Techgage has posted a look at what AMD's new HBCC feature in Vega is and how it will help you run games faster. HBCC allows your GPU to treat VRAM as a last-level cache, so that a request for data not currently located in VRAM can be pulled into Vega's HBC for immediate access while simultaneously flushing out data which is no longer needed. In addition to describing how the feature functions they also did quite a bit of testing to determine the real world effect of enabling HBCC in games and benchmarks. Drop by for a look.
"AMD’s Vega GPU architecture brings many notable features to the table, but the one to find its way into Radeon chief Raja Koduri’s heart is HBCC – or “high-bandwidth cache controller”. In this article, we’re going to take a look at what HBCC is, why it offers no benefit right this moment, and talk about what it could offer in the future."
Here are some more Graphics Card articles from around the web:
- AMD RX VEGA @ Kitguru
- AMD Radeon RX Vega 64 & 56 Best Playable Settings At 4K & Ultrawide @ Techgage
- Asus ROG STRIX RX Vega64 O8G Gaming @ Kitguru
- How The Radeon OpenGL Performance Has Evolved From The HD 2900XT To RX Vega @ Phoronix
- AMD Radeon RX Vega56 8GB @ Kitguru
Useless marketing
Useless marketing gimmick.
Everything about radeon is inferior.
Well the RX Vega 64 beat out
Well the RX Vega 64 beat out the Quadro P6000 on some interesting compute benchmarks over at that same techgage website, and I guess he does not have a RX Vega FE to test on any compute workloads.
Also there are quite a few folks excited at SIGGRAPH with that Vega HBCC/HBC ability, and the Vega micro-arch will do fine with the compute/AI and professional graphics markets. It’s no wonder that AMD did not live-stream Vega’s launch from SIGGRAPH because, if I where a professional, I would not want a professional Trade Show having any non technical focused events.
It’s the gamers’ fault for not being able to separate out the Professional Compute/AU marketing around Vega from the gaming marketing around Vega, and that compute/AU market can afford to pay some fine markups on the Professional Vega Variants.
The miners will pay top dollar for Vega in the consumer markets and that HBCC/HBC IP will come in handy if the compute requirements for coin mining surpass the 8GB memory limit for any non Vega SKUs. The Blender artists are some more folks that are very interested in Vega’s HBCC/HBC IP, as that means their 3d animation scenes can have huge amounts of high polygon count mesh models and not crash the GPU when the physical VRAM memory runs out! And is all because that Vega HBCC/HBC can use regular system memory/SSD/Hard-Drive space as Virtual VRAM up to 512TB in size.
The Vega cards are selling out and the Vega revenues are coming in for AMD/RTG. So AMD’s stock holders are happy with that news. Thanks Raja!
JHH over at Nvidia, on seeing how much Vega’s are selling for even over the GTX 1080Ti pricing, may just have a little extra “Demand” driven price increases on the Green Team’s pricing because JHH is all about the green and will demand that Nvidia’s consumer GPU prices be raised. And It’s not like JHH needs to lower his pricing to gain market share, as Nvidia already has the majority of the consumer GPU market share already. So look for Nvidia’s pricing to go higher until the Vega pricing starts to drop and that may take a while.
JHH is secretly very happy the miners are buying AMD’s consumer cards and raising average GPU pricing across the entire consumer GPU market because that gives him an excuse to raise Nvidia’s consumer GPU prices also! And the Green Teams Fanboys will gladly pay more and brag about paying more with a cherry on top right into JHH’s rather rotund bank account!
Vega is probably superior in
Vega is probably superior in several ways, just none of them are being taken advantage of by current games.
I am wondering how this ties
I am wondering how this ties into HSA. An APU with some HBM backed by a high performance SSD or NVDIMMs could be interesting for mobile, but an APU would already have virtual memory from the CPU component. Would GPU HBCC be active? Virtualized memory can make much better use of the available memory, so even a 4 GB HBM stack could be quite effective. I don’t know if they can make use of it to avoid memory copy for non-HSA aware applications.
