AMD have released their new HPC card, the Radeon Vega Frontier Edition, which Jim told you about earlier this week. The air cooled version is available now, with an MSRP of $999USD followed by a water-cooled edition arriving in Q3 with price tag of $1499.
The specs they list for the cards are impressive and compare favourably to NVIDIA's P100 which is the card AMD tested against, offering higher TFLOPS for both FP32 and FP16 operations though the memory bandwidth lags a little behind.
| Radeon Vega Frontier Edition |
Quadro GP100 | |
|---|---|---|
| GPU | Vega | GP100 |
| Peak/Boost Clock | 1600 MHz | 1442 MHz |
| FP32 TFLOPS (SP) | 13.1 | 10.3 |
| FP64 TFLOPS (DP) |
0.819 |
5.15 |
| Memory Interface | 1.89 Gb/s 2048-bit HBM2 |
1.4 Gbps 4096-bit HBM2 |
| Memory Bandwidth | 483 GB/s | 716 GB/s |
| Memory Size | 16GB HBC* | 16GB |
| TDP | 300 W air, 375 W water | 235 W |
The memory size for the Vega is interesting, HBC is AMDs High Bandwidth Cache Controller which not only uses the memory cache more effectively but is able to reach out to other high performance system memory for help. AMD states that the Radeon Vega Frontier Edition has the capability of expanding traditional GPU memory to 256TB; perhaps allowing new texture mods for Skyrim or Fallout! Expect to see more detail on this feature once we can get our hands on a card to abuse, nicely of course.
AMD used the DeepBench Benchmark to provide comparative results, the AMD Vega FE system used a dual socketed system with Xeon E5 2640v4s @ 2.4Ghz 10C/20T, 32GB DDR4 per socket, on Ubuntu 16.04 LTS with ROCm 1.5, and OpenCL 1.2, the NVIDIA Tesla P100 system used the same hardware with the CuDNN 5.1, Driver 375.39 and Cuda version 8.0.61 drivers. Those tests showed the AMD system completing the benchmark in 88.7ms, the Tesla P100 completed in 133.1 ms, quite an impressive lead for AMD. Again, there will be much more information on performance once the Vega FE can be tested.
Read on to hear about the new card in AMD's own words, with links to their sites.
Today, AMD unleashed the world’s fastest graphics card for machine learning development and advanced visualization workloads, Radeon Vega Frontier Edition, available now.
Designed to empower the next generation of data scientists, game designers and visualization professionals, Radeon Vega Frontier Edition allows users to tackle massive datasets with ease, and scored up to 33 percent faster than the competition in the DeepBench benchmark measuring the performance of basic operations involved in training deep neural networks.
The first graphics card on the market to feature AMD’s “Vega” GPU architecture, Radeon Vega Frontier Edition together with the power of Radeon Pro Software enables data scientists, creators and developers to pursue new frontiers in their fields with:
- Unmatched performance and TCO in machine learning applications together with AMD’s open-source, fully scalable ROCm software platform. Developers can now use the power of the “Vega” architecture for machine learning algorithm development on the Radeon Vega Frontier Edition faster than with any other GPU on the marketi, before deploying it out to massive servers equipped with Radeon Instinct accelerators.
- Advanced photo realistic rendering performance to drive increasingly large and complex models and to deploy real-time visualization and physically-based rendering. Together with Radeon ProRender, AMD’s powerful physically-based rendering engine that enables creative professionals to produce stunningly photorealistic images, Radeon Vega Frontier Edition allows professionals to render assets up to 172 percent faster than the comparable competitor card. A visualization powerhouse, the GPU offers exceptional multi-GPU scaling, with 91 percent faster rendering using two Radeon Vega Frontier Edition GPUs.
- Accelerated game design and immersive workflows by providing a single GPU that is optimized for every stage of a game developer’s workflow, including everything from asset production to playtesting and performance optimization. With the Radeon Pro Settings user interface, users can seamlessly switch between “Radeon Pro Mode” and “Gaming Mode” to alternate between development on animation applications like Autodesk Maya and performance optimizations with free, open source tools available through AMD’s GPUOpen initiative.
