FP64 Performance and Separating Radeon VII from Radeon Instinct MI50

One of the interesting and amusing consequences of the Radeon VII launch is that for the first time in quite a while, AMD has needed to seriously think about how they’re going to differentiate their consumer products from their workstation/server products. While AMD has continued to offer workstation and server hardware via the Radeon Pro and Radeon Instinct series, the Vega 20 GPU is AMD’s first real server-grade GPU in far too long. So, while those products were largely differentiated by the software features added to their underlying consumer-grade GPUs, Radeon VII brings some new features that aren’t strictly necessary for consumers.

It may sound like a trivial matter – clearly AMD should just leave everything enabled – but as the company is trying to push into the higher margin server business, prosumer products like the Radeon VII are in fact a tricky proposition. AMD needs to lock away enough of the server functionality of the Vega 20 GPU that they aren’t selling the equivalent of a Radeon Instinct MI50 for a fraction of the price. On the other hand, it’s in their interest to expose some of these features in order to make the Radeon VII a valuable card in its own right (one that can justify a $699 price tag), and to give developers a taste of what AMD’s server hardware can do.

Case in point is the matter of FP64 performance. As we noted in our look at the Vega 20 GPU, Vega 20’s FP64 performance is very fast: it’s one-half the FP32 rate, or 6.9 TFLOPS. This is one of the premium features of Vega 20, and since Radeon VII was first announced back at CES, the company has been struggling a bit to decide how much of that performance to actually make available to the Radeon VII. At the time of its announcement, we were told that the Radeon VII would have unrestricted (1/2) FP64 performance, only to later be told that it would be 1/8. Now, with the actual launch of the card upon us, AMD has made their decision: they’ve split it down the middle and are doing a 1/4 rate.

Looking to clear things up, AMD put out a statement:

The Radeon VII graphics card was created for gamers and creators, enthusiasts and early adopters. Given the broader market Radeon VII is targeting, we were considering different levels of FP64 performance. We previously communicated that Radeon VII provides 0.88 TFLOPS (DP=1/16 SP). However based on customer interest and feedback we wanted to let you know that we have decided to increase double precision compute performance to 3.52 3.46 TFLOPS (DP=1/4SP).

If you looked at FP64 performance in your testing, you may have seen this performance increase as the VBIOS and press drivers we shared with reviewers were pre-release test drivers that had these values already set. In addition, we have updated other numbers to reflect the achievable peak frequency in calculating Radeon VII performance as noted in the [charts].

The end result is that while the Radeon VII won’t be as fast as the MI60/MI50 when it comes to FP64 compute, AMD is going to offer the next best thing, just one step down from those cards.

At 3.5 TLFLOPS of theoretical FP64 performance, the Radeon VII is in a league of its own for the price. There simply aren’t any other current-generation cards priced below $2000 that even attempt to address the matter. All of NVIDIA’s GeForce cards and all of AMD’s other Radeon cards straight-up lack the necessary hardware for fast FP64. The next closest competitor to the Radeon VII in this regard is NVIDIA’s Titan V, at more than 4x the price.

It’s admittedly a bit of a niche market, especially when so much of the broader industry focus is on AI and neural network performance. But there’s none the less going to be some very happy data scientists out there, especially among academics.

AMD Server Accelerator Specification Comparison
  Radeon VII Radeon Instinct
MI50
Radeon Instinct
MI25
FirePro S9170
Stream Processors 3840
(60 CUs)
3840
(60 CUs)
4096
(64 CUs)
2816
(44 CUs)
ROPs 64 64 64 64
Base Clock 1450MHz 1450MHz 1400MHz -
Boost Clock 1750MHz 1746MHz 1500MHz 930MHz
Memory Clock 2.0Gbps HBM2 2.0Gbps HBM2 1.89Gbps HBM2 5Gbps GDDR5
Memory Bus Width 4096-bit 4096-bit 2048-bit 512-bit
Half Precision 27.6 TFLOPS 26.8 TFLOPS 24.6 TFLOPS 5.2 TFLOPS
Single Precision 13.8 TFLOPS 13.4 TFLOPS 12.3 TFLOPS 5.2 TFLOPS
Double Precision 3.5 TFLOPS
(1/4 rate)
6.7 TFLOPS
(1/2 rate)
768 GFLOPS
(1/16 rate)
2.6 TFLOPS
(1/2 rate)
DL Performance ? 53.6 TFLOPS 12.3 TFLOPS 5.2 TFLOPS
VRAM 16GB 16GB 16GB 32GB
ECC No Yes (full-chip) Yes (DRAM) Yes (DRAM)
Bus Interface PCIe Gen 3 PCIe Gen 4 PCIe Gen 3 PCIe Gen 3
TDP 300W 300W 300W 275W
GPU Vega 20 Vega 20 Vega 10 Hawaii
Architecture Vega
(GCN 5)
Vega
(GCN 5)
Vega
(GCN 5)
GCN 2
Manufacturing Process TSMC 7nm TSMC 7nm GloFo 14nm TSMC 28nm
Launch Date 02/07/2019 09/2018 06/2017 07/2015
Launch Price (MSRP) $699 - - $3999