The real benefit seems to mostly be for developers. GPU memory still seems to need a lot of micro-management. With something like an open worlds game and fully transparent hardware virtualized memory, you could technically just load everything required into memory as you go, and not have to worry about space. That which hasn’t been used recently will just get swapped out. If the player then goes back to a previously visited area, it will get swapped back in automatically. Virtual memory systems provide a huge benefit for programming on CPUs. Memory management is simplified significantly for the programmer. It will be great if this can provide similar features for GPUs, then developers should be interested. With an 8 GB card and developers still doing the micro-management, I wouldn’t expect much difference now. It will probably be used more for HPC in the near term.
I also have to wonder how much the Vega system interface is actually shared with Ryzen. There was a rumor (I think) of an APU with a Vega based graphics chip with HBM mounted on an MCM attached to 2 zeppelin die with 4 links. That makes sense since two die on an Epyc processor use 4 links to connect to the other two die. It would make an exceptional HPC device. I have to wonder if the Vega infinity fabric interface is actually nearly an exact copy of Ryzen’s interface with support for cache coherent memory access, the same as a zeppelin die would have. It would make sense if the virtual memory management system is essentially the same design. Running it over pci-express is probably not using much of its full potential. I don’t known if it makes sense for it to have the the full 8 x16 links that a zeppelin die has; perhaps it just has 4 though. Vega probably isn’t designed to connect multiple GPUs together. That will probably come with Navi; I expect it will be multi-die solution. It is unclear whether it will be multi-die on an MCM or multiple small interposers with GPU and HBM packaged onto an MCM, or just multiple cards. Pci-express 4.0 signaling will supply a lot of bandwidth though, and the high clock rate will reduce die-to-die latency, so just using multiple cards may still work out quite well.
Yes APUs already have VM
Yes APUs already have VM functionality that both the CPU/GPU cores share, and discrete GPUs can have that same ability as GPUs are just processors really with their own memory controllers on those PCIe card variants. Really some maker could build a whole computing platform on a PCIe card if they wanted to, and that may be coming sooner that you think. But AMD’s First Raven Ridge APUs may not even come with any HBM2, or AMD may have provided for in its first Raven Ridge mobile SKUs some eDRAM to take the place of HBM2 in that VRAM/VRAM used as HBC by that HBCC in Vega’s GPU micro-arch, and still be able to have a Vega APU variant with some HBC(eDRAM instead of HBM2) to make use of.
Remember also that Vega GPU’s make use of the Infinity Fabric IP so that APU will speak Infinity Fabric protocal between the CPU cores and the GPU cores on that Raven Ridge APU. This Infinity Fabric IP from AMD also has great implications for any Dual Vega/dies on one PCIe card variants communicating over that Infinity Fabric fully cache coherent protocal and those 2 Vega dies on the same PCIe card will be seen as one logical GPU to the graphics APIs and gaming software, no need for any that inefficent CF with its not so good scaling issues.
If the first Zen/Vega Raven Ridge APUs have no HBC(HBM2, eDARM/other) to make use of then the APU will have to work from DIMM Based DRAM only and AMD will have the same issues again with laptop OEMs only providing a single channel to DIMM based DDR4 DRAM. So AMD really needs to have some form of large last level VRAM/Cache on its APU SKUs or the Vega graphics will once again be starved of memory bandwidth.
AMD does have in development some workstation grade interposer based APU variants(Zen Dies, and larger Vega Die, and HBM2, all on an interposer) but that’s in the 2018+ time frame so that will be interesting to see once that class of workstation garde APU SKUs arrive.
Using system memory to
Using system memory to compliment VRAM would be of very limited use, mainly because VRAM to GPU is around 400GB/s whereas system memory to VRAM is around 15GB/s.
Streaming works well since
Streaming works well since the original Playstation.
But what they do is make the
But what they do is make the handling the different memory regions automatic just like your cpu.
The data is always first entering normal DRAM then its copied to graphics memory.
