AMD GPU Architectures pre-GCN Are Now Legacy

Subject: Graphics Cards | November 26, 2015 - 03:09 PM |
Tagged: amd, graphics drivers, GCN, terascale

The Graphics Core Next (GCN) architecture is now a minimum requirement for upcoming AMD graphics drivers. If your graphics card (or APU) uses the TeraScale family of microarchitectures, then your last expected WHQL driver is AMD Catalyst 15.7.1 for Windows 7, 8.x, and 10. You aren't entirely left out of Radeon Software Crimson Edition, however. The latest Crimson Edition Beta driver is compatible with TeraScale, but the upcoming certified one will not be.

View Full Size

GCN was introduced with the AMD Radeon HD 7000 series, although it was only used in the Radeon HD 7700 series GPUs and above. The language doesn't seem to rule out an emergency driver release, such as if Microsoft breaks something in a Windows 10 update that causes bluescreens and fire on older hardware, but they also don't say that they will either. NVIDIA made a similar decision to deprecate pre-Fermi architectures back in March of 2014, which applied to the release of GeForce 343 Drivers in September of that year. Extended support for NVIDIA's old cards end on April 1st, 2016.

I wonder why AMD chose a beta driver to stop with, though. If AMD intended to support TeraScale with Crimson, then why wouldn't they keep it supported until at the first WHQL-certified version? If they didn't intend to support TeraScale, then why go through the effort of supporting it with the beta driver? This implies that AMD reached a hurdle with TeraScale that they didn't want to overcome. That may not be the case, but it's the first thing that comes to my mind none-the-less. Probably the best way to tell is to see how people with Radeon HD 6000-series (or lower-end 7000/8000-series) cards work with Radeon Software Crimson Beta.

Likely the last drivers that users with Radeon HD 6000-series graphics need are 15.7.1 or Radeon Software Crimson Edition Beta. We will soon learn which of the two will be best long-term.

Or, of course, you can buy a newer GPU / APU when you get a chance.

Source: AMD

November 26, 2015 | 03:32 PM - Posted by Anonymous (not verified)

so does this mean an end to support of older AMD APU's?

November 26, 2015 | 04:13 PM - Posted by Scott Michaud

Yup. Llano, Trinity, and Richland are based on TeraScale and do not support upcoming AMD drivers. Kabini, Kaveri, and Carrizo will remain supported though, because they use GCN-based GPUs.

November 26, 2015 | 04:48 PM - Posted by Anonymous (not verified)

I run a 7950 with 5450 for additional DVI-I/VGA outputs (active adapters aren't a solution because they don't support the full resolution and refresh rate of the monitors). Up to now, the 5000 series and 7000 series were all supported by the same driver and everything worked as it should; I hope this change doesn't cause any problems, or I'm going to be stuck with current driver forever.

November 27, 2015 | 02:54 AM - Posted by JohnGR

You can sell the 5450 and buy a second hand cheap nvidia card if you end up having problems with that configuration. Other than the whole story of selling one card/buying another, you shouldn't have a problem and it shouldn't cost you any more money. I guess a second hand GT 610 sells for about the same as a second hand 5450. I have a HD 7850 and an GT 620 in my system for years. Just install the Nvidia driver first, AMD after.

Of course with many people having older APUs and GCN discrete graphics cards, I assume that both drivers will be able to co exist in the system or AMD is unprepared for this change.

November 26, 2015 | 04:50 PM - Posted by Kusanagi (not verified)

They had a very great overall run, though.

November 26, 2015 | 04:59 PM - Posted by Anonymous (not verified)

"GCN was introduced with the AMD Radeon HD 7000 series, although it was only used in the Radeon HD 7700 series GPUs and above."

Then GCN was not started at the 7000 series, it started with the 7700 series and above. What is the obsession with having GCN associated with increments of 1000, when GCN was introduced on an increment of 100, the 7700 hundred series and above. And I'm damn tired of series of GPU micro-architectures(VLIW4, VLIW5, GCN(1.0, 1.1, 1.2) being listed by arbitrary increments of numbers. Saying 7000 series is a misnomer, it's the 7700 series and above for GCN. AMD has had three GCN generations/revisions to date with the 1.0, 1.1, and 1.2 numbers arbitrarily associated with the generations/revisions.