Speaking of AI, it should be noted that machine learning performance is another area where AMD is throttling the card. Unfortunately, more details aren’t available at this time. But given the unique needs of the ML market, I wouldn’t be surprised to find that INT8/INT4 performance is held back a bit on the Radeon VII. Or for that matter certain FP16 dot products.

Also on the chopping block is full-chip ECC support. Thanks to the innate functionality of HBM2, all Vega cards already have free ECC for their DRAM. However Vega 20 takes this one step further with ECC protection for its internal caches, and this is something that the Radeon VII doesn’t get access to.

Finally, Radeon VII also cuts back a bit on Vega 20’s off-chip I/O features. Though AMD hasn’t made a big deal of it up to now, Vega 20 is actually their first PCI-Express 4.0-capable GPU, and this functionality is enabled on the Radeon Instinct cards. However for Radeon VII, this isn’t being enabled, and the card is being limited to PCIe 3.0 speeds (so future Zen 2 buyers won’t quite have a PCIe 4.0 card to pair with their new CPU). Similarly, the external Infinity Fabric links for multi-GPU support have been disabled, so the Radeon VII will only be a solo act.

On the whole, there’s nothing very surprising about AMD’s choices here, especially given Radeon VII’s target market and target price. But these are notable exclusions that are going to matter to certain users. And if not to drive those users towards a Radeon Instinct, then they’re sure to drive those users towards the inevitable Vega 20-powered Radeon Pro.

Vega 20: Under The Hood Meet the AMD Radeon VII
POST A COMMENT

297 Comments

View All Comments

  • i4mt3hwin - Thursday, February 07, 2019 - link

    So FP64 is 1:4 and not 1:8 or 1:2 as previously known? Reply
  • tipoo - Thursday, February 07, 2019 - link

    Yep, looks like they changed the cap in vBIOS based on feedback.

    Which also means they could have uncapped it, but it's still cool that they did that.
    Reply
  • Ganimoth - Thursday, February 07, 2019 - link

    Does that mean it could be potentially unlocked by some bios mod? Reply
  • tipoo - Friday, February 08, 2019 - link

    I hope so! Reply
  • Hul8 - Thursday, February 07, 2019 - link

    I don't think it was ever reported or assumed to be 1/2 - that best possible ratio is only for the pro MI50 part. Early reports said 1/16. Reply
  • Ryan Smith - Thursday, February 07, 2019 - link

    For what it's worth, when we first asked AMD about it back at CES, FP64 performance wasn't among the features they were even throttling/holding back on. So for a time, 1/2 was on the table. Reply
  • GreenReaper - Thursday, February 07, 2019 - link

    So it was *your* fault! ;-p Reply
  • PerterLustig - Thursday, February 07, 2019 - link

    Thanks Ryan!

    Any word why only 64 ROP Cluster? 64 ROPs...that's basically FERMI era.
    Reply
  • del42sa - Friday, February 08, 2019 - link

    it must have 64 ROP´s disabled because according Anandtech Vega 7 supposed to have 128 ROP´s. LoL

    and more secret weapons ( Primitive Shader and NGG fast Path )
    Reply
  • tipoo - Friday, February 08, 2019 - link

    It's essentially a die shrunk Vega 64. It's still GCN, so all the GCN limits are still in play. Reply

Log in

Don't have an account? Sign up now