What AMD done is making it automatic you no longer need to handle the tracking it is done in hardware just like it’s done for the CPU. So this is making it much easier to work with.
But yes if your working set for EVERY frame is larger than available graphics memory then its slow but that is true regardless. This is a great feature that makes handling big data much easier its not something that will make peak performance better.
Yes, but the HBCC/Memory
Yes, but the HBCC/Memory controller subsystems on Vega are managing that in the background and hiding any latency/bandwith issues with the transfers over PCIe from system memory into discrete GPU’s VRAM/cache. So with HBM2 being treated like a last level cache the Vega nCUs/shader cores will feed from L2 cache and indirectly from HBM2/last level cache mostly at no loss of bandwith or average latency.
That is the very definition of cache memory and HBC(HBM2) is where the GPU will work from the majority of the time and from its HBM2(HBC/Last level cache) on up into the Vega GPU’s higher cache levels like L2 cache that the HBCC is a direct client of under Vega on up into the shader cores’ L1 I$ and L1 D$ and ranks of registers.
HBM2 under Vega in fact becomes L4 cache to any larger amounts of slower DIMM based DRAM, or Paged memory on SSD/Hard-drive. AND that main function of cache is to hide latency because that’s why cache memeory was invented for in the first place on up the levels of cache into the instruction piplines on the shader cores, ditto for data cache!
Vega’s 8GB of HBM2/VRAM/Cache can leverage many times its 8GB of VRAM/Cache size and Vega’s HBCC can manage and hide any latency involved with using virtual VRAM on the DIMMs/DRAM or disk/SSD in the background with Vega’s shader cores(L1 I$, L1 D$, register files) working mostly from L2 cache<-HBM2 last level cache<-system memory<- SSD/hard-drive. Raja/His Teams of engineers engineered it that way and the IP is definitely the most interesting part of the new Vega GPU Micro-Arch with up to 512TB of directly adressable virtual VRAM(system memory-DIMM-DRAM/SSD/Hard-drive).
Yeah but who is going to
Yeah but who is going to program a 512tb game. Most don’t even reach 50gbs size these days. Until storage capacties ramp up, download speeds massively improve and cost is a lot less per terrabyte, games aren’t going to be much bigger. Face facts hbcc is not useful in gaming scenarios now.
That’s 512TB of directly
That’s 512TB of directly adressable memory space and no one expects any games or professional(for that matter) applications to need to use that much actual memory/paged memory. That size is just there for some future expansion latitude, and you never know what will be needed in the future. I’ll bet you probably thought greater than 512K of memory/memory adressing space would never be needed also.
And do read my reply to your other post those 2 articles I refrenced are very interesting reads!
You will now be known as Dafty Duck for your continued Daft replies, so keep trying! It’s so fun to show you the truth and see how you can even deny the very existence of that nose on your face!
This will not see any real
This will not see any real benefit for games, as using VRAM as a cache has been the standard way of utilising memory for the last decade.
When a level loads, games will load every texture (+geo, etc) possible into VRAM until either they run out of textures for that level, or run out of VRAM. All that cached data is flagged as low-priority, so it will get immediately overwritted if VRAM is full and space is needed for e.g. a buffer.
Basically, AMD have stuck their own label on something that everyone has already been doing for years.
This will not see any real
that is not even remotely a correct statement.
If you are managing that VRAM
If you are managing that VRAM to system DRAM transfer with the gaming software then yes that’s been around for a while(Allocating Buffers) but Vega has hardware methods to treat HBM2 as a VRAM/Cache like its a level 4 cache in Vega’s hardware with programmers letting Vega’s HBCC manage the paging via it’s hardware/firmware caching algorithms and with the programmers more freed up from having to manage the page movement using software methods that very often are inefficient methods done by some programmers than can be detrimental to gaming performance.
With Vega/HBCC-HBC(HBM2), programmers can ask for memory to be allocated for whatever texture amount that the games needs and the HBCC will track and keep what is most used/immediately needed paged into the HBM2/HBC and let the games programmers worry about the games’ logic and not the the paging of Virtual VRAM logic. And that’s the same as any VM paging done on CPUs with the programmer only asking for memory and not worrying about managing caches or paging tables, etc.
