AMD Radeon R300 series

AMD Fiji with the four high bandwidth memory chips

The Radeon R300-series is a series of desktop - graphics chip company AMD and successor to the Radeon R200-series . Their code name is "Pirate Islands".

description

The Radeon R300 series is a classic refresh series. The first models were released on May 5, 2015 and were only intended for the OEM market . The first models for the retail market followed on June 18, 2015. These were classic rebrandings, i.e. new editions of old models from the previous series under a new name, sometimes with marginal changes. So it was z. For example, the Radeon R9 380 is a Radeon R9 285 with slightly increased clock rates. As a rule, such rebrandings are viewed critically in the trade press, with the R300 series making it even more difficult that the competing series from Nvidia , the Geforce 900 series , appeared 9 months earlier and was based on a new architecture. The use of the Pitcairn GPU, now called Trinidad, which was already used in the Radeon HD 7800 , was criticized in particular . On the other hand, the two Radeon R9 390 (X) models were rated positively because they are equipped with 8 GB of video memory as standard and have a very good price-performance ratio. Only in terms of power consumption they were clearly inferior to their Nvidia rivals due to architecture (the R9 390X even set a new negative record with 303 watts).

The Fiji graphics processor is the only new development in the R300 series. From a technical point of view, it is a double Tonga GPU (called Antigua from the R300 series onwards), consisting of 4096 ALUs, 256 texture units and 64 ROPs while maintaining the GCN architecture in expansion stage 1.2. In addition to the UVD 6.0 video processor , there were significant changes to the memory interface: Instead of GDDR5 memory, the new HBM memory is being installed for the first time , which is why Fiji has a 4096 bit DDR HBM interface. AMD states that with HBM memory higher transfer rates are possible, energy consumption can be reduced and space can be saved (both on the PCB and on the GPU itself). The disadvantage, on the other hand, is that when using the HBM1 memory, the maximum memory expansion is limited to 4 GB (only with HBM2 memory are up to 32 GB VRAM possible). Since AMD, like its competitor Nvidia, has omitted the 20 nm manufacturing process at TSMC , Fiji continues to be manufactured using the 28 nm process. The chip consists of 8.9 billion transistors and is 596 mm² in size. While Nvidia has already had numerous graphics processors of this size produced, this is a novelty for AMD. Until now, the R600 and the Hawaii GPU were the largest graphics processors from AMD with 420 mm² and 438 mm², respectively.

With the Radeon R9 Fury X , which AMD presented on June 24, 2015, the Fiji graphics processor was used for the first time and is used in its full configuration. The card has some peculiarities, as can already be seen from the naming, which partly differs from the rest of the naming scheme of the R300 series. In the reference design, the Fury X received water cooling, which is normally only implemented by the board partners using their own design (such in-house designs were forbidden by AMD on the Fury X). AMD set the list price for the Fury X at $ 649, which is identical to Nvidia's Geforce GTX 980 Ti . However, the Nvidia model has around 10% higher performance under Full HD, and around 5% under WQHD. Equal performance can only be achieved under 4K (Ultra HD resolution). Since a "battle for the top performance" with the Geforce GTX Titan X was expected in advance due to the leaked hardware data and the naming , but not even the equally expensive Geforce GTX 980 Ti could be beaten, the rating in the trade press was accordingly mixed out. This effect was reinforced by the better energy efficiency of the Nvidia models, as well as a bug in the first batch of water cooling (which caused a pump whine) of the Fury X. The memory configuration of only 4 GB was seen as a problem in advance, but the test reports only showed a limitation for high resolutions in exceptional cases (the 4 GB turned out to be a marketing disadvantage, since the Geforce GTX 980 Ti had 6 GB).

With the Radeon R9 Fury , AMD presented the second model based on the Fiji graphics processor on July 10, 2015. In contrast to the Fury X, this is now partially deactivated, i.e. H. Only 56 of the 64 shader clusters are now active, leaving the Fury with 3584 shader and 224 texture units. This configuration is also known internally at AMD under the name Fiji Pro. With marginally lower clock rates compared to the Fury X, the Fury achieves around 4% better performance in Full HD compared to Nvidia's Geforce GTX 980 . With larger resolutions, the lead gradually increased (although the card's suitability for Ultra HD resolution was disputed). Since AMD has now dispensed with a reference design, the problems with water cooling as with the Fury X did not arise (most of the board partners resorted to their own designs with air cooling). In combination with the significantly better price-performance ratio, the Fury was also rated significantly more positively in the trade press than the Fury X a few days earlier. Nevertheless, the list price of US $ 549 for the Radeon R9 Fury was seen as too high, as Nvidia had previously lowered the list price for the almost equally fast Geforce GTX 980 to US $ 499.