What AMD needs is a master list with every GPU AMD/ATI has ever made, including its integrated graphics, in tabular form with each entry associated with the proper GPU micro-architectures, including the code names, and marketing names, and any other steppings, fixes and other modifications to the series associated with the exact GPU make/model name/number in question. None of this guess what was started on increments of 1000, and then guess the 100's where the new micro-architecture actually started being used!

Master naming disambiguation tables must be maintained with the proper code names, numerical names, etc. otherwise what is the use for numbers incremented by the 100s 1000s 10,000s, on and on! A series is not bound by increments of 1000s, or any number it's usually associated with a micro-architectural change according the scientific computing fields.

WTF is with the naming/numbering conventions that were established decades ago by the computing sciences disciplines, and promptly not adhered by the MBA/marketing department run “Technology” companies of today.

November 26, 2015 | 09:26 PM - Posted by Anonymous (not verified)

The numbers are just marketing names based on what is available at the time. There was some overlap when they were still selling both TerraScale and GCN based parts. They wouldn't have wanted to move the GCN based part up to 8000 series as this could have hurt sales of TerraScale based parts still sold as 7000 series. Bumping them all up to 8000 series parts doesn't change the situation. The numbers do give a relative performance indication for the available products, but it doesn't indicate the underlying architecture. I don't see this as that big of an issue. Most of the numbers listed with computer hardware are meaningless to the average buyer anyway, even when they refer to real quantities.

If you want a listing of marketing name and architecture, Wikipedia actually has a good table for most architectures that list the name the chip was sold as as well as the architecture and other specific information. The 7000 series chipset table on Wikipedia doesn't list the architecture though; it is indicated in a note on the top. See the wiki Radeon_HD_7000_Series#Chipset_table. With the Rx 200 series parts, the architecture is included in the table since there is a mixture of TerraScale2, GCN 1.0, and GCN 1.1. The remaining TerraScale2 parts are OEM only in the 200 series.

November 26, 2015 | 10:34 PM - Posted by Anonymous (not verified)

You are talking marketing jiber-jaber, I'm not talking about that. If GCN starts at the 7700 number then the 7000 series does not denote a series, the 7700 hundred number is the beginning of GCN series and anything below 7700 is TeraScale/other. This Big numbers means more performance mantra is for Joe six-pack, his mono-brow and road dragging knuckles. So I did not imply that the parts should have been started at 8000. All the "TerraScale, TerraScale1, TerraScale2" parts accordingly should be labeled next to the exact GPU name and model number, same for the GCN labeling that AMD uses GCN Gen 1, GCN GEN 2, and GCN Gen 3. AMD is the one that needs to make the tables, and not the press, or Wikipedia.

It's just fine for a new micro-architecture to start at the 7700(series), what is not fine is for the code names, and the engineering names, and the marketing names to not be properly documented and disambiguation tables for the code names/engineering names/marketing names not produced by the GPU maker, and published by the GPU maker. AMD's GPUs need to be research-able by the serial numbers on the chips/device, so the technical documentation and research can be properly done. That includes parts that are only sold to OEMs, that wind up in mostly laptops. There are many open source driver coders that need the proper documentation on the parts, including the OEM/laptop parts. Disambiguation tables are the responsibility of the device's manufacturer with every make and model number of GPU part having its own entry in the tables, none of this leaving it up to guess work, because marketing naming/terminology does not mesh up with engineering/terminology including code names.

November 27, 2015 | 06:14 PM - Posted by Anonymous (not verified)

Your missing the point. The "7000 series" type identifiers are marketing names, but they still identify the device for most consumer purposes. If more information is needed, a very specific device ID and vendor ID can be found on a sticker on the card or via the Windows system information tool. Since it is available through the windows device manager and system information tool, it is available via the PnP device ID which can be discovered programmatically. Just because you do not know how to find the information does not mean that it is not available. Although, AMD does appear to direct to a third party pcidatabase site to match device IDs with specific product strings like "AMD Radeon™ HD8530M". This isn't necessarily a useful thing to do for a driver developer unless you want to list the name to the user in some manner. This means that there isn't much reason to have a table to match device ID (the "one true identifier") with the model name.