Now there will be ways for the programmer to prioritize paging but that’s not new either, what is new with Vega is that sort of complicated page management to and from system DRAM/SSD/Hard-drive will be more in the GPU’s hardware managed by the HBCC/HBC memory-controller and associated subsystems. And that’s why GPU programmers/games programmers are all excited about Vega’s HBCC/HBC-HBM2 IP. No more programmer/gaming engine and shader programming that has to be always on guard for running out of VRAM and crashing the games. There will still be programmer control over the process via some added page priority functionality in the GPU opcodes, or APIs, concerned with memory movement control added to the Graphics APIs as Vega hardware specific extensions to the DX12/Vulkan APIs and hopefully that Vega Type of functionality will be used across the entire industry and all GPUs will have some similar functionality.
Seems like Jeremy was
Seems like Jeremy was listening to John Mellencamp when he wrote the title to this blurb.
Could be, could be 😉
Could be, could be 😉
Much Ado about nothing
Much Ado about nothing really. Video cards have had the ability to use system ram as video ram for a while.
https://www.techpowerup.com/forums/threads/is-it-possible-to-share-graphic-card-memory-with-ram.162038/
It’s much slower than accessing video ram obviously.
Simulating a 2 gig card and touting it’s performance when hbcc in use. Makes for an impressive 50% boost in frame rates and 100% for minimums. Marketing fluff at it’s finest.
LOL this feature is found on: 2 gig card nope ;a 4 gig card nope ;a 8 gig card yea. 8 gigs will be sufficient long after this card is obsolete. Unless we get some ultra cheap super fast high capacity storage medium for computers, I don’t see 8 gigs being outdated soon.
I have a 4gig 760 ftw and I’m currently playing games like Overwatch,Skyrim,and many others at 4k. The frame rate is often comparable no matter what resolution I run at because I’m GPU bound. If I had 2 gig version fat chance of it being remotely playable. The 4 gig was deemed a gimmick by Game Debate because card was too weak. I can confirm it uses the full 4 gig in games. All this from a 4 year old “gimped” Kepler card.
“Much Ado about nothing
“Much Ado about nothing really. Video cards have had the ability to use system ram as video ram for a while.”
And that 2012 TPU forum is a little out of date but the software is doing that buffer transfer work and not the hardware as is done with Vega’s HBCC/HBC-HBM2. And the bandwidth on that card’s VRAM is around 192.3 GB/s while Vega 64’s memory bendwidth is 483.8 GB/s on 2 stacks of HBM2. So that’s going to be much higher realtive to any dual or even quad channel DDR4 system memory’s effecive bandwith relative to Vega’s HBC/HBM2-cache’s bandwitth.
And it’s not the same as Vega’s HBCC/HBC hardware using the HBM2 as a last level cache and Vega able to run mostly from the HBM2 VRAM/cache! Now say for a discrete Vega Mobile with 2GB/4GB of HBM2 with the HBCC doing the transfers from System DRAM in the background over to the GPU’s HBM2/VRAM-cache and you, to get an idea, will have to look at that test that AMD performed by disabling half of the HBM2 in the Vega demo and the game not suffering any adverse effects. And you are not a programmer and you really do not uderstand how caches are supposed to work.
And really the games are going to query the GPU hardware and adjust accordingly! So some gaming benchmarks are going to have be developed where the games are modified to throw 8 gigs of assets at a 4GB GPU because the games as you tested on your old hardware where designed to adjust assets based on the GPU SKUs ability to deal with the game using that specific card’s hardware that the game detected.
But you will continue to try and spin negative without doing any objective testing at all, and the linked to article below is more of an indication of what your real intents are, objetivity be damned. Really more daft posts and no objsctive testing done with Vega 64, and no gaming benchmarks designed to stress test Vega’s HBCC/HBC-HBM2 or BIOS settings provided by AMD so half the HBM2 capacity can be disabled and still you look at 5+ tear old technology to make your “Informed Decision”!