- Groundbreaking VR content creation to provide the needed performance across all stages of VR content creation workflows, from the initial design stage to the final VR experience. AMD’s fastest Radeon VR Ready Creator graphics card ever, Radeon Vega Frontier Edition achieves the maximum possible score in the SteamVR benchmark, up to 21 percent higher than the multi-GPU Radeon Pro Duo solution. When used with Radeon Loom, creators can stitch high-resolution video in real time at up to 8K x 4K.
Radeon Vega Frontier Edition graphics cards are available from etailers in select regions today with an SEP of $999 USD for the air-cooled edition. The water-cooled edition is expected to launch in Q3 with an SEP of $1499. You can find more information here. Following the beta release of the Radeon ProRender plug-in for Blender and the Radeon ProRender add-in for SOLIDWORKS earlier this year, the first production versions are freely available for download here. AMD is also releasing a beta version of its Radeon ProRender Game Engine Importer, a tool that allows you to easily import your geometry and materials using Radeon ProRender from applications like SOLIDWORKS right into Unreal Engine and view them in VR. Check out the following blog to learn more.
Is that FP64 number correct
Is that FP64 number correct as that TFLOP count a bit pitiful as it would be comparable to a K6000. While there are limited dual precision workloads those who need that precision are usually the ones that have large amounts of capital to spend on these cards. Or are they targeting the titan Xp/1080ti instead of GP100?
I think they are!
I’ve heard
I think they are!
I’ve heard rumors about next gen AMD card having roughly half FP64 of FP32 number… Which would be darn impressive. But maybe Instinct may have higher FP64 than this version
AMD did not include any
AMD did not include any double precision numbers but Jeff over at Tech Report did some asking and heard "the card will perform FP64 operations at 1/16 the FP32 rate, or 819 GFLOPS."
They are conveniently picking
They are conveniently picking hardware and software to make it shine against competitor product, like comparing Vega FE to the Titan Xp in professional software where a GP104 based Quadro P5000 is faster than a Titan Xp or Vega FE and that with a 180W TDP while having a similar price (it was the same before AMD dropped the price by $200).
And this one may be the same as a Titan Xp is considerably faster (have seen more than 20%) than a P100 in some of the DeepBench tests.
They also tested P100 with uuDNN 5.1 while version 6 is available since March and we don’t know which test those number represents, DeepBench is not like 3DMark where you are given a single score, was the test done for inferencing or learning? at which matrix size? was pseudo FP16 or mixed precision used on GP100 or was done with FP32?
There’s really nothing
There’s really nothing convenient about that, they’re marketing it against the TITAN Xp and pricing is reflective of this. You’re comparing a professional (application certified) Quadro against the prosumer model from AMD which costs far less. Those products will come later from AMD. Whether or not they’re competitive remains to be seen.
Is the HBCC 48bit(256TB) or
Is the HBCC 48bit(256TB) or 49bit(512TB)?
Ive read seemingly conflicting things.
It’s 512TB of virtual address
It’s 512TB of virtual address space according to AMD, and I just wonder if there will be any benchmarks to test Vega’s HBC/HBCC subsystems because it will be great to have a reletively unlimited Texture/Geomentry space that can be swapped out of the HBM2/CACHE to regular DRAM, or even Virtural memory swap file on an SSD, so some Graphics Workloads with large and very detailed scenes/meshes can be utilized.
I’m sure that Vega’s HBCC will probably be able to handle having whatever the GPU’s HBM2 size is paired with maybe 3 times the HBM2’s allotment swapped to regular DRAM/SSD and allow any game to run from the HBM2/CACHE with no GPU performance degredation. That HBCC looks like it has the ability to manage any regular DRAM/SSD to HBM2/CACHE traffic in the background and keep the Vega NCUs feeding from HBM2 and the L2/l1 Cache levels above.