The tables on Wikipedia are useful if you want to know more specific information, but it isn't really information that a general consumer requires; it doesn't give you any idea of relative performance. To the common consumer, most of those numbers are meaningless. Also, the information listed on Wikipedia is probably not useful to driver developers at all. You don't need to know the number of transistors on a GPU to write a driver. They would be going by the device ID. You seem to be trying to create a problem that does not exist. As far as I know, AMD is more open than Nvidia as far as open source driver development. The problem is, Nvidia's proprietary Linux drivers seem to be better than AMD's proprietary Linux drivers.

July 2, 2017 | 04:18 PM - Posted by Anonymous2 (not verified)

That's right! What is the point of naming/numbering conventions if they only confuse people?

November 26, 2015 | 07:35 PM - Posted by Anonymous (not verified)

Are we getting a new base architecture with the 14/16 nm parts?

November 26, 2015 | 11:23 PM - Posted by AS118 (not verified)

Yes, next year with HBM and GDDR5X memory too. It'll probably a whole new lineup top to bottom with GCN 2.0 or something, and new drivers.

Both AMD and Nvidia have said that it may be the latter part of the year when launches happen though. Probably close to the holidays.

November 26, 2015 | 09:44 PM - Posted by Anonymous (not verified)

According to this Article it's an all-new micro-architecture, and an All new instruction set architecture, but that "Mark Papermaster, chief technology officer of AMD, calls the next iteration of GCN" so maybe the new ISA will just be a superset of the previous GCN ("1.0, 1.1, 1.2") GCN ISA/Revisions redone to be more efficient in the Arctic Islands micro-architecture.

It's not unheard of to revise the underlying hardware to execute an ISA more efficiently and add ISA extensions from generation to generation, or even on simple revisions to the ISA on GPUs. So hopefully the new ISA on Arctic Islands is adding more asynchronous compute abilities that will allow for the ACE units to do their job more efficiently with some extra instructions added to do things more quickly and efficiently. It is always possible that some of the older GCN ISA instructions will be retired in favor of newer instructions, but that always happens in GPU micro-architectures that are constantly being revised to better perform. So if AMD completely reworked the underlying hardware to make GCN/whatever they decide to call it run better with a new more efficient ISA to improve performance then will they decide to call in something new, or just continue with GCN naming.

I don't think that the move to Arctic Islands is going to be as drastic of a change as AMD's moving from VLIW to GCN, but maybe there will be more VR types of Instructions added to GCN, and an overall reworking of the underlying hardware to run the latest GCN/Arctic Islands ISA faster and more efficiently. Maybe the ratios of SPs to tessellation to ROP/etc. will change and some better cache algorithms/larger caches, and wider internal buses in the ACE units/other areas will be added.

If the Greenland GPU tape-outs have already happened then the benchmarks are not to far off from leaking.

http://www.kitguru.net/components/graphic-cards/anton-shilov/amd-readies...

November 27, 2015 | 06:38 PM - Posted by Anonymous (not verified)

One of the advantages of GPUs over CPUs with a fixed SIMD ISA is that you do not need to use the ISA outside of the driver. For the GPU, the shader code gets compiled specifically for each different GPU. Even if they do make drastic changes, the shader code would just get recompiled for the new architecture when it is executed. I doubt it will be drastic changes though. AMD's GCN architecture seems to have been designed to be very forward looking with how well even the older GCN based GPUs handle DX12 and asynchronous compute.

I think that the asynchronous compute resources will be very important for reducing latency for VR. HSA may be important also. It is a lot lower latency to just pass a pointer to the GPU rather than calling on an entire subsystem to copy the memory from system memory to GPU memory over the PCIe bus. You need unified memory, like what would be offered by an APU with HBM or a GDDR5 derivative, to take full advantage of this though. It is going to be 2017 or 2018 before we get a high-end APU though. A lot of consumers will think of it as integrated graphics at first, but a high-end APU with HBM should be able to outperform a solution with graphics connected via PCIe. We may have to wait for AMD's HPC APU. They could do something similar to what Intel does with Xeons. They sell salvaged Xeons as Extreme Edition parts.

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.