AMD R9 Nano

The Radeon R9 Nano , which AMD presented on September 10, 2015, is a special version of the Fury X in mini-ITX format. In order to comply with the necessary specifications, the Nano is only 15.3 cm long and specially designed for the so-called "SweetSpot" of the built-in Fiji graphics processor (ie the point in the clock spectrum where the highest efficiency is achieved from graphics card performance to power consumption) . For this, AMD uses the Fiji GPU again in full configuration, but with the clock rates of the Fury (1,000 MHz) and not the Fury X (1,050 MHz). However, hardware tests have shown that AMD's official clock rates are never achieved in practice; the actual average clock rate under 3D load was around 875 MHz. The card achieves roughly the same performance as the Radeon R9 Fury or the Geforce GTX 980. Compared to these, the Nano is actually not competitive with a list price of 649 US- $ (Radeon R9 Fury: 549 US- $, Geforce GTX 980 : 499 US- $), but both models are not available in mini-ITX format. The fastest models available there to date were the Geforce GTX 970 and the Radeon R9 390X, but both were significantly slower than the Nano, which is why it was the fastest graphics card without competition in the Mini-ITX market segment.

On April 26, 2016, AMD released the Radeon Pro Duo . This is a dual graphics card that uses two Fiji GPUs in full and replaced the Radeon R9 295X2 . From a technical point of view, the Radeon Pro Duo represented two nanos on a PCB including water cooling. Since Nvidia dispensed with such a dual card in the Geforce 900 series , the Radeon Pro Duo assumed a monopoly position as the fastest card on the market. AMD again installed a water cooling system from Cooler Master on the Radeon Pro Duo, which only came on the market in a reference design, and set the list price at US $ 1,499.

Data overview

Graphics processors

Graphics chip	archi- tecture	production			units						L2 cache (in kB )	API support					True audio	Video pro- cessor	Bus interface stelle
		production process	transis- interfere	The surface	ROPs	Unified shaders			Texture units			DirectX	OpenGL	OpenCL	Mantle	volcano
						ALUs	Shader units	Shader cluster	TAUs	TMUs
Oland	GCN 1	28 nm	1.04 billion	090 mm²	08th	0384	024 × Vec16- SIMD	06th	024	024	k. A.	11.1	4.5+	1.2+	Yes	1.0	No	UVD 3.1	PCIe 3.0
Tobago (Bonaire)	GCN 2		2.08 billion	160 mm²	16	0896	056 × Vec16-SIMD	14th	056	056		12.0		2.0+		1.1	Yes	UVD 4.2
Trinidad (Pitcairn)	GCN 1		2.80 billion	212 mm²	32	1280	080 × Vec16-SIMD	20th	080	080	0512	11.1		1.2+		1.0	No	UVD 3.1
Antigua (Tonga)	GCN 3		5.00 billion	359 mm²	32	2048	128 × Vec16-SIMD	32	128	128	0768	12.0		2.0+		1.1	Yes	UVD 5.0
Grenada (Hawaii)	GCN 2		6.20 billion	438 mm²	64	2816	176 × Vec16-SIMD	44	176	176	1024	12.0					Yes	UVD 4.2
Fiji	GCN 3		8.90 billion	596 mm²	64	4096	256 × Vec16-SIMD	64	256	256	2048	12.0					Yes	UVD 6.0