AMD’s driver engineers are probably doing 80 hour weeks for the next 5 weeks up to SIGGRAPH and RX Vega’s release. And I fully expect at least 3 more months of intense GPU driver updates owing to the new IP introduced with the Vega GPU micro-arch. All that HBC/HBCC technology, NCU tweaking, and primative shaders/other IP probably will see things on the driver side from AMD on a intense and constant update path for the remainder of 2017 into 2018.
1600 MHz is not the base
1600 MHz is not the base clock. It would have been a miracle if it was. 1600 MHz is the boost clock. Also, as Anandtech points out, 1382MHz is the *typical* clock, not base clock. AMD haven’t given a base clock.
4 years ago when the technical press was 200% Intel/Nvidia fans searching for reasons to bash AMD, there was a major attack against AMD’s R9 290 series cards for not keeping a stable 1GHz frequency. It will be interesting to see tech press’ reaction if the air cooled Frontier Edition throttles under that 1382MHz frequency. The tech press today treats AMD way much better.
Well, how do you think they
Well, how do you think they would have treated other companies if they released a product that wasn’t sustaining the advertised frequency because of the crappy cooler or for power consumption going out of control while gaming, causing framerate drop and inconsistency?
It was … interesting …
It was … interesting … trying to get the information for this post. The PR did not have a lot of numbers in it but you are right, that should be 'peak clock'. Fixing now.
Well this Vega SKUs appears
Well this Vega SKUs appears to be like that Radeon Pro Duo with some switchable professional drivers and gaming drivers for some non safety essential workloads and not really tuned in the hardware for gaming workloads.
I kind of cringe at the thought of this SKU being used for Professional workloads where the Radeon Pro “WX”(formally branded FirePro) SKUs or Nvidia Quadro SKUs are are really necessary.
You are not going to ever see any structural engineering firms using any SKUs but the Xeon or Epyc Workstation CPU SKUs alongside any AMD Radeon Pro WX or Nvidia Quadro GPU SKUs. And that includes any Xeon or Epyc Motherboard SKUs that are fully certified for error free(As much a possible) workloads.
That certification is what makes the AMD Radeon Pro WX and Nvidia Quadro SKUs cost so much and each and every professional software package has to be fully certified to work with any Radeon Pro WX and Quadro GPU SKUs at the cost of millions and many millions of extra programmer hours. Ditto for the motherboard UEFI/BIOS, and Professional WX/Quadro GPU BIOSs, on the certified motherboard/PCI card hardware, and all that extra QC/QA runs in the millions for software and hardware OFFICAL certification processes.
For some students this Vega Frontier Edition is low cost(Relative to the WX/Quadro SKUs) with at least the Pro drivers available for learning and programming. So I can see this used by colleges for training and development for software that targets the Professional drivers used in the production legal “WX” branded Real Professional GPUs from AMD. This Radeon Pro Frontier Edition will never be legally used for any Engineering Production workloads as that will legally require the WX version of the GPU(And Pro CPU platforms), but for students and software development(Non production workloads) it’s similar in reach to the Radeon Pro Duo that AMD released last year.
There is supposed to be a Vega Radeon Pro WX version of this card and it will very likely be demonstrated at SIGGRAPH 2017. It’s kind of funny that a Gaming card is being premired at SIGGRAPH(A professional Trade Show for the professional market) but AMD has been very busy lately and in need of some better better techical/writing staff. So hopefully EPYC sales will bring in the revenues for AMD to begin to Hire more Technical writing/engineering folks who can begin producing the better whitepaers and such that are beyond the pay/brain grade of any marketing monkyes.
“The memory size for the Vega
“The memory size for the Vega is interesting, HBC is AMDs High Bandwidth Cache Controller which not only uses the memory cache more effectively but is able to reach out to other high performance system memory for help. AMD states that the Radeon Vega Frontier Edition has the capability of expanding traditional GPU memory to 256TB…”
(should be a comma after “effectively” above, btw.) – & also, methinks the correct term for the above is HBCC, not HBC?
I cant believe there isnt more of a fuss about ssd ~= vram?
Its a pity it isn’t a striped array?
How much is the ssd option for vega?
Its always mentioned, but never any specifics?