself, being wider and having more ALUs.

    Nonetheless, I'm sure we're still going to see comments like "Polaris does async in hardware. Pascal is hopeless with its software async hack".
  • Matt Doyle - Wednesday, July 20, 2016 - link

    Typo in the lead sentence of HPC vs. Consumer: Divergence paragraph: "Pascal in an architecture that..."

    "is" instead of "in"
  • Matt Doyle - Wednesday, July 20, 2016 - link

    Feeding Pascal page, "GDDR5X uses a 16n prefetch, which is twice the size of GDDR5’s 8n prefect."

    Prefect = prefetch

Log in

Don't have an account? Sign up now