Model data

model	Official launch	Graphics processor (GPU)							Graphics memory			Performance data
		Type	Active units				Clock in MHz		size	Clock in MHz	Storage interface	Computing power (in GFlops )		Fill rate		Memory band- width in GB / s
			ROPs	Shader - cluster	ALUs	Texture units								Pixels in GP / s	Texel in GT / s
							default	Boost				SP  (MAD)	DP  (FMA)
Radeon R5 330 (OEM)	May 5, 2015	Oland	08th	05	0320	020th	0830	0855	2 GB DDR3	0900	0128 bit	00531.2	0033.2	006.64	016.6	0028.8
Radeon R5 340 (OEM)	May 5, 2015	Oland	08th	06th	0384	024	0800	0825	2 GB DDR3	0900	0128 bit	00614.4	0038.4	006.4	019.2	0028.8
									2 GB GDDR5	1125						0072
Radeon R7 340 (OEM)	May 5, 2015	Oland	08th	06th	0384	024	0730	0780	2 ... 4 GB DDR3	0900	0128 bit	00560.6	0035	005.8	017.5	0028.8
									1 ... 2 GB GDDR5	1125						0072
Radeon R7 350 (OEM)	May 5, 2015	Oland	08th	06th	0384	024	1000	1050	2 GB DDR3	0900	0128 bit	00768	0048	008th	024	0028.8
									1 GB GDDR5	1125						0072
Radeon R7 360	Jun 18, 2015	Tobago	16	12	0768	048	1000	1050	2 GB GDDR5	1625	0128 bit	001536	0096	016	048	0104
Radeon R7 370	Jun 18, 2015	Trinidad	32	16	1024	064	0950	0975	2 GB GDDR5	1400	0256 bit	01945.6	0121.6	030.4	060.8	0179.2
									4 GB GDDR5
Radeon R9 360 (OEM)	May 5, 2015	Tobago	16	12	0768	048	1000	1050	2 GB GDDR5	1625	0128 bit	01536	0096	016	048	0104
Radeon R9 370 (OEM)	May 5, 2015	Trinidad	32	16	1024	064	0925	0975	2 GB GDDR5	1400	0256 bit	01894	0118.4	029.6	059.2	0179.2
									4 GB GDDR5
Radeon R9 380 (OEM)	May 5, 2015	Antigua	32	28	1792	112	k. A.	0918	2 GB GDDR5	1375	0256 bit	03290	0206	029.8	102.8	0176
Radeon R9 380	Jun 18, 2015	Antigua	32	28	1792	112		0970	2 GB GDDR5	1425	0256 bit	03476.5	0217.3	031	108.6	0182
									4 GB GDDR5
Radeon R9 380X	Nov 19, 2015	Antigua	32	32	2048	128		0970	4 GB GDDR5	1425	0256 bit	03973.1	0248.3	031	124.2	0182
Radeon R9 390	Jun 18, 2015	Grenada	64	40	2560	160		1000	8 GB GDDR5	1500	0512 bits	05120	0640	064	160	0384
Radeon R9 390X	Jun 18, 2015	Grenada	64	44	2816	176		1050	8 GB GDDR5	1500	0512 bits	05913.6	0739.2	067.2	184.8	0384
Radeon R9 Fury	10 Jul 2015	Fiji	64	56	3584	224		1000	4 GB HBM1	0500	4096 bits	07168	0448	064	224	0512
Radeon R9 Nano	Sep 10 2015	Fiji	64	64	4096	256		1000	4 GB HBM1	0500	4096 bits	08192	0512	064	256	0512
Radeon R9 Fury X	Jun 24, 2015	Fiji	64	64	4096	256		1050	4 GB HBM1	0500	4096 bits	08601.6	0537.6	067.2	268.8	0512
Radeon Pro Duo	Apr 26, 2016	2 × Fiji (Gemini)	2 × 64	2 × 64	2 × 4096	2 × 256		1000	2 × 4 GB HBM1	0500	2 × 4096 bits	16384	1024	128	512	1024

Power consumption data

model	Type	Consumption ( watt )				additional power plug
		TDP	Readings
			Idle	3D load	Maximum load
Radeon R5 330 (OEM)	Oland	050 W	k. A.		k. A.	no
Radeon R5 340 (OEM)	Oland	065 W
Radeon R7 340 (OEM)	Oland	050 W	k. A.		k. A.	no
Radeon R7 350 (OEM)	Oland	065 W
Radeon R7 360	Tobago	100 W				1 × 6 pin
Radeon R7 370	Trinidad	110 W	11 W.	108 W.
Radeon R9 360 (OEM)	Tobago	085 W	k. A.		k. A.	1 × 6 pin
Radeon R9 370 (OEM)	Trinidad	150 W
Radeon R9 380 (OEM)	Antigua	190 W				2 × 6-pin
Radeon R9 380	Antigua	190 W	14 W.	179 W
Radeon R9 380X	Antigua	190 W	14 W.	188 W
Radeon R9 390	Grenada	275 W	15 W	250 W		1 × 6-pin 1 × 8-pin
Radeon R9 390X	Grenada	275 W	15 W	303 W
Radeon R9 Fury	Fiji	275 W	15 W	256 W		2 × 8 pin
Radeon R9 Nano	Fiji	175 W	13 W.	183 W		1 × 8 pin
Radeon R9 Fury X	Fiji	275 W	21 W	289 W		2 × 8 pin
Radeon Pro Duo	2 × Fiji (Gemini)	350 W	k. A.			3 × 8 pin