If you want to see that HBM2/HBCC on Vega in action look at that techgage article on Vega 64’s compute and its better performance relative to the Quadro p6000(24GB) SKU and there you will see what Vega and its 8GB of VRAM and greater compute can really do, realtive to a $5000 Nvidia pro SKU. But for gaming the benchmarks will have to be created to stress test Vega’s HBCC/HBC IP but that’s not available in the here and now outside of AMD’s testing labs and that will be fixed shortly so the proper tasting can be done on Vega.
But Your post is so disingenuous and I can’t see the forest for the FUD lately! I’m only looking at your constant spin negative with objsctivity thrown out the window mostly to advance your illogical cause!
[Meta]62% of Vega reviews on Amazon rate it 1-star. Not a single Vega review on Amazon is from a verified purchase.
https://www.reddit.com/r/Amd/comments/6veus7/62_of_vega_reviews_on_amazon_rate_it_1star_not_a/
You are a compute groupie. I
You are a compute groupie. I could care less about that I’m concerned with gaming performance. HBCC doesn’t do much in fact it actually may lower performance right now with it on. Any gains or losses are within margin of error so how useful is it now out the door at launch. Yes AMD might address this in the future or it might become more unsupported tech, who knows.
Doesn’t matter if it’s hardware accessing system ram to use as video ram or not. System ram is slower than Vram period. Integrated graphics which use system ram as videoram are much slower. It is much better to use built in Vram on video card than using outside sources for it’s cache.
A 2tb nvme ssd ($1000) isn’t going to be included with a budget 2gig video card ever. This is for Vega’s more costly professional market only.
So making comparison with 2 gig card is disingenous at best. 3/4 gig is new standard amount of video ram for a gaming laptop these days as well.
The 62% is where it should really be using up to 2x the wattage to reach basically the same performance as a 1.5 year old Nvidia 1080. Reviewer got their golden samples and have to be extremely kind to AMD in their review if they want to receive them in the future.
AMD is having quality control issues as not every Vega is produced the same. Sad really late to the party and still can’t get it right.
Do you need me to do your work and cite sources for this common info a 6 year old and search engine of choice can find?
If you want me to continue to trash Vega I can’t do a worse job than AMD already has rushing it out the door and not supporting features touted to be on the card. False advertising IMO.
AMD actually disabled gaming drivers for VegaFE in latest drivers so one could not compare gaming performance on latest driver to Vega 64. They said gaming would be better on Vega64 but that can’t be proved either way. Shady.
You don’t need new gaming benchmarks just crank the details up to use more Vram in a game and bench it’s not rocket science. Oh maybe 8gigs will be plenty for the near future and is overkill for now which makes hbcc kinda useless for gaming now.
If Nvidia released Vega64 in it’s current state, let’s just say people wouldn’t be nearly so kind towards them as they are to AMD.
Cheer up AMD has Ryzen and Threadripper and Epyc for a few more months at least until Intel gets serious and responds in kind.
Ha! Epyc will get AMD back to
Ha! Epyc will get AMD back to its Opteron levels of Server market share, and higher, and Vega will be right there for any Epyc/Vega compute for the HPC markets without the need for the Zen micro-arch needing any power hungry AVX 512 units. Epyc does just fine for the majority of server workloads that do not need any heavy AVX units crowding the cores and making them drink the power. Vega will be able to provide that FP compute for any Epyc based HPC solutions. The compute market can afford to pay more than any consumre market and Nvidia is with that GPU compute market also.
There really in no large amount of revenue growth potential in the consumer gaming markets so it’s Just going to be AMD and Nvidia fighting it out for that stagnent consumer/GPU hardware markets share of business. The only part of the Gaming Market that is even growing revenue wise is in the games themselves and the micro-transactions and subscription services around the gameing markets and that does not help the GPU hardware sales as much.
The consumer market percentage of revenues for both Nvidia and AMD will continue to shrink relative to Nvidia’s and AMD’s higher revenue producing professional compute/AI markets and the other non gaming compute markets where the serious revenue growth potential is.
AMD has just closed some big deals in the far east markets with all 3(one new Epyc partner just announced) of those companies that have market caps in the $70 billion to $400+ billion dollar market cap range, so that’s some big companies all looking to use Epyc/Vega based systems for their compute needs. AMD will no longer be dependent on any fickle gamers for AMD’s very survival!
AMD once had a $92(May 25 2000) stock value and share of the server market and that was on CPU sales alone, so AMD has no worries if Gamers do not rush to Vega(Currenty Out of Stock after introduction).
AMD’s will still be tweaking Vega 64/FE and Vega 56 is scheduled to arrive with a little less compute and still have remaining some TPU/ROP resources that are very close the the GTX 1080, so we will se how Vega 56 and some overclocking works out, maybe even some undervolting but still high enough clocks push the GTX 1080 boundry in performance. AND the DX12/Vulkan gaming title conversion away from DX11 continues to accelerate as the entire gaming market moves on.
So enjoy your antiquated hardware anecdotes, but no one will take you seriously!
And this article from servethehome:
”
.
.
.
“Here is the AMD Vega10 graphics overview. Suffice to say, this is a significant step forward from Fiji.
We wanted to focus on SR-IOV virtualization features of the AMD Vega10 platform. AMD exposes the GPU as an SR-IOV device to up to 16 VMs.
The key here is that AMD’s engine works to partition off compute and memory resources to VMs. Important to note here is that AMD does not have an expensive software licensing model like NVIDIA GRID has. This market has seen several releases ahead of VMworld 2017 with the NVIDIA Tesla P6 recently released.” (1) [See Hot Chips slides from article]
And this article from EnterpriseTech:
“AMD (NASDAQ: AMD) announced deployments of its EPYC server processor during an event in Beijing, including new customer JD.com. The Chinese e-commerce giant along with Internet search giant Baidu (NASDAQ: BIDU) said they would deploy the AMD server processor during the second half of this year while media and web services provider Tencent (HKG: 0700) rolls out the AMD platforms in its datacenters by year’s end.”(2)
(1)
“AMD Vega10 Update from Hot Chips 2017”
https://www.servethehome.com/amd-vega10-update-from-hot-chips-2017/
(2)
“AMD Lines Up New China Datacenter Partners ”
https://www.enterprisetech.com/2017/08/23/amd-lines-new-china-datacenter-partners/
Its pleasing to see the usual
Its pleasing to see the usual drones expertly rebuffed – so… WHAT THEY SAID…, but I would add, these arguments are not extended to hbc managed virtualized nvme raid arrays.
They should be. If hbcc can minimise latency for relatively slower system ram, so it can for near memory speeds for seqential nvme arrays.
Yes its lower on the latency/speed heirachy of caches to DDR4, (doh), but its eons from what we historically regard as virtualised storage speeds – and it’s too all intents, limitless (512TB).
Its also ~affordable.
~$150usd per samsung 256GB evo nvme. Expect ~3200MB/s seq read & 2100MB/s seq write.
~$50 for an asus 4x nvme adapter (16 lane).
speed multiples on native threadripper/epyc nvme raid 0 are ~seamless, so raw speeds of ~12GB/s can be expected from the above 1TB array
Its not a good substitute allways. Thats what hbcc is for – to intelligently prefetch into faster cache.
A more mainstream context e.g. is the rumored 28CU raven ridge apu w/ 2GB of HBM2 (and why wouldn’t they for a 95w+ am4 power envelope?).
The cost of an ssd is a given in a new pc.
For less money than a single 500GB ssd, they can boot on a striped pair of 256GB drives, and double speeds to ~6GB/s
So even a very frugal 8GB APU entry level pc, could punch far above its weight.
Thats a low end example, but everywhere i see the cost of ram being a deal breaker for folks who could use an upgrade. Ram shortages are here for all of 2018 and beyond, and only amd & hbcc seem to offer hope of making do with less.