Remarks

1. ↑ The date indicated is the date of the public presentation, not the date of availability of the models.
2. ↑ The specified performance values ​​for the computing power via the stream processors, the pixel and texel fill rate, as well as the memory bandwidth are theoretical maximum values ​​(with standard clock, if available), which are not directly comparable with the performance values ​​of other architectures. The overall performance of a graphics card depends, among other things, on how well the available resources can be used or fully utilized. There are also other factors that are not listed here that affect performance.
3. ↑ ^{a b} The specified clock rates are the reference data recommended or specified by AMD, the I / O clock (in brackets the effective clock) is specified for the memory clock. However, the exact clock rate can deviate by a few megahertz due to different clock generators, and the final definition of the clock rates is in the hands of the respective graphics card manufacturer. It is therefore entirely possible that there are or will be graphics card models that have different clock rates.
4. ↑ ^{a b c d e f g} The model is an OEM product that is not available on the retail market and, despite the partially identical designation, can differ significantly from these variants.
5. ↑ The TDP value specified by AMD does not necessarily correspond to the maximum power consumption. This value is not necessarily comparable with the “MGCP” value of the competitor Nvidia.
6. ↑ The measured values ​​listed in the table relate to the pure power consumption of graphics cards that correspond to the AMD reference design. A special measuring device is required to measure these values; Depending on the measurement technology used and the given measurement conditions, including the program used to generate the 3D load, the values ​​can fluctuate between different devices. Therefore, measured value ranges are given here, each representing the lowest, typical and highest measured values ​​from different sources.
7. ↑ The value given under 3D load corresponds to the typical game usage of the card. However, this is different depending on the 3D application. As a rule, a modern 3D application is used to determine the value, which, however, limits the comparability over longer periods of time.
8. ↑ The maximum load is usually determined with demanding benchmark programs, the loads of which are significantly higher than those of "normal" 3D applications.

Web links

• Manufacturer website for the Radeon R9 OEM series
• Manufacturer website for the Radeon R7 OEM series
• Manufacturer website for the Radeon R5 OEM series

Individual evidence

1. ↑ ^{a b c d e f g h i j k l m n o p q} Power consumption of current and past graphics cards. 3DCenter.org, February 23, 2014, accessed August 8, 2015 . 
2. ↑ ^{a b} AMD Radeon R9 Fury X launch analysis (page 3). 3DCenter.org, June 25, 2015, accessed August 11, 2015 . 
3. ↑ AMD Radeon R9 Fury X in the test: 4 gigabytes for a high-end card. PC Games Hardware, June 28, 2015, accessed August 12, 2015 . 
4. ↑ AMD Radeon R9 Fury X in the test - page 21: Memory usage with HBM. hardwareLUXX.de, June 24, 2015, accessed on August 12, 2015 . 
5. ↑ Even with HBM, 4 GB is only 4 GB. Golem.de, June 24, 2015, accessed on August 12, 2015 . 
6. ↑ Launch analysis of the AMD Radeon R9 Fury. 3DCenter.org, July 11, 2015, accessed May 16, 2016 . 
7. ↑ The Nano / Fury prices have to go down. 3DCenter.org, November 26, 2015, accessed May 16, 2016 . 
8. ↑ ^{a b} AMD Radeon R9 Nano launch analysis. 3DCenter.org, July 11, 2015, accessed May 16, 2016 . 
9. ↑ AMD Radeon R9 Nano in the test: The fastest small graphics card for Mini-ITX (page 3). ComperBase, July 10, 2015, accessed May 16, 2016 . 
10. ↑ Radeon Pro Duo: 1,600 euros for two Radeon R9 Nano on one PCB. ComperBase, April 26, 2016, accessed May 16, 2016 . 
11. ↑ AMD Radeon Pro Duo officially released: $ 1,499 for two Fiji XTs. PC Games Hardware, April 30, 2016, accessed May 16, 2016 .